▶ Watch the full episode and read the transcript: Lucas Thelosen on the Data Faces Podcast

At almost every data, analytics, and AI conference I’ve attended lately, someone on stage has announced that BI is dead, so I asked Lucas Thelosen directly: is BI dead? He built one of the largest Looker consultancies, helped lead Looker into Google, and now builds an AI analyst. “It’s always evolving,” he said, but one thing is different this time. If you ask the product leaders at the major dashboarding products, “they will tell you in private that actual usage numbers are really low.”

That matches my previous experience. I’ve spent years marketing analytics software at Dell, TIBCO, Alteryx, and other vendors, and we all promised analytics for everybody. Yet most companies ended up with thousands of dashboards and little insight. The data is slightly dated but still relevant: when BARC and Eckerson Group measured it, about 25 percent of employees actively used BI and analytics tools, a figure that had barely moved in seven years.^[1]

So, is this the end of BI as we know it? Probably, and to steal a line from R.E.M., I feel fine. Why? Well, an AI analyst can do the digging most employees never had time for. Still, it builds on the foundations BI teams already know: trustworthy math, numbers that agree, and documentation someone keeps current and maintains.

About Lucas Thelosen

is the founder and CEO of Orion by Gravity, the company behind Orion, an AI analyst. Lucas started as an analyst during the 2008 financial crisis and put his first AI into production in 2010 for Countrywide. After founding DAS42, one of the largest Looker consultancies, he joined Looker’s executive team and, when Google acquired Looker, went on to lead Google Cloud’s data and AI product organization.

In this episode, Lucas and I discuss:

- Whether BI is dead, and why so few employees open the dashboards their company builds

- Why Orion by Gravity put deterministic code between the AI and the math

- How an AI analyst can personalize every answer and still keep everyone’s numbers the same

- What “ambient learning” means for keeping AI context up to date

- What happens to analysts, and why Lucas thinks administering AI agents is next year’s big topic

Watch the full conversation here:

Who does the digging when nobody opens the dashboard?

Lucas doesn’t expect dashboards to disappear, since teams still need them for weekly meetings and quarterly business reviews. What changes is who does the digging. Most people care about beating their quarterly goals, and an AI analyst lets them state that goal and then “run the 120 queries and permutations” to find what matters. “You don’t have the time to dig deep,” Lucas said. “You don’t have the time to understand the tools to ask the questions.” His best example is something he calls “red blindness.” People build spreadsheets as wide as their screen and dashboards as long as their screen, then look at the red numbers because that’s where the problems seem to be.

An AI analyst looks at both colors, Lucas said, “because it has infinite time.” He used to race Orion to find the $10 million insight in a new customer’s data, and two years ago he usually won. “But it really quickly changed,” he said. The data got too messy for him to handle, while “Orion doesn’t give up. It keeps going. It finds the nuggets.” An AI analyst will keep checking green numbers long after a person stops, which also means it will find every weakness in the data underneath.

Keep the AI away from the math

But can you trust what an AI analyst finds? When I told Lucas that AI makes up a lot of crap, he agreed. Gravity started building Orion about two years ago, when the models “were even dumber than they are now,” so the team put deterministic code wherever it could. SQL and Python run the calculations, the AI interprets the question and builds the output, and agents from different AI companies check each other’s work. That design is slower and costs more to run, and Lucas thinks the trade is easy in analytics: if a correct dashboard costs ten dollars to produce rather than three, it’s worth it.

The design also makes the work auditable, which matters in finance, healthcare, and insurance. Lucas calls it a glass box, where you can see which assumptions the AI analyst made, how confident it is in each one, and every query it ran. That log matters because deterministic code protects only the math, and if the AI misreads the question, you catch it in the assumptions.

One metric tree, many personal answers

In the dashboard era, I sat in plenty of meetings where someone presented an answer, and someone else immediately had a different number from a different dashboard. An AI analyst for each person sounds like a recipe for more of the disaster we’ve already lived through, a hyper-personalized, confidently wrong answer for every employee. So how do you keep everyone’s numbers in agreement?

Lucas describes analytics as a linear workflow, with personalization at both ends. The question going in and the answer coming out are both tailored to each person. Everything in between is shared, so everyone with similar questions, metrics, and teams runs through the same SQL and Python, organized as a metrics tree. “For me, right? I’m German. I need three bullet points, four words each,” Lucas said. “For you, it might be a little different, but the numbers will be the same.”

Data teams have long put that shared middle in a semantic layer, and Lucas thinks a semantic layer covers less than its vendors claim. Along with ontologies and data dictionaries, he said, “semantic layers maybe get you 40 percent there.” Orion doesn’t need one to start. It reads the query log, checks its numbers against past reports, and asks people to clarify definitions such as revenue. Malcolm Hawker of Profisee made a related argument on an earlier episode about the limits of the semantic layer when a company holds 15 versions of the same customer.

Context is documentation with a maintenance plan

Everyone is talking about context right now, and Lucas thinks we’ve been here before. “We called it documentation back then,” he said, to which I’d add that most of it was outdated wiki pages nobody read. Context can also be overdone, like handing a new hire a stack of books on their first day and asking, “But what do you really want me to focus on?” His rule is the right context at the right time.

Context also needs ongoing maintenance, which Orion by Gravity handles with what Lucas calls ambient learning. Orion sits in team chat channels, and many customers also connect their meeting note-taker summaries. When it detects a change in a team’s thinking, it proposes an edit to the knowledge base. One large retailer decided to focus on apparel rather than shoes for a quarter. Without that context, the AI analyst would keep asking why the money wasn’t going to shoes, and a quarterly goal like that never makes it into a semantic layer. On Ep 53, Ken Sanford argued that tribal knowledge is the AI context layer nobody names, and ambient learning is one way to keep it current.

Dashboard sprawl becomes agent sprawl

When ChatGPT came out, every software company added a chatbot. Lucas wondered why nobody built an administrator, since he saw the biggest pain in enterprise software as administering it. “Why didn’t Salesforce, the first thing they did, build an AI that administers Salesforce for you?” he asked. Orion by Gravity built one early. It flags dashboards and slide decks nobody uses anymore and folds them into the ones people do use. It also highlights conflicting definitions for an administrator to resolve.

He expects administration to be next year’s big conversation, because agents are about to sprawl the way dashboards did and nobody is asking who governs them. Some were set up by employees who have since left the company, and as Lucas put it, “nobody knows anymore why the agent is still running.” I made a similar argument after the BARC Data & Analytics Retreat in June, where the takeaway was that governance now decides whether AI delivers value. There’s also a second kind of sprawl I don’t think anyone has answered yet. A shared metric tree keeps everyone’s numbers the same, but when everyone can generate a personalized dashboard on demand, each person decides which numbers to highlight. What happens when two executives walk into the same meeting with the same data and different stories? And who decides which view is right?

Why I feel fine

What happens to the people who do this work today? Lucas is direct about it: “There’s going to be fewer and less data scientists and just straightforward analysts, because the AIs are really good at it.” He suggests two new job titles. A context architect designs and maintains context at every level, from one person up to the whole company. A data product manager runs data like a product, “with a roadmap and a vision,” moving the organization from what happened to why it happened, then to optimization. Capital One’s data team described that role in practice on Ep 48, when they talked about building data products for the hundredth use case. For anyone who wants neither, Lucas offered a third title, “senior human in the loop,” and I’m considering it for my LinkedIn profile.

Which brings me back to R.E.M. Keeping calculations deterministic and metrics consistent is old BI work, and so is keeping documentation current and pruning what nobody uses. Any team that did that work in the dashboard era is ready for an AI analyst, and the rest will learn about their inconsistencies quickly, because an AI analyst digs through everything. Here are three checks I’d run this quarter:

1. Find out which dashboards anyone opens: Pull the usage logs from your BI tool. Anything nobody has opened in a quarter is a candidate for retirement or for an AI analyst.

2. Decide what has to stay deterministic: List the calculations you’d have to defend to an auditor or your CFO, such as revenue, churn, and forecast logic, and require any AI analyst you evaluate to run them in code it can show you.

3. Name an owner for the context and for every agent: Someone has to keep the knowledge base current, and every agent needs a person who can explain why it’s still running.

Get those three right, and you can sing along with Michael Stipe.

Listen to the full conversation with Lucas Thelosen on his Data Faces Podcast episode page.

Based on insights from Lucas Thelosen, Founder and CEO of Orion by Gravity, featured on the Data Faces Podcast.

Share

Leave a comment

Podcast highlights

- [0:06] David introduces Lucas Thelosen and the topic: the future of analytics

- [1:02] From an analyst in the 2008 financial crisis to Looker and Google

- [3:52] Is BI dead?

- [6:42] “AI makes up a lot of crap,” and why Orion by Gravity keeps the AI away from the math

- [8:50] Personalized answers with the same numbers across teams

- [11:12] Context architects and data product managers

- [13:45] Can you have too much context?

- [19:14] Red blindness: nobody drills into a green number

- [20:31] The glass box audit trail

- [21:45] The AI administrator nobody built

- [23:32] Ambient learning, and the retailer that moved from shoes to apparel

- [28:10] Do you need a semantic layer to start?

- [31:14] Repeatable workflows and the 7 a.m. executive podcast

- [32:32] Administering the agents: next year’s conversation

- [33:39] When Breath Becomes Air

Frequently asked questions

Is BI dead?

Business intelligence (BI) isn’t dead, but the way people use it is changing. Lucas Thelosen, founder and CEO of Orion by Gravity and former head of product for data and AI at Google Cloud, says dashboards will stay for team meetings and quarterly business reviews. The digging moves to AI analysts, because few employees open dashboards: a 2022 BARC and Eckerson Group study found about 25 percent of employees actively use BI and analytics tools. This figure barely moved in seven years.

What is an AI analyst?

An AI analyst is software that answers business questions by querying a company’s data, investigating the results, and producing slide decks, dashboards, or written answers. A person can give it a goal, such as hitting a quarterly target, and it runs the queries a human analyst would run. Orion by Gravity is one example. Lucas Thelosen says an AI analyst’s advantage is patience, since it checks the green numbers that people skip.

How can you trust an AI analyst’s numbers?

Keep the AI away from the math. Gravity built its Orion AI analyst so that deterministic SQL and Python run the calculations, while the AI model interprets the question and formats the answer. Agents from different AI companies check each other’s work. Lucas Thelosen also recommends a “glass box” audit trail that logs the assumptions, confidence levels, and queries behind every answer, which matters in regulated industries like finance and healthcare.

Do you need a semantic layer to use an AI analyst?

A semantic layer helps but isn’t required to start, according to Lucas Thelosen of Orion by Gravity, who estimates that semantic layers, ontologies, and data dictionaries get an AI analyst about 40 percent of the way. An AI analyst can read the database query log, match its numbers to past reports, and ask people to clarify definitions such as revenue or a new customer. Business context, like a quarterly goal, rarely lives in a semantic layer at all.

What is ambient learning in AI analytics?

Ambient learning is how an AI analyst keeps its business context up to date. Gravity’s Orion sits in team chat channels and reads meeting note-taker summaries. When it detects that a team’s priorities have changed, it proposes an edit to its knowledge base. In one example, a retailer moved its quarterly focus from shoes to apparel, and the AI analyst needed that context to stop recommending more investment in shoes.

What will data analysts do when AI does the analysis?

Lucas Thelosen of Orion by Gravity expects fewer data scientists and traditional analysts, and two growing roles. A context architect designs and maintains the business context that AI agents use, from the individual to the company level. A data product manager runs data like a product, with a roadmap, and moves the organization from reporting what happened to explaining why and then optimizing. Analysts still need business knowledge to correct and train an AI analyst.

About David Sweenor

David Sweenor is the founder and host of the Data Faces podcast, where he talks with the people who are making data, analytics, AI, and marketing work in the real world. He is also the founder of TinyTechGuides and a recognized top 10 AI thought leader and international speaker who specializes in practical business applications of artificial intelligence and advanced analytics.

With over 25 years of hands-on experience implementing AI and analytics solutions, David has supported organizations including Alation, Alteryx, TIBCO, SAS, IBM, Dell, and Quest. His work spans marketing leadership, analytics implementation, and specialized expertise in AI, machine learning, data science, IoT, and business intelligence. David holds several patents and consistently delivers insights that bridge technical capabilities with business value.

Books

- Artificial Intelligence: An Executive Guide to Make AI Work for Your Business

- Generative AI Business Applications: An Executive Guide with Real-Life Examples and Case Studies

- The Generative AI Practitioner’s Guide: How to Apply LLM Patterns for Enterprise Applications

- The CIO’s Guide to Adopting Generative AI: Five Keys to Success

- Modern B2B Marketing: A Practitioner’s Guide to Marketing Excellence

- The PMM’s Prompt Playbook: Mastering Generative AI for B2B Marketing Success

Follow David on Twitter @DavidSweenor and connect with him on LinkedIn.

^[1]Bange, Carsten, and Wayne Eckerson. “Strategies for Driving Adoption and Usage with BI and Analytics.” BARC, April 5, 2022. https://barc.com/news/new-study-identifies-drivers-of-bi-and-analytics-adoption-in-companies-today/

▶ Watch the full episode and read the transcript: Why De-Biasing AI Is Impossible | Leigh Felton

Ask most executives what generative AI is for, and you’ll hear some version of the same answer. It makes people more productive and faster. Leigh Felton spent 20 years at Microsoft, where she headed the company’s first ethical AI enablement program, and she thinks that answer is where the trouble starts. Speed is easy to see, so it becomes the measure, and the thinking that should happen between the prompt and the published result drops out of the process.

I sat down with Leigh, now president of the AI for Job Security Foundation, on location at the CDOIQ Symposium in Cambridge, Massachusetts.^[1] She has a name for the habit she sees spreading through offices, and she calls it AI think.

It’s an uncomfortable argument for anyone who uses gen AI every day, and I include myself in that group. So, what does AI think cost you? Leigh’s answer covers your productivity and your voice, along with the next generation of people who would otherwise question how work gets done in the age of AI.

About Leigh Felton

is president of the AI for Job Security Foundation, a 501(c)(3) nonprofit that helps communities, students, and faculty understand what AI is, how to control it, and how to keep and create jobs for the future workforce. She spent 20 years at Microsoft, where she headed the company’s first ethical AI enablement program.

She is also the author of the two-volume I Could’ve Been CEO Yesterday series.^[2] Volume one, But AI Algorithms Say I’m a DEI Risk Today, looks at how AI trained on historical data limits who gets an opportunity. Volume two, Does AI Decide Who’ll Be CEO Tomorrow?, argues that autonomous AI shouldn’t be an organization’s end goal. She writes a newsletter on Substack, puts out a weekly video message every Friday, and invites anyone who disagrees with her to start a conversation on LinkedIn.

In this episode, Leigh and I discuss:

- Her two-volume book series on AI, hiring, and who gets to lead

- Why she treats AI as an environment that shapes how people think

- How to talk about AI productivity in the language a chief data officer uses

- AI think, the sea of sameness, and why she won’t let AI genericize her

- Why junior employees will see what 30-year experts stopped questioning

Watch the full conversation here:

What “AI think” means

AI think is the habit of accepting what a gen AI tool produces and passing it along without using your noggin and applying your own critical thinking. Leigh’s examples are familiar to anyone who works with these tools: a developer takes the code a model returns and moves it forward, or a writer publishes the first draft a chatbot hands back. The work gets done, and nobody did much thinking along the way.

She traces the habit back to how the technology was introduced. ChatGPT arrived as artificial intelligence, and Leigh doubts millions of people would have signed up overnight if OpenAI had described it the way she does: “a probabilistic engine that is statistically guessing based on patterns that it recognizes.” When you believe the thing on the other side of the prompt is sentient and intelligent, checking its work starts to become optional. But if you use it regularly, you see its foibles and shortcomings.

Research on gen AI and critical thinking illustrates the point. In a survey of 319 knowledge workers who shared 936 examples of using gen AI at work, researchers from Microsoft Research and Carnegie Mellon University found that higher confidence in gen AI was associated with less critical thinking, while higher self-confidence was associated with more.^[3] The more people trusted the tool, the less they questioned its output. But the people who trusted their own judgment kept checking.

That second finding helps explain Leigh’s stance. She has been in the industry for 30 years, and she told me plainly that AI will never write, strategize, or decipher better than she can. You could hear that as bravado, but the survey suggests it’s protective, because confidence in your own expertise keeps you reviewing the machine’s work.

Why faster output can mean less productivity

When I asked Leigh how she gets a chief data officer (CDO) to care about any of this, she said you have to speak their language. Tell a CDO they need governance and rules, and you’ll get “blah, blah, blah” and a request for the box to check for the regulators. Talk about productivity, the language most large companies speak right now, and you get their attention.

The 19% figure comes from METR, a nonprofit research group. In a randomized controlled trial run in early 2025, 16 experienced open-source developers worked through 246 real tasks, and when they were allowed to use AI tools, they took 19% longer to finish.^[4] Before the study, they expected AI to speed them up by 24%. Even after being slowed down, they believed it had sped them up by 20%.

In February 2026, METR published new data on late-2025 AI tools and estimated an 18% speedup for the developers who returned from the original study, although it called the new results unreliable because so many developers now refuse to code without AI at all.^[5] In the original study, developers felt faster while running slower, which is AI think in a single data point. The work felt fast, so they never checked whether they were coding any faster.

I told Leigh I see the same thing in marketing. Someone sends over an AI draft, fixing it takes longer than writing it would have, and the AI slop cleanup just moves to a different desk. Then I offered what I thought was the fix: stop accepting the first output and use the model as a challenger, asking it for a thousand reasons why something is right and a thousand more why it’s wrong. Leigh pushed back and went a step further, telling me not to ask the AI for the reasons at all. Work through the problem with your own expertise, she said, and then ask the model what you’re missing. Her point stung a little, because my challenger approach still lets the model set the terms and outsources your brain.

The proverbial “sea of sameness”

Leigh and I finished each other’s sentences when the conversation turned to AI-written content. Scroll LinkedIn for five minutes, and you’ll find the same “it’s not this, it’s that” construction in post after post, along with the same rhythm and the same tidy three-part lists. Her diagnosis is that people let AI set the parameters. Even when they ask for a thousand options, the model decides what the options look like, and the writer picks one. Leigh won’t work that way, and when a draft doesn’t sound like her, she rants to her AI about it until it does.

Her picture of good thinking comes from old movies. A writer at a typewriter makes a mistake and throws the crumpled page on the floor. Five minutes later, the writer picks it back up because something on it triggered an idea. Working through a problem looks like that pile of crumpled paper, and she worries that people have stopped doing it. They ask the model for a thousand reasons and take what comes back, and nobody goes back to the floor.

The person who asks why

Near the end of our conversation, I asked Leigh whether she’s optimistic about the future of work. She is. Leigh believes AI will never be as smart and intelligent as people. She also knows that belief has to spread beyond her into schools, universities, and organizations, especially the ones whose leaders think they can replace their associate and junior roles and keep a few 30-year experts to run things.

Then she told the turkey story. Experienced people can see what’s right and what’s wrong and can organize just about anything. Still, they also cut the turkey in half before it goes in the oven, because their mothers did it that way, and so did their grandmothers, and nobody has asked why in generations. Then a junior person walks in and asks. The answer is that their ovens were smaller.

Leigh sees junior employees, especially Gen Z, as the people who notice what experts stopped questioning, as long as they don’t give in to AI think themselves. Jim Jinright made the hiring side of this case on his Data Faces episode about the tech job market. As AI takes over the small, well-defined problems that junior people used to learn on, the bottom rungs of the career ladder disappear. Put the two episodes together, and an organization that cuts its junior roles loses its training pipeline along with the people most likely to ask why the turkey gets cut in half.

How to avoid AI think

How do you keep yourself and your team out of AI think? Here’s what I’ve learned works, and most of it comes straight from Leigh.

1. Set the parameters before you prompt: Decide what good looks like, from your own expertise, before the model offers its version. For writing, that can be a style guide, and for analysis, it’s the hypothesis and the questions you want answered. When I ran a cold-prompt test against my own battlecard workflow, a fresh Claude out-wrote my 2025 prompts, and what still mattered was my judgment about which jobs to automate and how to check the results in the field.

2. Treat the first output as a draft to argue with: Leigh rants at her AI until the draft sounds like her. You can skip the ranting, as long as you say what’s wrong and make the model try again, which is the collaborating and adapting she counts as productivity.

3. Ask the model what you’re missing: Do your own thinking first, then ask the model, “What don’t I see?” Leigh says that kind of back-and-forth expands what your AI can do for you, and the co-adaptation makes it stronger over time.

Team leaders should also look at what they measure. When cycle time is the only AI metric on the dashboard, the team gets rewarded for AI think, so pair it with a measure of quality and protect the junior roles where people still learn to ask why.

Leigh’s first piece of advice to me was simple: don’t let AI replace your brain. Faster tools make that advice more important, since the faster the output arrives, the easier it is to skip checking it.

Listen to the full conversation with Leigh Felton on her Data Faces Podcast episode page.

Based on insights from Leigh Felton, President of the AI for Job Security Foundation, featured on the Data Faces Podcast.

Share

Leave a comment

Podcast highlights

- [0:00] David welcomes Leigh Felton to Data Faces, on location at the CDOIQ Symposium in Cambridge, Massachusetts

- [0:25] What Leigh wanted to be when she grew up

- [1:06] Inside the AI for Job Security Foundation

- [1:36] Volume one of I Could’ve Been CEO Yesterday, and AI trained on historical data

- [3:29] Volume two, and why autonomous AI shouldn’t be the end goal

- [4:48] AI as an environment that shapes how people think

- [6:55] How to get a chief data officer to care, in the language of productivity

- [8:39] Faster versus more productive, and what Leigh calls AI think

- [9:41] Why accepting the first output is the most common mistake

- [10:09] Don’t let AI genericize you, and the sea of sameness

- [13:30] Will AI ever be smarter than people?

- [14:14] The turkey story, and why junior employees matter

- [15:45] Where to find Leigh’s books, newsletter, and weekly video

Frequently asked questions

What is AI think?

AI think is a term Leigh Felton, president of the AI for Job Security Foundation, uses for accepting what a generative AI tool produces and publishing it without applying your own critical thinking. A developer who moves model-generated code forward unchecked, or a marketer who posts a chatbot’s first draft, is doing AI think. Felton argues that people mistake the habit for productivity because it makes work faster, even though no one did the thinking.

How does generative AI affect critical thinking?

A 2025 survey of 319 knowledge workers by Microsoft Research and Carnegie Mellon University found that higher confidence in generative AI was associated with less critical thinking, while higher self-confidence was associated with more. Workers who trusted the tool questioned its output less often, and workers confident in their own expertise kept checking it. The findings support Leigh Felton’s argument that people who rely on their own judgment are less likely to fall into AI think.

Does AI make developers less productive?

The evidence has changed over time. METR’s early-2025 randomized trial found that experienced open-source developers took 19% longer on tasks when they could use AI tools, even though they believed AI had sped them up by 20%. In February 2026, METR estimated an 18% speedup with late-2025 tools for returning developers but called that data unreliable. The finding that has held up is that developers’ sense of speed didn’t match their measured speed.

Why does AI-generated content all sound the same?

AI-generated content tends to sound the same because writers let the model set the parameters. Even when people ask for many options, the model decides what those options look like, so posts converge on the same constructions, such as “it’s not this, it’s that.” Leigh Felton calls the result the sea of sameness. Her fix is to bring your own voice and expertise first and push back on drafts that don’t sound like you.

Should companies replace junior roles with AI?

Leigh Felton argues against it. Experienced employees often keep doing things the way they always have, and junior employees, especially Gen Z, are more likely to ask why. Her example is a family that cuts the turkey in half for generations because an earlier oven was too small, until someone new asks. Cutting junior roles removes the people most likely to question inherited habits, along with the small problems they learn on.

About David Sweenor

David Sweenor is the founder of TinyTechGuides and host of the Data Faces Podcast. He is an international speaker, advisor, and the author of eleven books on artificial intelligence, analytics, and B2B marketing, including Generative AI Business Applications, The CIO’s Guide to Adopting Generative AI, and Modern B2B Marketing. With more than twenty-five years in analytics and AI at companies including Alteryx, Tableau, TIBCO, SAS, IBM, and Dell, David advises technology companies on product marketing, content strategy, and go-to-market execution. He holds several patents and has been named a top influencer in data and analytics by Onalytica, Thinkers360, and Analytics Insight.

Connect with David on LinkedIn and subscribe to the Data Faces Podcast for conversations with the people shaping enterprise data and AI.

Footnotes

^[1]CDOIQ Symposium. “The 20th Annual CDOIQ Symposium.” July 21–23, 2026, Hyatt Regency Cambridge, Massachusetts.

https://2026cdoiq.org/

^[2]Felton, Leigh. I Could’ve Been CEO Yesterday series (V1: But AI Algorithms Say I’m a DEI Risk Today; V2: Does AI Decide Who’ll Be CEO Tomorrow?). Amazon author page, accessed October 1, 2026. https://www.amazon.com/stores/Leigh%20Felton/author/B0C442RJ6R

^[3]Lee, Hao-Ping (Hank), Advait Sarkar, Lev Tankelevitch, Ian Drosos, Sean Rintel, Richard Banks, and Nicholas Wilson. “The Impact of Generative AI on Critical Thinking: Self-Reported Reductions in Cognitive Effort and Confidence Effects From a Survey of Knowledge Workers.” Proceedings of the 2025 CHI Conference on Human Factors in Computing Systems, Association for Computing Machinery, April 2025. https://doi.org/10.1145/3706598.3713778

^[4]Becker, Joel, Nate Rush, Beth Barnes, and David Rein. “Measuring the Impact of Early-2025 AI on Experienced Open-Source Developer Productivity.” METR, July 10, 2025. https://metr.org/blog/2025-07-10-early-2025-ai-experienced-os-dev-study/

^[5]METR. “We Are Changing Our Developer Productivity Experiment Design.” METR, February 24, 2026. https://metr.org/blog/2026-02-24-uplift-update/

In October, a client blog passed every check that my system put it through and still wasn’t ready to publish. A dozen pieces of internal shorthand borrowed from a sales deck, phrases like “the objections” and “the edge,” got through two automated style checkers, an editorial review, and a copy edit. I caught them all in one read because I hadn’t worked from the deck while writing the draft. The style checkers looked for banned words, and the AI reviewers before me had all worked from the same deck the draft came from, so the shorthand looked like plain English to all of them. The more go-to-market (GTM) content an AI system produces, the more a miss like that costs you, because nothing in the process was built to catch it.

Catching problems like that systematically and reliably takes more than a good prompt. I’ve now written more than a hundred Claude skills across four Claude Code workspaces, one for my own business and the rest for client work, and the same four layers keep turning up underneath every GTM system that I’ve built that works:

1. Skills: packaged jobs, such as building a battlecard, that you trigger by name.

2. Context: the written-down facts, rules, and voice that every job reads before it starts.

3. Orchestration: what decides when a job runs and what has to finish before it, plus where its output goes.

4. Evals: the checks that run on the output before it reaches a rep or a customer.

Most teams that I talk to have built one of these and stopped — usually skills. Any one layer on its own is a toy, something that impresses in a demo and breaks in daily use, and each layer fails in its own predictable way. Of the three components I described in What AI GTM engineering is, context comes first; skills and orchestration make up the second; evals check the output of both; and enablement is the third, focused on the people who run the system. These four layers are the stack that the competitive intelligence system I promised in my last post will run on. That post described one piece of it, a battlecard system, which I’ll use as the example here.

A battlecard skill waits for somebody to remember it

My battlecard skill is a good example of the skills layer. It’s one packaged job, triggered by name, and it knows which sections a card needs, which deal scenarios to cover, and which claims a rep should never make. In September, I asked a fresh Claude session, with no files or history, to write a battlecard prompt. It wrote one about as good as mine, which shows how cheap this layer has become.

On its own, though, a skill runs only when somebody remembers it exists, and it runs on whatever that person pastes in. A competitor that changes its pricing in March can go without a new card until June, when a rep finally asks for one, and the skill builds it from whichever notes the rep has open. That’s a prompt library with better packaging, and it keeps most marketing teams at the personal stage of the AI GTM maturity model, where AI is episodic and depends on someone to run it.

Context only counts when the jobs read it

Before a job writes a word, it should know your positioning, voice, and the corrections that come back from sales. It should also know which file is the master copy of each list that you maintain, such as competitors or target accounts. I’ve written about the components that hold that context and what Claude’s memory adds over time, and context engineering gets its own post later in this series. Here, I’ll stick to the bugaboo that bites most teams.

Writing context down isn’t enough, because each job has to find the master copy and read it the way you meant it. In early September, a job of mine updated my content calendar using a column that it mistook for the row number. That column held an old article ID, the sheet wasn’t sorted by it, and the job overwrote two rows and created a duplicate before I caught it. The skill worked as written, and the context existed. Nothing in the context said which column identifies a row, so the job guessed. Every calendar job now reads a rule that names the ID column before it writes, along with an instruction to read the row back afterward.

Orchestration repeats your mistakes on schedule

For one client, a market intelligence scan runs on the 16th of every month at 4 a.m., without anyone starting it. The day before, a Slack message gives me a heads-up so that I can add anything that the scan should cover. Once the scan runs, it files the results in Notion and Google Docs and sends me a summary email. In the battlecard system from my last post, a site monitor watches competitors’ web pages, and a pricing change that it catches is one of four events that tell the battlecard skill to rebuild a card. The other three are a new win-loss analysis, a major deal won or lost, and a correction from the field.

Orchestration also repeats whatever is underneath it, mistakes included. When the corrections that a scheduled rebuild reads from are wrong, the rebuild copies the wrong claim into every refreshed card without anyone touching it. Scheduled jobs also fail where you can’t see them. In September, one of my publishing jobs ran an hour early because of a time-zone mismatch, and I didn’t see its desktop notification for a day, so those publishing jobs now email me when they fail.

For a CMO, or a private equity operating partner overseeing several portfolio companies, the cost of a wrong claim grows with how much the team has automated. A wrong claim in a hand-built battlecard reaches the reps who open it. The same claim in an orchestrated system reaches every rep, every refreshed feature matrix, and every web page built from it, every time it runs, and it looks just as polished as the true ones.

Evals are the layer teams build least

An eval checks a system’s output before it reaches a rep or a customer. It compares the output to something you know is true, and a named person owns the result. In the State of Agent Engineering survey from LangChain, which sells tools for monitoring and evaluating AI agents, 89% of organizations reported some form of observability on their agents, while 52.4% ran offline evaluations on test sets.^[1] Watching a system run and testing whether its answers are correct are separate jobs, and I’d bet marketing teams trail software teams on the second by a wide margin.

Most GTM teams already collect the raw material for one eval: corrections from reps. In the battlecard system from my last post, every correction from a rep or the product team goes into a field-intelligence log that records who made it, when, and how confident they are. When the log and the competitor research disagree, the log wins. In that post, a client’s sales materials listed four deployment options when the product had three, and the correction became a test that every battlecard, feature matrix, and web page had to pass. A card that gets a claim wrong in front of the prospect’s own engineers costs the rep credibility with the buyer, and after that, the rep stops opening the card. That’s why checking every claim against the log pays for itself first.

Evals on their own break too. A check with nothing to trigger it runs only when somebody remembers it, which puts you back at the skills problem, and a check against a log that nobody updates passes claims that stopped being true months ago. The check needs the other three layers to tell it when to run and what’s currently true.

The client blog from the start of this post had plenty of checks and no eval of what a reader would understand. Now, a separate AI model reads each draft with nothing but the file, like a first-time reader, and lists every phrase it can’t resolve. The same check applies to the numbers you report up, since a note in my own tracker claimed for three years that the YouTube analytics API couldn’t separate paid views from organic ones, and one test would have shown that it could.

With all four layers in place, here’s how the battlecard from my last post works. The context holds the rule that field corrections outrank the research, and the skill knows what a card needs. Orchestration rebuilds the card when a pricing change or a lost deal makes it wrong, and evals check every claim against the field-intelligence log before a rep sees it.

Four questions for your stack

So, how do you know which layer you’re missing? Answer these for an ordinary week, since your best project will flatter you:

1. Skills: Which of your recurring GTM jobs is packaged as a skill, and which still lives in someone’s prompt history?

2. Context: When two jobs need the same fact, such as a competitor’s pricing, do they read the same file, and does that file record who last corrected it?

3. Orchestration: What event would make your most-used sales asset wrong, and does anything watch for it?

4. Evals: Before a rep says it on a call, would your system catch a claim that could cost the deal?

Most teams can answer the first question and get stuck on the third. In my next post, I’ll build a competitive intelligence system in four layers and show what each layer reads and hands off, plus how its output gets checked. If you’d like to compare notes sooner, I’m happy to spend thirty minutes walking through your stack with you and naming the layer worth building next.

You can buy the tooling for every layer. You can’t buy the record of which claims your reps stopped trusting, and that record is what the context and evals are built from.

Interested in learning more? Book a meeting.

Share

Leave a comment

Frequently asked questions

What is an AI GTM stack?

An AI go-to-market (GTM) stack is the set of layers a GTM team needs for AI work to run as a system rather than as one person’s habit. Skills are packaged jobs, like building a battlecard, and context is the facts, rules, and voice those jobs use. Orchestration decides when each job runs, and evals check AI output before it reaches a rep or a customer.

What is an eval in an AI marketing system?

An eval checks an AI system’s output before it reaches a rep or a customer. It compares the output to something known to be true, such as a log of sales corrections, and a named person owns the result. In LangChain’s State of Agent Engineering survey, 89% of organizations reported some form of observability on their AI agents, and 52.4% ran offline evaluations on test sets.

How is orchestration different from a Claude skill?

A Claude skill is one packaged job, such as building a battlecard, that runs when somebody triggers it. Orchestration decides when that job runs without anyone asking, and what runs first. When a site monitor catches a competitor’s pricing change and tells a battlecard skill to rebuild the card, the skill becomes part of a system. Orchestration also repeats mistakes, so it needs evals to check its output.

Why isn’t a library of Claude skills enough for a marketing team?

A Claude skill runs only when somebody remembers to run it, and it works from whatever that person pastes in. A battlecard for a competitor that changed its pricing in March may not be rebuilt until a rep asks for it in June. Without context that the skill reads, orchestration that runs it, and evals that check it, a skill library is a prompt library with better packaging.

Where should a marketing team start building an AI GTM stack?

Answer four questions for an ordinary week. Which recurring GTM jobs are packaged as skills? Do jobs that need the same fact read the same file? What event would make your most-used sales asset wrong, and does anything watch for it? Would your system catch a claim that could cost a deal before a rep says it? Most teams can answer the first and get stuck on the third.

How does the AI GTM stack relate to AI GTM engineering?

AI GTM engineering is the practice of building AI systems into how a go-to-market team works, and it has three components: written-down standards and context, workflows that run without anyone starting them, and enablement for the people who operate them. Context is the first component. Skills and orchestration together make up the second, since skills are the jobs and orchestration runs them without anyone starting them, and evals check both outputs. Enablement covers the people who run the stack.

About David Sweenor

David Sweenor is the founder and host of the Data Faces podcast, where he talks with the people who are making data, analytics, AI, and marketing work in the real world. He is also the founder of TinyTechGuides and a recognized top-10 AI thought leader and international speaker specializing in practical business applications of artificial intelligence and advanced analytics.

With over 25 years of hands-on experience implementing AI and analytics solutions, David has supported organizations including Alation, Alteryx, TIBCO, SAS, IBM, Dell, and Quest. His work spans marketing leadership, analytics implementation, and specialized expertise in AI, machine learning, data science, IoT, and business intelligence. David holds several patents and consistently delivers insights that bridge technical capabilities with business value.

Books

- Artificial Intelligence: An Executive Guide to Make AI Work for Your Business

- Generative AI Business Applications: An Executive Guide with Real-Life Examples and Case Studies

- The Generative AI Practitioner’s Guide: How to Apply LLM Patterns for Enterprise Applications

- The CIO’s Guide to Adopting Generative AI: Five Keys to Success

- Modern B2B Marketing: A Practitioner’s Guide to Marketing Excellence

- The PMM’s Prompt Playbook: Mastering Generative AI for B2B Marketing Success

Follow David on Twitter @DavidSweenor and connect with him on LinkedIn.

Footnotes

^[1]LangChain. “State of Agent Engineering.” LangChain, June 12, 2026. https://www.langchain.com/state-of-agent-engineering.

▶ Watch the full episode and read the transcript: Gerry Van Zandt on the Data Faces Podcast

Anyone who has worked on a Gartner Magic Quadrant, Forrester Wave, or IDC MarketScape submission knows the questionnaire. It runs to hundreds of detailed rows, and I’ve never been sure who on the other side could read it all. Now my clients use AI to help draft those responses, and I leaned on Claude for a client’s Magic Quadrant submission myself this summer. Some analyst firms have started using AI to pre-fill the questionnaire before it reaches the vendor. So, is AI just talking to AI? If so, what of that?

That question sits inside a bigger one that marketing leaders keep asking: are industry analysts still relevant when a buyer can ask a chatbot for a shortlist? Forrester’s 2025 Buyers’ Journey Survey found that 94 percent of business buyers use AI in their purchasing decisions, and twice as many buyers named gen AI as a meaningful source of information than any other source.^[1] Numbers like those fed a meme that ran from last year into this one: AI would displace analysts, and Gartner was dead. Investors took it seriously enough that Gartner’s stock fell more than 50 percent from its 2025 highs.^[2] Gerry Van Zandt has spent close to 30 years in analyst relations (AR), and he argues the opposite. When anyone can generate a vendor comparison in minutes, he thinks the people who can tell you which comparison to trust matter more than they did before.

“AI can tell you what the market appears to believe, but an experienced analyst can tell you whether you should believe that or not.”

— Gerry Van Zandt, Senior Director of Analyst Relations, Alteryx

So what does AI do well in analyst relations, and what does it still need a person for? Gerry is a heavy AI user himself, and his answer bifurcates AR into two jobs.

About Gerry Van Zandt

Gerry Van Zandt is senior director of analyst relations at Alteryx, where he has run AR for five and a half years. Before that, he spent about nine years in AR at Intel, seven and a half at HP, and six at Oracle, and he has served on the board of the Institute of Industry Analyst Relations (IIAR). We worked together at Alteryx, and he has briefed more Gartner analysts than almost anyone I know.

In this episode, Gerry and I discuss:

- What influences a technology buyer in 2026, and how much that list has grown in a decade

- How AI cuts the time and people needed to answer an analyst questionnaire

- Whether Gartner and the other big firms are pay-to-play

- How a startup with no paid relationship landed in a Gartner Cool Vendors report

- What companies still get wrong about analyst relations

Watch the full conversation here:

What AI already does well in analyst relations

Gerry has used AI on analyst requests for information (RFIs) for the last 18 months, including the big, deep ones. By his estimate, AI cuts about 70 percent of the time needed to complete an RFI and reduces the number of internal experts who review a response from 15 or 20 to five or six, saving hours across the company. His method matters as much as the savings. Gerry doesn’t feed a PDF into ChatGPT and wait for a finished questionnaire; he works question by question and section by section, and he curates every answer that comes back. A few years ago, Alteryx branded its early AI capabilities as AiDIN, and that name still turns up in some internal large language models (LLMs). An answer built on a product name that’s two or three years out of date reads fine to a reviewer who doesn’t know better, and an analyst will catch it.

AI handles most of his preparation work as well. Before an executive meets an analyst, Gerry can build a one-page brief in about five minutes that covers the analyst’s focus and mindset, the questions and choke points to expect, and the three things the executive needs to get across. When relevant analyst reports come out, he has Gemini write a one-paragraph summary of each and sends them to internal teams in about three minutes, a job that used to take him an hour or more.

Each of these tasks pulls together existing information and drafts something useful from it. I’d call that the aggregation job, and it’s the kind of recurring go-to-market work I’ve been writing about as AI GTM engineering: a task becomes a system with a built-in human review step.

What an analyst knows that a model doesn’t

A technology buyer needs judgment from an industry analyst as much as aggregation, and Gerry gives three reasons why an AI assistant can’t supply that judgment.

Gerry starts with access, because analysts often receive briefings under nondisclosure agreements (NDAs) on roadmaps and strategy that never appear in public, so no scraper or LLM trained on external data can see them. When a vendor briefs an analyst on what’s coming next year, that context shapes the analyst’s view of where the vendor is headed. Gerry counts that as another benefit of a strong analyst relationship, because you can brief a person on what’s coming, and no AI can do the same.

Trust comes next, and Gerry is candid about how the profession has changed. Twenty years ago, a Gartner analyst was often a practitioner with two decades of experience who became an analyst and retired as one. Today, some analysts join right out of school or with three to five years of experience. Even so, he says the big firms “still possess a very high level of institutional trust,” and independent experts and former analysts have a harder time earning it.

“Nobody is going to make an organizational, major platform buying decision based on an AI query and an AI shortlist. They’re going to know they need a throat to choke.”

— Gerry Van Zandt, Senior Director of Analyst Relations, Alteryx

Accountability rounds out his case. When a platform decision goes wrong, a buying committee can point to a published evaluation, and a chatbot can’t answer for anything. Forrester’s own buyer research points the same way. When AI answer engines deliver incomplete or unreliable information, Forrester found that buyers compensate “by seeking validation from trusted sources, emphasizing the value of human contact in the buying process.”^[3]

Is Gartner pay to play?

Before we recorded, I told a few people I was interviewing an AR expert, and nearly all of them asked the same thing: is it pay to play? Gerry has heard the question for 30 years, and his answer for the major firms, and even the minor ones, is “unequivocally no.” He doesn’t believe any Magic Quadrant, Wave, or MarketScape placement depends on how much a vendor spends with the firm. However... money does buy access.

A paid relationship gets you inquiries, briefings, and time with analysts at conferences, which Gerry says “helps grease the skids.” He has also watched analysts shut down sales topics on the spot when a rep joins a call, deferring to the account team because it isn’t their area. I agree with him, and I say that as someone who called the Magic Quadrant process more theater than methodology. Vendors pay handsomely for the right to quote their placement in marketing, but in my experience the research, sales, and events sides of these firms don’t talk to each other.

Gerry’s best evidence comes from early in his career, when he worked AR on the agency side. He got a startup client with no paid Gartner relationship into a Cool Vendors report. This Gartner report usually profiles four to six early-stage companies doing something innovative in a market. He didn’t know the analyst beforehand, so his job was to figure out who the right analyst was and brief them. Within two days of the report coming out, the startup’s CEO told him that Michael Dell personally called to learn more about what the company was doing.

For a marketing leader with a small budget, that story matters more than the pay-to-play debate. Gartner, Forrester, and IDC will brief a vendor that has no contract with them, especially one doing something interesting, and no AI tool can get that access on your behalf.

“Good relationships are the currency with which effective analyst relations is achieved.”

— Gerry Van Zandt, Senior Director of Analyst Relations, Alteryx

Where marketing teams get analyst relations wrong

Analysts have written plenty of columns about what they dislike in AR people, and Gerry says the complaint has changed. Twenty years ago, companies expected AR to act as a gatekeeper that metered every contact between an analyst and an executive. He thinks that model stopped working once people became easier to reach, and he now sees AR’s job as preparing executives and shaping messages so that each engagement is productive.

Today, analysts more often complain that AR people don’t know the business well enough to speak for the products, so every substantive question gets routed to an expert. Gerry has deliberately built his own product knowledge at Alteryx so he can represent the company himself and bring in an expert when the conversation needs one.

The list of people who shape a technology purchase has also grown. A decade ago, an executive making a platform decision would rely on a handful of inputs, mainly analyst reports, peers, and the trade press. Now there are podcasters and influencers with larger followings and opinions, former analysts running their own consultancies, and vendor shortlists that gen AI chatbots produce on request, and Forrester counts nine external influencers in a typical business purchase.^[4] Gerry maps each type to a stage in the buying cycle, with press, blogs, and LinkedIn driving awareness and analysts mattering most when a buying team narrows its shortlist. An analyst’s view also travels through the journalists and independent experts who repeat it, which he calls a multiplier effect.

Forrester, IDC, and Gartner have each responded to AI differently. Forrester launched an agent for Microsoft 365 Copilot in April 2026 and offers a Model Context Protocol (MCP) connector so clients can bring its research into their own systems.^[5] IDC offers a plugin that puts its research and data inside Claude.^[6] Gartner has built its gen AI tool, AskGartner, into its own client platform, and Gerry reads the firm’s recent signals as a commitment to keeping clients in that portal.^[7] He admits AskGartner was “pretty crummy at first” and says it’s “pretty good now.” Either way, a marketing team’s AR plan has to account for analyst research that now reaches buyers inside the AI tools they already use.

What to do with the hours AI gives back

So, where should a marketing leader put AR time and budget now? Gerry thinks AR is more relevant than ever, and that it has to cover far more ground than Gartner, Forrester, and IDC. Here’s how I’d act on that:

1. Turn the aggregation work into systems: RFI drafting, executive briefing prep, and research summaries come around every quarter, so build each one once as a repeatable workflow. Write Gerry’s curation step into it, along with a current list of product names, so a retired brand like AiDIN never reaches an analyst. I wrote earlier this month about moving from one-off prompts to GTM systems.

2. Spend the saved hours on relationships: Gerry calls relationships the one constant he has seen in 30 years of AR. Use the time you get back for briefings, inquiries, and roadmap updates under NDA, since that context never reaches an AI assistant, and invest in the product knowledge that lets your AR lead field analyst questions without escalating every one.

3. Map every influencer to a stage of the buying cycle: List the analysts, independent experts, and podcasters your buyers follow, and note which stage each one influences. That map shows where a briefing or a sponsorship does the most good, and it keeps the big three firms in proportion.

Let AI draft the submission, and spend the time it saves making sure the analyst reading it already knows your roadmap and trusts the person who briefed them.

Listen to the full conversation with Gerry Van Zandt on his Data Faces Podcast episode page.

Based on insights from Gerry Van Zandt, Senior Director of Analyst Relations at Alteryx, featured on the Data Faces Podcast.

Share

Leave a comment

Podcast highlights

- [0:05] David introduces Gerry Van Zandt and the question for the episode: are analysts still relevant in the age of AI?

- [4:30] The Air Force pilot dream that turned into a technology career

- [5:41] A two-minute primer on what analyst relations is and why it matters

- [7:23] What influences a technology buyer in 2026, compared with five years ago

- [9:57] Veteran analysts aging out, and why institutional trust has held

- [12:42] “AI can tell you what the market appears to believe”

- [14:54] AI pre-filled RFIs, and the 70 percent time savings

- [17:44] Relationships as the currency of analyst relations

- [19:17] Is it pay to play?

- [20:41] The NDA information no AI can reach

- [23:24] Briefing Gartner with no subscription, and the Cool Vendors story

- [25:09] Weighing a Gartner analyst against an independent with 50,000 followers

- [27:07] What companies get wrong about analyst relations

- [29:38] Five-minute briefing docs and three-minute report summaries

- [32:21] Genealogy and a World War II howitzer battalion

- [35:21] Where CMOs and CEOs should invest in AR

- [37:11] How Forrester, IDC, and Gartner are responding to AI

Frequently asked questions

Are industry analysts still relevant in the age of AI?

Industry analysts remain relevant because AI tools aggregate information, while analysts supply judgment that buyers can trust. Gerry Van Zandt, senior director of analyst relations at Alteryx, argues that AI can tell you what the market appears to believe. Still, an experienced analyst can tell you whether to believe it. Analysts also see roadmap briefings under nondisclosure agreements that no AI model can access, and large platform decisions still need an accountable, independent evaluation to point to.

Is Gartner pay to play?

According to Gerry Van Zandt, who has worked in analyst relations for close to 30 years, major analyst firms aren’t pay-to-play. He doesn’t believe any Gartner Magic Quadrant, Forrester Wave, or IDC MarketScape placement depends on how much a vendor spends. A paid relationship buys access, such as inquiries, briefings, and time with analysts at events, which makes engagement easier. Vendors also pay for the right to quote their placement in marketing materials.

How is AI changing analyst relations?

AI is taking over the aggregation work in analyst relations (AR). Gerry Van Zandt of Alteryx estimates that AI cuts about 70 percent of the time needed to complete an analyst request for information (RFI) and reduces the internal reviewers from 15 or 20 to five or six. He also uses AI to write a one-page executive briefing in about five minutes. He reviews every answer question by question, because AI tools can return out-of-date product names.

Can a startup brief Gartner without a paid subscription?

Yes. Gartner, Forrester, and IDC will take briefings from vendors with no paid relationship, especially vendors doing something innovative. Gerry Van Zandt got a startup with no Gartner contract into a Cool Vendors report, a Gartner report that usually profiles four to six innovative early-stage companies in a market. According to the startup’s CEO, Michael Dell called to learn more within two days of the report’s publication. The work is to find the right analyst and brief them well.

What do companies get wrong about analyst relations?

Twenty years ago, the common mistake was treating analyst relations (AR) as a gatekeeper that controlled every analyst conversation. Today, according to Gerry Van Zandt of Alteryx, analysts’ most frequent complaint is that AR people don’t know the business well enough to speak for the products. He recommends that AR leads build enough product knowledge to represent the company themselves and bring in executives or subject matter experts when a conversation needs them.

How are Gartner, Forrester, and IDC using AI?

The three largest analyst firms are taking different approaches to AI. Forrester launched an AI agent for Microsoft 365 Copilot in April 2026 and offers a Model Context Protocol (MCP) connector for clients’ own systems. IDC offers a plugin that brings its research and data inside Claude. Gartner has built its AskGartner gen AI tool into its own client platform, which Gerry Van Zandt of Alteryx reads as a commitment to keeping clients in the Gartner portal.

About David Sweenor

David Sweenor is the founder and host of the Data Faces podcast, where he talks with the people who are making data, analytics, AI, and marketing work in the real world. He is also the founder of TinyTechGuides and a recognized top 10 AI thought leader and international speaker who specializes in practical business applications of artificial intelligence and advanced analytics.

With over 25 years of hands-on experience implementing AI and analytics solutions, David has supported organizations including Alation, Alteryx, TIBCO, SAS, IBM, Dell, and Quest. His work spans marketing leadership, analytics implementation, and specialized expertise in AI, machine learning, data science, IoT, and business intelligence. David holds several patents and consistently delivers insights that bridge technical capabilities with business value.

Books

- Artificial Intelligence: An Executive Guide to Make AI Work for Your Business

- Generative AI Business Applications: An Executive Guide with Real-Life Examples and Case Studies

- The Generative AI Practitioner’s Guide: How to Apply LLM Patterns for Enterprise Applications

- The CIO’s Guide to Adopting Generative AI: Five Keys to Success

- Modern B2B Marketing: A Practitioner’s Guide to Marketing Excellence

- The PMM’s Prompt Playbook: Mastering Generative AI for B2B Marketing Success

Follow David on Twitter @DavidSweenor and connect with him on LinkedIn.

^[1]Buten, John. “B2B Buyers Make Zero-Click Buying Number One.” Forrester, January 22, 2026. https://www.forrester.com/blogs/b2b_buyers_make_zero_click_buying_number_one/

^[2]Investing.com. “Gartner Earnings in Focus: AI Threat Tests Research Model.” Investing.com, May 4, 2026. https://www.investing.com/news/earnings/gartner-earnings-in-focus-ai-threat-tests-research-model-93CH-4656676

^[3]Forrester. “Forrester’s 2026 Buyer Insights: GenAI Is Upending B2B Buying As Leaders Face Mounting Pressure To Justify Every Dollar Spent.” Forrester, January 21, 2026. https://www.forrester.com/press-newsroom/forrester-2026-the-state-of-business-buying/

^[4]Forrester. “Forrester’s 2026 Buyer Insights: GenAI Is Upending B2B Buying As Leaders Face Mounting Pressure To Justify Every Dollar Spent.” Forrester, January 21, 2026. https://www.forrester.com/press-newsroom/forrester-2026-the-state-of-business-buying/

^[5]Forrester. “Forrester Launches Forrester AI Agent For Microsoft 365 Copilot To Deliver Research-Driven Insights Into Leaders’ Workflows.” Forrester, April 28, 2026. https://www.forrester.com/press-newsroom/forrester-ai-agent-microsoft-365-copilot/

^[6]IDC. “Quanta in Claude and MCP.” IDC Developer, accessed September 29, 2026. https://www.idc.com/developer/quanta/

^[7]Gartner. “AskGartner FAQ: Answers About Gartner’s AI-Powered Research Tool.” Gartner, accessed September 29, 2026. https://www.gartner.com/en/products/ask-gartner-faqs

▶ Watch the full episode and read the transcript: Industry Data Quality Gets a D Plus | Stacie Christensen

Most stalled data projects get the same postmortem. Something goes amok with the platform, you get lots of “help” to try to fix it, and then the budget is yanked. Unfortunately, the next project fizzles in roughly the same manner. Stacie Christensen has spent 22 years in the data industry watching the vicious cycle repeat, and she built her CDOIQ session around it.

I caught up with Stacie on the third day of the CDOIQ Symposium in Cambridge, Massachusetts, where her talk covered why data initiatives fail, usually long before technology is ever the problem.^[1] Technology comes and goes, she told me, and the failures look the same no matter what platform is underneath.

“What usually happens is we have initiatives, and they start struggling, but tech is the first thing that we’re blaming. And it’s usually not what’s actually at fault. What’s at fault is our foundation that’s under that technology.”

— Stacie Christensen, Senior Data Leader, H-E-B

So, if the tool were the problem, why does the same project stall again on the replacement? Stacie’s answer is that the foundation under the tool never changed, and she has a practical way to find out where it’s busted.

About Stacie Christensen

Stacie Christensen is a senior data leader at H-E-B with more than 22 years in the data industry. After an undergraduate degree in computer science and mathematics, she went back for an MBA and is now deep into a doctorate specializing in data. Computer science was her choice because everyone around her called it the hardest class they had ever taken, so she decided to take it, earn an A, and prove them wrong. It turned out to be a wonderful experience, and that’s when she knew technology was her path.

She is also the author of the Data Alignment Framework, a three-book series. Built on Data came out in May, Built on Simplicity came out without an announcement the week before CDOIQ, and Built on Alignment is due this fall.^[2][3] You can find more of her work at dataisnotpersonal.com.

In this episode, Stacie and I discuss:

- Why initiatives that struggle get blamed on technology first, and what is usually at fault instead

- The people-owned issues underneath every data foundation, from accountability to governance

- Circular data flows, and why data moving back upstream is a red flag for hidden complexity

- Where governance belongs in a data pipeline, and why the end of it is too late

- How to change a process without making the people who run it resist you

Watch the full conversation here:

The easy button keeps changing names

Why do organizations keep reaching for the tool? Stacie thinks it comes down to wanting an easy button, and she includes herself in that. Technology is the thing we interact with our problems through, so it’s the natural place to point the finger, and a new system feels easier than confronting the amorphous people-owned issues that caused the trouble in the first place. Every project I’ve worked through, from data warehousing to self-service analytics to data catalogs, arrived with the same promise that this platform would be the panacea that would solve world hunger.

Today, the easy button has a new name. I asked Stacie whether we could just AI it and solve all of our ills, fully expecting her to play along. She said no, flatly, and then explained that AI scales whatever foundation you already have, which means it has nothing positive to scale when the foundation is missing. I opened my CDOIQ field notes with that line, because nobody at the event said it more plainly.

The research agrees with her about where the fault usually sits. RAND interviewed 65 experienced data scientists and engineers about why AI projects fail, and it notes that by some estimates, more than 80 percent of them do, twice the rate of IT projects that don’t involve AI.^[4] Eighty-four percent of those interviewees named leadership-driven causes, such as optimizing for the wrong business problem, as a primary reason. RAND calls them the most frequent cause of failure. In other words, the technology tends to take the blame because it’s the easiest part of a project to replace.

What “foundation” means in practice

When Stacie says foundation, she means a short list of words that make most people wince: accountability, ownership, alignment, and governance. I tried to summarize that on air as “people are the problem,” and she gently corrected me. She pins the trouble on people-owned issues, which is different, and when I brought my line back up later in the conversation, she was quick to point out that those were my words, not hers.

I think her wording matters, because an issue that someone owns is an issue that someone can fix, and blaming people usually just makes them defensive. The same idea runs through her first book, Built on Data, whose subtitle makes the argument for her: How Discipline Creates Usable Data, Not Technology.^[5] She means organizational discipline, the kind of foundational work that keeps governance, data quality, and investment decisions honest.

“It’s the stuff that still needs that human touch that AI cannot come in and answer for us today.”

— Stacie Christensen, Senior Data Leader, H-E-B

Part of that discipline is deciding what data is worth the effort of cleaning up and streamlining. Not every dataset earns that investment, and she says the answer depends on the company, the organization, and the industry. That’s the kind of call that no tool can make on your behalf.

Three questions that expose a foundation problem

Before anyone evaluates a platform, Stacie asks questions that separate a technology problem from a people-owned one, and three of them came up in our conversation. None of them requires a vendor, a budget request, or a steering committee to answer.

Does your data ever flow back upstream?

Circular data flows are the complexity marker Stacie comes back to again and again in her books. A circular data flow happens when data reaches the end of the pipeline, usually analytics and reporting, and someone changes or creates data there and then pushes it back to the top of the pipeline. The consumer does this because the data they received wasn’t what they needed, so they transformed it themselves and now want it back in the warehouse. In Stacie’s view, sending data upstream means something is wrong with the process.

She admits that she gets “a little judgy” when she hunts for complexity, and she thinks you have to. She tests it by asking whether anyone can explain why the loop exists. When the best answer is “we just decided to do this down here,” she treats it as a red flag, and she tackles those outer edge cases before touching the harder work of the pipeline itself.

“Simplicity is probably the most underrated skill in all of technology.”

— Stacie Christensen, Senior Data Leader, H-E-B

That belief is the premise of Built on Simplicity, which covers how to make processes, data flows, and data models simple enough to describe and understand, and from there to govern and scale.^[6] It’s also why the question matters more now. Once those flows are handed to AI, a loop that nobody can explain gets scaled right along with everything else.

Where does governance sit in your pipeline?

Stacie has worked at both extremes. One company invested heavily in data governance and had trustworthy data throughout the pipeline. At another, the term was so taboo you didn’t dare say it in a meeting. Her view is that governance belongs at the beginning of the pipeline, at data creation, and at the critical transitions where data is transformed, because bad data becomes exponentially more expensive the further downstream it travels.

She pushes back hardest on governance that lives only at the end, in analytics and reporting, which she considers far too late in the process. She has seen two structures work well, a centralized governance team deployed upstream and distributed ownership broken up by stage, and she’s cautious about calling either one the right answer. She cares most about whether governance exists where the data is created and changed. I covered why governing at creation keeps coming up in my CDOIQ data quality write-up, so I won’t repeat the whole case here.

Are you collecting data nobody uses?

When I pushed Stacie for a letter grade on the industry’s data quality, she eventually settled on somewhere between a D plus and a C minus, and she pinned it on sheer volume. She thinks we gather so much data that most of it never informs the critical business decisions we need to make. It turns into noise, and then it turns into a backlog of cleanup that nobody needs to run the company.

I told her that people are hoarders, especially with data, because they might need it someday. She agreed that the habit also introduces risk. Her advice on unstructured data followed the same logic. Before investing in it, ask what it’s being used for, because some companies have important data sitting unstructured, and others leave it that way because it isn’t worth the investment. Malcolm Hawker made a similar case on the show when he argued that AI-ready data is the wrong question, since readiness depends on the use case in front of you.

Fixing the foundation without a revolt

Once you find the problem, you still have to fix it with the people who own it. As Stacie put it, people hate to be governed, and walking into an organization to announce that its processes look ridiculous is the fastest way to guarantee resistance. Some of those processes will look ridiculous at first, and she thinks it’s important not to approach them that way.

Her method starts with understanding. She spends time with the people who run a process to learn why they do it that way, because there’s always a reason an organization got where it is. Then she partners with them and helps lead them to the conclusion, so they own the decision rather than having it handed to them. She calls herself a go-with-the-flow person, and she’s wary of any one-size-fits-all answer for how a company should function.

I’ve seen plenty of data programs stall because the change arrived as a mandate from the top, and her approach runs the other way. The next time an initiative stalls, ask a people question before you buy anything. Who owns this data, who can explain why it moves the way it does, and who was never asked?

Listen to the full conversation with Stacie Christensen on her Data Faces Podcast episode page.

Based on insights from Stacie Christensen, Senior Data Leader at H-E-B, featured on the Data Faces Podcast.

Share

Leave a comment

Podcast highlights

- [0:00] David welcomes Stacie Christensen to Data Faces, on location on day three of the CDOIQ Symposium in Cambridge, Massachusetts

- [0:29] Stacie on 22 years in data, computer science, and math, an MBA, and a doctorate in progress

- [1:23] Why she took the class everyone said was the hardest on campus

- [1:58] Her CDOIQ session, and why technology gets blamed before the foundation underneath it

- [2:49] Accountability, ownership, alignment, and governance, the people-owned issues

- [3:10] Why everyone wants an easy button

- [3:35] “Just AI it,” and why AI scales a foundation rather than fixing it

- [4:01] Built on Data and the Data Alignment Framework series

- [4:35] Organizational discipline, and the work that still needs a human touch

- [5:16] Grading industry data quality, and whether we gather too much data

- [6:50] Unstructured data, and deciding what is worth the investment

- [7:35] Data governance at both extremes

- [8:30] Governance at data creation, and why the end of the pipeline is too late

- [9:51] Centralized or distributed governance teams

- [10:59] Built on Simplicity, and simplicity as the most underrated skill in technology

- [11:40] Circular data flows, and getting “a little judgy” about complexity

- [13:16] Getting people on board by understanding why they do things

- [14:26] Where to find Stacie’s books and her website

Frequently asked questions

Why do data initiatives fail?

Most data initiatives fail because of the foundation under the technology, not the technology itself. Stacie Christensen, a senior data leader at H-E-B with 22 years in the industry, points to people-owned issues such as accountability, ownership, alignment, and governance. Organizations blame the tool first because it is the easiest part to replace. RAND’s 2024 study of AI projects found that 84 percent of the practitioners it interviewed named leadership-driven causes as a primary reason for failure.

What is a circular data flow?

A circular data flow happens when data reaches the end of a pipeline, usually analytics and reporting, and someone changes or creates data there and pushes it back to the top of the pipeline. It usually means the data that consumers received wasn’t what they needed, so they transformed it themselves. Stacie Christensen treats circular data flows as a red flag for hidden complexity, especially when nobody can explain why the loop exists.

Where should data governance sit in a data pipeline?

Data governance works best at data creation and at the critical transitions where data is transformed, according to Stacie Christensen of H-E-B. Bad data becomes exponentially more expensive the further downstream it travels, so governing it before it enters the pipeline prevents the most costly errors. She considers governance that exists only at the end of the pipeline, in analytics and reporting, far too late. Both centralized upstream teams and stage-by-stage ownership can work.

Can AI fix a weak data foundation?

AI does not fix a weak data foundation. Stacie Christensen describes AI as a multiplier and a scaler of the foundation an organization already has, so it has nothing positive to scale when governance, ownership, and data quality are missing. Complexity, such as circular data flows, scales along with everything else once those processes are handed to AI, which is why she recommends simplifying data flows before automating them.

How do you get people to accept data governance?

People resist being governed, so the approach matters as much as the policy. Stacie Christensen recommends taking time to understand why a team runs a process the way it does before changing it, because there is always a reason. Then partner with the people who own the process and lead them to the conclusion, so they own the decision rather than having it imposed on them. She avoids one-size-fits-all answers.

What is the Data Alignment Framework?

The Data Alignment Framework is a three-book series by Stacie Christensen, a senior data leader at H-E-B. Built on Data (May 2026) argues that organizational discipline, not technology, creates usable data. Built on Simplicity (July 2026) covers reducing complexity in processes, data flows, and data models before technology scales it. A third book, Built on Alignment, is planned for fall 2026. More of her work is at dataisnotpersonal.com.

About David Sweenor

David Sweenor is the founder of TinyTechGuides and host of the Data Faces Podcast. He is an international speaker, advisor, and the author of eleven books on artificial intelligence, analytics, and B2B marketing, including Generative AI Business Applications, The CIO’s Guide to Adopting Generative AI, and Modern B2B Marketing. With more than twenty-five years in analytics and AI at companies including Alteryx, Tableau, TIBCO, SAS, IBM, and Dell, David advises technology companies on product marketing, content strategy, and go-to-market execution. He holds several patents and has been named a top influencer in data and analytics by Onalytica, Thinkers360, and Analytics Insight.

Connect with David on LinkedIn and subscribe to the Data Faces Podcast for conversations with the people shaping enterprise data and AI.

Footnotes

^[1]CDOIQ Symposium. “The 20th Annual CDOIQ Symposium.” July 21–23, 2026, Hyatt Regency Cambridge, Massachusetts.

https://2026cdoiq.org/

^[2]Christensen, Stacie. Built on Data: How Discipline Creates Usable Data, Not Technology. The Data Alignment Framework. Independently published, May 6, 2026. https://www.amazon.com/dp/B0GXCFL5ZH

^[3]Christensen, Stacie. Built on Simplicity: Rebuilding Data Systems Without the Complexity. The Data Alignment Framework. Independently published, July 15, 2026. https://www.amazon.com/dp/B0H91663NS

^[4]Ryseff, James, Brandon De Bruhl, and Sydne J. Newberry. “The Root Causes of Failure for Artificial Intelligence Projects and How They Can Succeed: Avoiding the Anti-Patterns of AI.” RAND Corporation, August 13, 2024. https://www.rand.org/pubs/research_reports/RRA2680-1.html

^[5]Christensen, Stacie. Built on Data: How Discipline Creates Usable Data, Not Technology. The Data Alignment Framework. Independently published, May 6, 2026. https://www.amazon.com/dp/B0GXCFL5ZH

^[6]Christensen, Stacie. Built on Simplicity: Rebuilding Data Systems Without the Complexity. The Data Alignment Framework. Independently published, July 15, 2026. https://www.amazon.com/dp/B0H91663N

▶ Watch the full episode and read the transcript: Ken Sanford on tribal knowledge and AI skills

The morning I recorded this episode, I interviewed a partner of one of my clients about data quality. When I asked what was missing, the partner answered with two words: tribal knowledge. I had been circling the same idea since July, when 24 interviews with data and AI leaders at the CDOIQ Symposium left me convinced that almost nobody is focused on unstructured data.

Guests on this show keep adding to the list of what AI needs beyond a better model. Josh Howard argued that your AI is dumb without your data, and Malcolm Hawker added that context without truth just produces confident, wrong answers. Semantic layers and knowledge graphs get their turn too, but I rarely hear anyone name the judgment that sits in the heads of the people who have done the job for 25 years.

I have a personal stake in this one. I run most of TinyTechGuides on Claude Code and a folder of skills, the instruction files that tell the AI how I write, publish, and produce the podcast. Those skills are my tribal knowledge written down, and they work because I wrote them myself. So I asked Ken Sanford, an old colleague from our SAS days, how a company gets that kind of knowledge out of its experts and into a form that any AI harness, the tool that runs the model, can use.

About Ken Sanford

Ken Sanford, Ph.D., leads go-to-market and commercial strategy at Clarifeye, a Paris-based startup that captures the unwritten knowledge in experts’ heads and makes it available to AI. A self-described reformed academic economist with a Ph.D. from the University of Kentucky, Ken spent years in data science at SAS, H2O.ai, and Dataiku, and he and I worked together at SAS about 15 years ago.

In this episode, Ken and I discuss:

- Why near-perfect document retrieval still leaves AI without the judgment to interpret what it finds

- How an AI interviewer that hunts for edge cases pulls expertise out of people who never wrote it down

- Why the most tech-forward people in a company are the wrong ones to write its AI skills

- How the memories your AI harness builds are your judgment, and why you should own them

- Where the next generation of experts comes from, and why Ken hires for agency

Watch the full conversation here:

The judgment that never got written down

Ken starts from a premise that sounds backward at first: document retrieval is close to solved. When AI can find nearly any passage in any document, the hard part becomes deciding what that passage means for the decision in front of you. A solid retrieval-augmented generation (RAG) system is only the starting position, in his view, and every company still has to add context around it.

One way to get that context is indirect capture, which means mining Slack, Teams, and Confluence to reconstruct how people think. Ken sees value in it, and he also sees where it stops. Those tools hold what people typed, and the reasoning behind a senior person’s decisions rarely makes it into a chat thread.

“Most of the really important things, most of the intuition, the judgment, was never written down.”

— Ken Sanford, Go-to-Market and Commercial Strategy, Clarifeye

When I suggested that this is a third layer, sitting on top of the data and the context, Ken’s answer was immediate: “That’s exactly it.” He divides a company’s knowledge into the explicit and the implicit, and until recently, the implicit half lived only in experts’ heads. Clarifeye’s job, as he describes it, is to elicit the parts of that knowledge that were never documented.

Interview for the exceptions

Every company I’ve worked at had a wiki that was out of date the day someone finished writing it, because a wiki waits for people to push knowledge into it. Ken’s team started from the other end. Clarifeye’s assistant, Clara, interviews experts the way a consultant would, and she arrives already knowing what the existing documents say, so she only asks about what’s missing.

Clara spends her questions on the edges of a process. She pushes edge cases and hunts for exceptions, often by posing a hypothetical for the expert to rule on, and each ruling makes the AI’s behavior more predictable. The sessions run five or ten minutes a day. When two experts would handle the same situation differently, Clara flags the disagreement, and a designated knowledge owner decides which answer stands.

I asked Ken why Clarifeye’s output looks like old-school enterprise architecture for a generative AI startup: process blueprints, versioned business rules, and reasoning workflows, rather than everything stuffed into a vector store. His answer comes down to readability. People need to be able to read the rules, and the simpler the representation, the more likely a large language model (LLM) is to follow them. For unwritten knowledge, his advice is to “write it old school, pen and paper, write it down.”

The wrong people are writing your AI skills

When I asked whether leaders ever consider that AI could replace them, Ken turned the question toward who builds the AI skills in the first place. Inside most organizations today, skills and custom agents come from the most tech-forward employees, the people staying up late experimenting with ChatGPT. Ken doesn’t think that’s all bad, but it creates a problem he came back to more than once.

“Those people are likely not the ones with the 25 or 30 years of intuition.”

— Ken Sanford, Go-to-Market and Commercial Strategy, Clarifeye

I’ve seen this from the inside. My own skills behave the way they do because of the judgment I put into them, and a colleague writing them on my behalf would have gotten a different result. A skill encodes the judgment of whoever wrote it, so when the author is the person who knows the tools best, the company scales that person’s judgment. Ken’s own example was close to home, since he would love to capture the tribal knowledge at a former employer we share, where plenty of veterans are approaching retirement age.

Ken wants capture to work as a pull. When a system asks the veteran targeted questions, the veteran never has to sit down and author a skill, and the intuition comes out anyway. He sees an upside for the expert, since it’s a chance to document how you think about problems. “Being able to hand that intuition over to other people, I think, is a superpower,” he told me.

Your harness already has your memories

I asked Ken whether the experts that Clarifeye’s assistant interviews push back. He sees mild resistance early on, and it tends to fade once people understand what their AI tools already do behind the scenes. Every time you work with an LLM inside a harness such as Claude or ChatGPT, the harness builds memories, and Ken argues that those memories amount to your judgment.

In Ken’s telling, Clarifeye differs by doing that capture in the open and handing the artifacts back to the customer as its own intellectual property (IP). He adds that many of Clarifeye’s customers get higher-quality answers after moving to a smaller model and using fewer tokens. I felt the cost of not owning those artifacts this summer. When Claude went down, I found out how much of my business lived inside a single harness.

“So a company now has the ability to store their own IP and move between harnesses.”

— Ken Sanford, Go-to-Market and Commercial Strategy, Clarifeye

Ken’s best example of keeping those artifacts current involves a word I’ve banned from my own writing. He was using Claude with one of Clarifeye’s internal knowledge stores to draft blog content, and it kept using the word “ship,” which drives him nuts. He told Claude never to do that again. Claude then sent a signal back to the knowledge store, and the next time he opened Clarifeye, Clara asked him to confirm the ban. He said yes, and she corrected every artifact that used the word. That made me laugh, because “ship” sits on my own banned list for the same reason.

Where the next experts come from

If an agent can run on a senior expert’s encoded judgment, where does the next generation of experts come from? Ken expects AI to speed up the climb onto the shoulders of giants and to produce ultra-specialized experts who waste less time rehashing work that has already been done. Curiosity matters, and when he hires, he looks for agency, the people who try things and ask forgiveness rather than permission.

Ken applies the same test to companies. He still talks to organizations stuck behind a VPN with nothing but Copilot, and his advice to the people inside them is blunt: “Just quit your company.” His own team is 12 people, and he believes it “could grow 50-fold and not hire a single person.”

Here’s what I took from the conversation. Most of the work on the AI context layer goes into the data, the definitions, and the retrieval, while the judgment that decides what the answer should be still lives in a few people’s heads. Get it out while those people are still around, and make sure the skills your AI runs on come from the people who earned that judgment.

Listen to the full conversation with Ken Sanford on his Data Faces Podcast episode page.

Based on insights from Ken Sanford, who leads go-to-market and commercial strategy at Clarifeye, featured on the Data Faces Podcast.

Share

Leave a comment

Podcast highlights

- [0:05] Introduction and welcome to the Data Faces Podcast

- [1:26] Clarifeye and capturing unwritten knowledge for AI

- [3:57] Iron Man 3 extras and delivering real estate listing books in Naples

- [5:58] Tribal knowledge, CDOIQ, and the missing focus on unstructured data

- [6:46] Near-perfect retrieval and the judgment that was never written down

- [9:26] Voice mode, five minutes a day, and context-aware interviews

- [12:12] Do experts resist? Harness memories as your judgment

- [13:48] Where customers start: onboarding and software migrations

- [16:36] Keeping knowledge current, and banning the word “ship”

- [19:50] Why the wrong people are building your AI skills

- [22:53] Where the next generation of experts comes from

- [26:16] Hiring for agency, and why Ken says “just quit”

- [29:22] Why process blueprints beat a vector store

- [31:40] End of days? Ken’s techno-optimist case

- [34:58] West of Jesus and the science of flow

- [37:15] Where to find Ken

Frequently asked questions

What is tribal knowledge in AI?

Tribal knowledge is the judgment that experts carry in their heads and never write down, such as how to handle an exception or which rule wins when two policies conflict. For AI, it is the context that retrieval-augmented generation (RAG) cannot supply, because it exists in no document. Ken Sanford of Clarifeye calls it the implicit half of a company’s knowledge and argues that capturing it is the last step to AI making decisions the way a company’s experts would.

Who should build AI skills inside a company?

The people with the deepest domain judgment should shape a company’s AI skills. Ken Sanford points out that skills usually come from the most tech-forward employees, who rarely have 25 or 30 years of intuition. Because a skill encodes the judgment of whoever wrote it, he recommends making knowledge capture a pull, where a system interviews the veteran so the expert shapes the skill without having to write it.

How do you capture tribal knowledge from experts?

Interview them about exceptions. Clarifeye’s assistant, Clara, reads the existing documentation first, then asks experts short questions about edge cases and hypotheticals for five or ten minutes a day. When two experts disagree, a knowledge owner decides which answer stands. The results become readable process blueprints, versioned business rules, and reasoning workflows, which Ken Sanford says large language models follow more reliably than loosely retrieved text.

Can AI skills and knowledge move between AI harnesses?

They can when the company owns the artifacts. Ken Sanford argues that every AI harness, the tool that runs the model, such as Claude or ChatGPT, already builds memories that amount to a user’s judgment. Clarifeye captures that knowledge in the open and returns it to the customer as intellectual property, so a company can switch harnesses or models when it chooses. Ken adds that many customers get higher-quality answers after moving to a smaller model.

About David Sweenor

David Sweenor is the founder and host of the Data Faces podcast, where he talks with the people who are making data, analytics, AI, and marketing work in the real world. He is also the founder of TinyTechGuides and a recognized top 10 AI thought leader and international speaker who specializes in practical business applications of artificial intelligence and advanced analytics.

With over 25 years of hands-on experience implementing AI and analytics solutions, David has supported organizations including Alation, Alteryx, TIBCO, SAS, IBM, Dell, and Quest. His work spans marketing leadership, analytics implementation, and specialized expertise in AI, machine learning, data science, IoT, and business intelligence. David holds several patents and consistently delivers insights that bridge technical capabilities with business value.

Books

- Artificial Intelligence: An Executive Guide to Make AI Work for Your Business

- Generative AI Business Applications: An Executive Guide with Real-Life Examples and Case Studies

- The Generative AI Practitioner’s Guide: How to Apply LLM Patterns for Enterprise Applications

- The CIO’s Guide to Adopting Generative AI: Five Keys to Success

- Modern B2B Marketing: A Practitioner’s Guide to Marketing Excellence

- The PMM’s Prompt Playbook: Mastering Generative AI for B2B Marketing Success

Follow David on Twitter @DavidSweenor and connect with him on LinkedIn.

Despite all the hoopla about pacing the frontier, LLMs keep evolving at breakneck speed. This got me thinking about the prompt workflows I previously created, so a few days ago I ran an experiment that I fully expected to win. I opened a fresh Claude session with no project files, no memory, and none of my skills, and I typed a single sentence: “Write me a prompt workflow for building a competitive battlecard for B2B sales.” Then I compared it with the Strategic Battlecard Prompt Workflow I published on Substack in July 2025, one of the roughly one hundred prompt workflows I have written over the past couple of years.

Claude’s version was better. It made win-loss notes and call recordings the core input, and it added landmine questions that expose a competitor’s weak spots without naming the competitor. It also ended with a list of claims reps should never say and a red-team review written from the perspective of the competitor’s VP of Sales. My 2025 version had none of those, and I copied two of them into my own battlecard skill that same afternoon. My battlecard was designed to be printed and fit on one page (back-to-back). Claude’s workflow asked for one page too, but in practice that’s one of the downfalls of LLMs: they generate verbose, copious amounts of text that sales reps can’t consume. In fact, no one can, because the outputs are too long.

Having been in this business for a while, I could see it was missing a critical ingredient. What was that, you ask? Well, it was the notion that key messaging or field validation from reps outranks everything the research says. If a rep loses credibility because you created an erroneous battlecard, they stop trusting it and stop trusting you. So after a hundred workflows, here’s what I’ve learned.

Prompts are cheap now, and they get better with every model release, but a library like mine still maps which GTM jobs feed which. Customer interviews feed the messaging framework, for example, and the messaging framework feeds both the launch plan and the cheat sheet a rep keeps open on a call. The library also holds rules you only pick up by doing the work, and the rule about field validation is one of them. AI GTM engineering is the work of turning those connections into systems that run.

A fresh model writes a better prompt than I did last year

Here’s how they compared, row by row, with the battlecard skill I run today added as a third column. The fresh model had nothing to work from except that one sentence, so you can paste it into any new chat window and repeat this cold-prompt test yourself in about two minutes. Rows marked “added after this test” are the two steps I copied from the model that afternoon.

The fresh model matched or beat my 2025 workflow on each of the top six rows. The bottom four rows tell a different story, because each one depends on something a chat window can’t see, such as the win-loss files, the deals in the pipeline, and the corrections reps have made along the way. Each one came out of building this process for clients.

In fairness, I pitted a 2026 model against a prompt workflow I wrote with 2025 tools, so the comparison says as much about how far the models have come as it does about my prompt. That’s exactly why it matters to anyone sitting on a prompt library, and I’m one of them. The step-by-step instructions improve on their own every time a new model comes out, so a folder of well-written prompts stops being much of an advantage, no matter how many hours went into it.

What the model couldn’t know

The rule that field observations outrank the research came from a battlecard system I built for a client earlier this year. Alongside the battlecards, I created a field-intelligence log, a running file of corrections from the client’s sales reps and product team. Every entry records the claim, who made the correction and when, how confident they are, and which assets have to change. When an entry conflicts with a competitor’s documentation, an analyst report, or my own research, the field-intelligence entry wins.

One correction shows how the log works in practice. The product team pointed out that the client’s materials listed four deployment options when there were only three, because one of the four was a product feature that had been mistaken for a way to deploy the software. That single entry had to reach every battlecard, feature matrix, and web page that mentioned deployment, and the log tracked which of them were still waiting on a fix.

“Field validation from reps outranks everything the research says. If a rep loses credibility because of an erroneous battlecard, they stop trusting it, and they stop trusting you.”

— David Sweenor, Founder/CEO, TinyTechGuides

Nobody writes a rule like that into a prompt until they have seen what happens without it. At another client, the person who built a battlecard against a large platform vendor told the team he checked every claim to make sure none overstated anything. He was happy to stand behind those claims on a call. He worried about anything stronger, such as claiming the competitor lacked a capability it had announced only two weeks earlier. The prospect’s own engineers often know the competitor’s product better than the seller does, and when they catch an overstated claim like that, it’s an argument a rep can’t win.

When a battlecard gets something wrong in front of a customer, fugettabout it. Reps go back to asking the sales engineer down the hall, and the deals the card should have helped win get lost for reasons that never appear in a pipeline report. For a CMO, that’s the real cost of a battlecard nobody maintains, and no model can see it from inside a chat window.

From a skill to a system

A Claude skill is one packaged job that you trigger with a slash command, and on its own, it waits for somebody to remember it exists. My battlecard skill is more useful than that because it depends on other jobs. It won’t build a card until my site monitor, a market research report, or a win-loss analysis has produced something about that competitor, and it reads the log of rep and product corrections last, so those corrections win.

Dependencies like that turn a prompt library into something more valuable than a set of prompts. Read my hundred workflows as a group, and they sketch a map of GTM work in which competitive intelligence feeds the battlecard, the battlecard feeds the rep, and the rep’s wins and losses feed the next round of competitive intelligence. When I built the B2B PMM Field Kit, the workflows sorted into five regions of that map: foundations, intelligence, competitive, launch and content, and the buying committee. Tool vendors tend to draw the map from what their own products touch, so it rarely matches the one a practitioner draws from the work.

A system also needs to know when to run. A battlecard goes out of date for predictable reasons, and my skill now watches for four of them. Two come from the market: a competitor pricing change caught by the site monitor and a new win-loss analysis, and two come from the field: a major deal won or lost and a high-confidence correction that names the card. Each is an event a system can act on without anyone’s calendar reminder, so the battlecard passes the test I set in What AI GTM engineering is: whether something you built still works the week you’re on holiday.

Three questions to ask of your own prompt library

Whether your prompt library is a shared Google Doc or a hundred Substack posts, three questions will tell you what it’s worth now. None of them takes more than an afternoon to answer, and together they show which workflows are ready to become part of a system.

1. Would a fresh model write it just as well? Run the cold-prompt test on your best workflow by asking a new chat session, with no context, for the same thing. If the model ties or wins, the prompt is your starting point rather than your advantage, so borrow whatever the model added and move on to the next two questions.

2. What does it read, and where does its output go? Name the inputs a workflow depends on and the place its output ends up. A workflow with neither is still a prompt someone has to remember, and connecting it to the jobs on either side makes it part of a system.

3. What should make it run again? Write down the events that would make its output wrong, such as a pricing change, a lost deal, or a correction from the field. Those events become the triggers, and the corrections become rules the system enforces every time it runs.

Over the next few posts, I’ll build these systems one at a time, starting with competitive intelligence, and show what each one reads, what it hands off, and what makes it run. For anyone who’d rather not wait, I’m happy to spend thirty minutes looking at your prompt library with you and picking the workflow most worth turning into a system first. A fresh model will keep writing better prompts every year, and it still won’t know which of your battlecards a rep stopped trusting last quarter.

Interested in learning more? Book a consultation.

Share

Leave a comment

Frequently asked questions

What is the cold-prompt test for prompt workflows?

The cold-prompt test compares your prompt workflow with one a fresh AI model writes with no project files or memory, given a single sentence of instruction, such as “Write me a prompt workflow for building a competitive battlecard for B2B sales.” When David Sweenor ran it in September 2026, a fresh Claude session wrote a competitive battlecard workflow that beat his July 2025 published version. The model’s version added win-loss inputs, landmine discovery questions, a list of claims reps should never say, and a red-team review. The test takes about two minutes and shows whether a prompt is still an advantage.

Are prompt workflow libraries still worth anything now that AI models write good prompts?

A prompt workflow library is worth less for its prompts and more for what it records about the work. A library of about a hundred B2B marketing workflows sketches a map of which go-to-market (GTM) jobs feed which, such as competitive intelligence feeding the battlecard and win-loss results feeding the next round of intelligence. It also holds rules learned from doing the work. AI GTM engineering uses that map and those rules as the raw material for systems that run.

What is the difference between a Claude skill and a GTM system?

A Claude skill is one packaged job that you trigger with a slash command, and it waits for someone to call it. A GTM system connects several skills so one job’s output becomes the next job’s input, and it runs when specific events happen. A battlecard skill becomes part of a system when it waits for site monitoring, market research, or win-loss analysis to produce something about a competitor, applies field corrections last, and rebuilds when a pricing change or major deal outcome makes the card wrong.

What is a field-intelligence log for competitive battlecards?

A field-intelligence log is a running file of corrections from sales reps and the product team that outranks a competitor’s documentation, analyst reports, and internal research. Each entry records the claim, who made the correction and when, how confident they are, and which sales assets have to change. When a product team corrects a claim, such as the number of deployment options a product supports, the log tracks the fix across every battlecard, feature matrix, and web page that mentions it.

When should a competitive battlecard be refreshed?

Refresh a competitive battlecard whenever an event makes it wrong; four events cover most cases. Two come from the market: a competitor pricing change caught by site monitoring and a new win-loss analysis, and two come from the field: a major deal won or lost against that competitor and a high-confidence correction from a rep or the product team. A GTM system can watch for all four and rebuild the card without anyone reminding it.

Where should a marketing team start turning prompt workflows into GTM systems?

Start by asking three questions of your best prompt workflow. First, would a fresh AI model write it just as well? If so, treat the prompt as a starting point. Second, what does the workflow read, and where does it send its output? Third, what events should make it run again? Workflows with clear inputs, clear destinations, and clear triggers, such as competitive battlecards, are the best first candidates for an AI GTM engineering system.

About David Sweenor

David Sweenor is the founder and host of the Data Faces podcast, where he talks with the people who are making data, analytics, AI, and marketing work in the real world. He is also the founder of TinyTechGuides and a recognized top 10 AI thought leader and international speaker who specializes in practical business applications of artificial intelligence and advanced analytics.

With over 25 years of hands-on experience implementing AI and analytics solutions, David has supported organizations including Alation, Alteryx, TIBCO, SAS, IBM, Dell, and Quest. His work spans marketing leadership, analytics implementation, and specialized expertise in AI, machine learning, data science, IoT, and business intelligence. David holds several patents and consistently delivers insights that bridge technical capabilities with business value.

Books

- Artificial Intelligence: An Executive Guide to Make AI Work for Your Business

- Generative AI Business Applications: An Executive Guide with Real-Life Examples and Case Studies

- The Generative AI Practitioner’s Guide: How to Apply LLM Patterns for Enterprise Applications

- The CIO’s Guide to Adopting Generative AI: Five Keys to Success

- Modern B2B Marketing: A Practitioner’s Guide to Marketing Excellence

- The PMM’s Prompt Playbook: Mastering Generative AI for B2B Marketing Success

Follow David on Twitter @DavidSweenor and connect with him on LinkedIn.

▶ Watch the full episode and read the transcript: No Business Value From AI? Shut It Down | Randy Bean

Randy Bean has been asked where the data function should sit for 45 years, and he has an answer that makes technology leaders wince. Put the chief data officer under the chief sales officer. He tells it as a joke on stage, and the audience usually reacts the way you just did, but the reasoning underneath it is serious. The chief sales officer owns revenue and customer responsibility, and in Randy’s view, those two things are the whole reason a data function exists.

I caught up with Randy on location at the CDOIQ Symposium in Cambridge, Massachusetts, a few hours after he moderated the chief data officer panel that he has organized for the event for 12 years running.^[1] He was a guest on the show back in January, when we talked about why culture eats AI for breakfast, so this was a mid-year check on the role he has watched longer than almost anyone. A few of his lines from this conversation made it into my field notes on the CDO role at 20 years, and this is the fuller version.

“If you’re not getting measurable business value or have a clear path to measurable business value from your data and AI investments, I suggest you go back to the office this afternoon and shut them down.”

— Randy Bean, Founder and CEO, Data & AI Leadership Exchange

Companies are asking the same questions that they asked when he started his career, and he means that as an observation rather than a complaint. So, why is the reporting-line debate still unresolved after two decades of chief data officers? I think it’s because the debate is really about what the function is for, and Randy has an unusually clear answer.

About Randy Bean

Randy Bean is the founder and CEO of the Data & AI Leadership Exchange and one of the longest-serving chroniclers of the chief data officer role. He founded the consultancy NewVantage Partners, wrote Fail Fast, Learn Faster: Lessons in Data-Driven Leadership in an Age of Disruption, Big Data, and AI, and publishes the annual AI & Data Leadership Executive Benchmark Survey, now in its 15th year.^[2][3] He has organized the CDOIQ Symposium’s chief data officer panel for 12 years, which gives him a front-row view of how the job has changed.

He was trained as a COBOL and assembler programmer, and he told me that he wanted to be a poet or a rock star before he found that he was more interested in data and how organizations could use it to make better decisions. Four and a half decades later, that interest is still what he does for a living.

In this episode, Randy and I discuss:

- The business-value test that he uses to provoke data leaders, and what to do if your investments fail it

- Why technology transformations only get serious when an organization has an existential reason to change

- Where chief data officers actually report today, and why he argues for the business side anyway

- Why every chief data officer from last year’s CDOIQ panel will be out of the role by the end of this year

- His mid-year read on the AI bubble, overestimated in the short term and underestimated in the long term

Watch the full conversation here:

The same questions, 45 years later

I asked Randy for a letter grade on the industry’s data quality, since we have been talking about it for as long as I have been in the business. He declined to grade it, and the reason is instructive. You can spend all the time in the world on data quality, but the only measure that he cares about is the business value that investments in data, analytics, and AI deliver. That’s why he tells organizations to shut down anything without measurable value or a clear path to it, and he admits that he says it partly to provoke a conversation.

The value question has only gotten more pointed since AI arrived. This year’s panel featured chief data officers from Lowe’s, Ford, and Humana, and Randy traced the role’s arc from a defensive mandate built on risk, regulation, and compliance toward an offensive, business-focused one. Then AI showed up over the past four years and changed more than the focus. It changed how organizations structure data management as an asset and where AI fits in that picture. Plenty of data leaders have a value story for their AI work. How many could defend it if the CFO asked them to shut down everything that couldn’t?

Transformation waits for an existential reason

When I asked whether companies are using AI for productivity or to reinvent their business, Randy said they are all over the place, then explained why most stay there. Technology transformations take time because they involve organizational change, people change, and changes to roles and processes. During the pandemic, the chief digital officer of the nation’s largest insurer told him that the company had done more to execute on its digital transformation in six months than in the previous 20 years.

“It’s often not until there’s an existential requirement to do something that organizations really fundamentally undergo the transformation that they need to do. Otherwise, it’s experimentation, lip service, carving off pieces that are low-hanging fruit.”

— Randy Bean, Founder and CEO, Data & AI Leadership Exchange

I recognized all three from my own career. Experimentation looks like progress on a slide, lip service satisfies the leadership offsite, and low-hanging fruit produces a case study that never scales. None of them touches how the company operates. Randy’s point is that AI should be an occasion to reinvent the business from top to bottom, and that most companies won’t do it until something forces them to.

Where the CDO reports, and where Randy would put it

Randy asks about the reporting line in his benchmark survey every year, and this year’s numbers are a three-way split. Thirty-three percent of chief data officers report into business leadership, 42 percent report into technology leadership, and another 23 percent report somewhere else entirely.^[4] Technology is the plurality, and it isn’t where he would put the job.

“The closer you are to the end business and the end customer, the more value you can deliver when the day’s done.”

— Randy Bean, Founder and CEO, Data & AI Leadership Exchange

He argues that technology, data, and every function like them exist to serve the business and its customers, and if they don’t improve the customer experience, then why are they there? He carefully adds that technology leaders can be business leaders, so this isn’t a swipe at CIOs. It’s a claim about proximity, and it’s where the chief sales officer joke comes from.

That joke has a backstory. As a young programmer, Randy listened to his colleagues complain that the business people could never articulate their requirements, until he finally told them to stop, because the business people were the reason any of them had a job. I sat on the IT side of that conversation for the first half of my career, and the complaint was just as common at IBM in 1999 as it was for him in the early 1980s. My own reason for agreeing with Randy is a little different. IT is built to be risk-averse, with processes and controls that make sense for systems of record, and models, analytics, and AI don’t work that way. They fail, and a failed model should be a normal part of the process, not an incident. The business side can tolerate that, and IT is structurally set up to treat it as a problem.

A role being rewritten in real time

Last year’s CDOIQ panel included Teresa Heitsenrether, the chief data and analytics officer at JPMorgan Chase, and Randy held her up as a model for what the role can be. She sat on the bank’s operating committee, reporting to Jamie Dimon, which put data and AI at the top table of the largest bank in the country. She is retiring at the end of this year, and her responsibilities will move back under the technology organization.^[5] Randy’s larger point was about the panel as a whole. A year ago, he had four chief data officers on stage, and by the end of this year, none of them will be in those roles. Three have already gone.

“There’s a battle, a fight going on to control AI and shape the AI future of organizations, and data is a component of that. I think the next year to two years will be very interesting, and I think the role could look very different in a way that I don’t really imagine right now.”

— Randy Bean, Founder and CEO, Data & AI Leadership Exchange

When we spoke in January, Randy was advocating for the combined chief data and AI officer role, and I wanted to know how that debate had moved in six months. He said things are changing in real time, and he declined to predict where the role will end up. That restraint is notable coming from someone who has watched the job for 12 years of panels. When the person with the longest view says that he can’t picture the role two years out, the safe assumption for anyone in the seat is that the job description is being written by whoever wins the fight over AI.

Overestimated now, underestimated later

I closed with the question that everyone asks at conferences this year. Is there an AI bubble? Randy answers that AI is real and inevitable, but its expectations and valuations are overinflated. He sums it all up in one line. AI is overestimated in the short term and underestimated in the long term. Its transformational effect on society, on jobs, and on how the world operates will exceed what most people can imagine today.

I told him my own version of the short-term half. The promised 100x productivity mostly looks like work being moved around: someone hands you AI-generated content to review, and the labor ends up on your desk rather than theirs. Randy agreed that the productivity story has been more personal than corporate so far, and then he widened the lens. Technology is a tool, and every transformational technology in history has been disruptive in positive and negative ways. He sees enormous upside in health care and human longevity, and real harm in kids glued to AI and phones. He was candid that after a movie like Oppenheimer, the jury is still out on whether the benefits of a technology this powerful outweigh the downside.

If you take one thing from Randy’s 45 years, take the test, not the joke. Everything in your data and AI portfolio should have measurable business value or a clear path to it, and the reporting line that gets you closest to the customer is the one most likely to keep it that way.

Listen to the full conversation with Randy Bean on his Data Faces Podcast episode page.

Based on insights from Randy Bean, Founder and CEO of the Data & AI Leadership Exchange, featured on the Data Faces Podcast.

Share

Leave a comment

Podcast highlights

- [0:07] David welcomes Randy Bean back to Data Faces, on location at the CDOIQ Symposium in Cambridge, Massachusetts

- [0:50] Randy on wanting to be a poet or a rock star, and 45 years of caring about the data instead

- [2:07] Randy on 12 years of CDOIQ panels, this year’s Lowe’s, Ford, and Humana lineup, and the role’s move from defensive to offensive

- [3:47] The business-value test, and why he tells data leaders to shut down what doesn’t deliver

- [4:57] Why transformation waits for an existential requirement, and the insurer that did 20 years of work in six months

- [6:33] Panels from A to Z, and Teresa Heitsenrether on JPMorgan’s operating committee

- [7:50] The survey split, 33 percent business, 42 percent technology, and 23 percent elsewhere, and why Randy argues for the business

- [9:21] The chief sales officer joke, and the young programmer who told his colleagues to stop blaming the business

- [10:56] Every chief data officer from last year’s panel will be gone by year end, and the fight to control AI

- [12:07] Overestimated in the short term, underestimated in the long term

- [13:21] Health care, human longevity, kids and phones, and the Oppenheimer question

About David Sweenor

David Sweenor is the founder of TinyTechGuides and host of the Data Faces Podcast. He is an international speaker, advisor, and the author of eleven books on artificial intelligence, analytics, and B2B marketing, including Generative AI Business Applications, The CIO’s Guide to Adopting Generative AI, and Modern B2B Marketing. With more than twenty-five years in analytics and AI at companies including Alteryx, Tableau, TIBCO, SAS, IBM, and Dell, David advises technology companies on product marketing, content strategy, and go-to-market execution. He holds several patents and has been named a top influencer in data and analytics by Onalytica, Thinkers360, and Analytics Insight.

Connect with David on LinkedIn and subscribe to the Data Faces Podcast for conversations with the people shaping enterprise data and AI.

Footnotes

^[1]CDOIQ Symposium. “The 20th Annual CDOIQ Symposium.” July 21–23, 2026, Hyatt Regency Cambridge, Massachusetts.

https://2026cdoiq.org/

^[2]Bean, Randy. Fail Fast, Learn Faster: Lessons in Data-Driven Leadership in an Age of Disruption, Big Data, and AI. Wiley, 2021. https://www.wiley.com/Fail+Fast%2C+Learn+Faster%3A+Lessons+in+Data-Driven+Leadership+in+an+Age+of+Disruption%2C+Big+Data%2C+and+AI-p-00310033

^[3]Data & AI Leadership Exchange. “2026 AI & Data Leadership Executive Benchmark Survey.” https://www.dataaiex.com/leadership

^[4]Data & AI Leadership Exchange. “2026 AI & Data Leadership Executive Benchmark Survey.” https://www.dataaiex.com/leadership

^[5]Bloomberg. “JPMorgan AI Chief, Longtime Top Trader Exits After Four Decades.” July 1, 2026. https://www.bloomberg.com/news/articles/2026-07-01/jpmorgan-ai-chief-longtime-top-trader-exits-after-four-decades

On September 18, I joined Eric Kavanagh and David Linthicum on Inside Analysis for an episode Eric dubbed the “AI Slowdown Showdown.” Every major outlet that week was asking whether the frontier AI labs would slow down, and whether they’d do it for our safety or their bottom line. Eric opened the show with fists swinging. “These models are amazing,” he said. “You could talk all day about the amazing things we know they do without having to manufacture these storylines like rogue agents.”

During the first break, I made the point that the slowdown everyone was debating has already happened inside companies, where most generative AI projects stalled long before anyone talked about pumping the brakes. So, if you lead a data or AI program, which slowdown should worry you more? The full episode is below, and here’s how I’d answer that question.

Would an AI slowdown agreement be a cartel?

The debate started with Dario Amodei’s essay “We Must Pace the Frontier,” published on September 12, which called for common safety benchmarks and agreed limits on how fast AI capabilities can grow. Sam Altman, Elon Musk, and Demis Hassabis each voiced their agreement that same day. I read the essay as a repositioning play, since Anthropic already owns the “safe” position in the public’s mind, and I doubt any of it is altruistic in any way, shape, or form. Then I got a laugh out of the panel by pointing out that when three companies restrict output, “I think there’s a word for that. Maybe we’ll call it cartel.”

Eric agreed that it’s “essentially collusion” if the labs are coordinating behind the scenes. Linthicum was less sure about collusion, but he thought the labs were trying to “manipulate the market in their direction. And guess what? It may work.”

I didn’t expect a federal judge to get the question the same day the episode aired. Four paying AI subscribers filed a proposed class action in federal court in California, Buist v. Anthropic PBC, accusing Anthropic, OpenAI, SpaceXAI, and Google of an illegal agreement to slow AI development under Section 1 of the Sherman Act.^[1] The complaint argues that agreeing to progress “slower than competition would otherwise produce” harms paying subscribers.^[2] A complaint is only an allegation, and the frontier might speed up, slow down, or spend the next two years in discovery. Whatever happens, the stalled projects inside your company are really what you need to worry about.

The AI slowdown nobody voted on

When I look at enterprise AI adoption, the brakes went on a while ago. PwC’s 29th Global CEO Survey of 4,454 chief executives in 95 countries and territories found that 56% had seen neither higher revenues nor lower costs from AI, and only 12% reported both.^[3] Those numbers shape the budget conversation for every data and AI leader I know, far more than anything a lab announces.

“I think the slowdown’s already happened. If you look at enterprise adoption, companies aren’t getting value from it. Nobody’s proposed it, nobody’s voted on it, nobody’s signed anything.”

— David Sweenor, Founder/CEO, TinyTechGuides

Linthicum said enterprises are “not doing core applications. They’re not doing inventory control.” They’re using generative AI for productivity work, such as managing a calendar. That work rarely touches the order-to-cash process where a CEO looks for a return, and the split between personal and core use also helps explain how the 3% of companies that win with AI differ from the rest.

On the show, I compared the whole debate to limiting the delivery truck when the warehouse is the bottleneck. Most companies haven’t yet organized the warehouse, which holds the data, processes, and decisions that a model is supposed to plug into. When the shelves are a mess, a faster or slower truck won’t change how many orders go out the door.

What part of your business can be wrong 5% of the time?

Eric asked a question that every CIO should tape to their monitor. Say your model is 80% accurate, you add retrieval-augmented generation (RAG) to reach 90%, and with a lot of work, you push it to 95%. At that point, the question becomes where your business can live with the remaining errors.

“Let’s say it’s 95%. What part of your business can be wrong 5% of the time? Fulfillment, no. Accounting, no.”

— Eric Kavanagh, CEO, The Bloor Group

I offered marketing coupons as a paradigm of hope, and Eric acquiesced on that point. Accuracy is only half the problem with putting a frontier large language model (LLM) into a core business system. Linthicum raised the other half, which is traceability. Core business systems get audited, and an answer that nobody can explain won’t survive a compliance review, however often it happens to be right.

“Every time I’ve tried to look into how I got a bad answer within an LLM, I can’t do it. And guess what? That means I can’t use it for business systems.”

— David Linthicum, Linthicum Research

Hallucination is a defect, and AI agents don’t go rogue

Part of the reason enterprises reach for the wrong fixes is the vocabulary our industry chose for AI. On the show, I asked what would happen if we’d called a hallucination a defect or a bug from the start, since your engineering team would think about it differently. The word “agent” works the same way, because it suggests something with intentions of its own. As I put it on the show, “We’ve given all these humanizing terms, but they’re just computers, computer code doing stuff.”

Eric made the same case about the “rogue agent” stories. In the incidents he described, someone deployed an agent in an environment with open ports, and it followed its objective function with no intent behind it. It didn’t break any rules, because it never decided anything.

“No agent goes rogue. To go rogue, you have to be sentient and decide to go against the rules.”

— Eric Kavanagh, CEO, The Bloor Group

The words matter because each one points a team toward a different fix. A defect gets an owner, a root cause, and a regression test. When teams call the same error a hallucination, the usual response is a shrug and a rewritten prompt, which is why I argued last year that teams should stop chasing hallucinations and focus on agentic quality.

Human in the loop is a process you haven’t designed yet

When Eric asked for final thoughts, I went after the most comfortable phrase in enterprise AI governance. When a team isn’t sure what an AI system will do, the default answer is to put a human in the loop. I called it a cop-out on the show.

“Let’s have a human in the loop. That means we haven’t figured out the process yet. And if you’re making decisions at scale, a human in the loop doesn’t work.”

— David Sweenor, Founder/CEO, TinyTechGuides

At thousands of decisions a day, the person reviewing each one either becomes the bottleneck or starts approving things without reading them. Many of the data and AI leaders I interview on the Data Faces podcast describe a version of this problem, too. Everybody wants guardrails on these probabilistic machines, and the guardrails keep getting narrower until I have to ask why the team didn’t write deterministic code in the first place. That tension showed up in the lessons I took from 27 data leaders in 2025, and nobody has resolved it yet.

Eric described a better way to build these systems. “You have to couch them into a workflow,” he said, calling the model for the one step that needs it “and then funnel that back into your deterministic workflow.” Linthicum went further for core systems.

“The core model is going to be hosted by us. We’re going to run it. We’ll build it using our data. It’s going to have a very narrow focus.”

— David Linthicum, Linthicum Research

Both describe a designed process that settles three things before anything goes live. It identifies the single step where a probabilistic model earns its place and wraps that step in deterministic checks that catch out-of-range outputs. It also logs each model call in enough detail to trace a bad answer back to its source. With those in place, the person in the loop has a specific decision to make, at a specific point, with the evidence in front of them.

Where to start with enterprise AI adoption

If your AI program feels stalled, designing the process will move it faster than waiting for a better model. Here are three places I’d start.

1. Name the decision behind every human in the loop: Wherever a person sits in the loop in your AI plans, write down the specific decision they make and the information they need to make it. When you can’t, designing that process is your first project.

2. Mark the one probabilistic step in each workflow: Identify where the model does work that ordinary code can’t, and put a deterministic check on either side of it.

3. Log model errors as defects: Give every bad output an owner, a root cause, and a test, and you’ll build the audit trail your compliance team will ask for.

Fix the warehouse first

When Eric asked why the AI companies are behaving the way they are, I went back to a business school professor who liked to say it always comes down to the fundamental economics of the situation. The frontier labs will speed up or slow down based on what their investors tell them, and none of us gets a vote on that. What happens inside your own company is a different story, because the warehouse belongs to you, and so does the decision to fix it first.

If you want to work through these ideas in person, join Eric, me, and a room full of practitioners at the Agentic Road Show on Thursday, October 22, at the University of Pittsburgh. Dr. Mark Brady, former chief data officer of the U.S. Space Force, will present his six laws for safe AI, and Carnegie Mellon’s Philip Koopman will close the day on autonomous vehicle safety. Workflow design is most of what I do in my advisory work, and it’s where I see stalled programs start moving again, so I invite you to join us in Pittsburgh.

Share

Leave a comment

Frequently asked questions

Is enterprise AI adoption slowing down?

In most companies, enterprise AI adoption slowed well before the frontier labs talked about slowing down. PwC’s 29th Global CEO Survey, published in January 2026, found that 56% of 4,454 chief executives had seen neither higher revenues nor lower costs from AI, and only 12% reported both. Most enterprise use of generative AI sits in personal productivity tasks, such as email and calendars. It has not reached core systems like inventory control or order entry, which is where CEOs look for a return.

Are AI companies being sued over the AI slowdown?

Yes. On September 18, 2026, four paying subscribers to ChatGPT, Claude, Grok, and Gemini filed a proposed class action, Buist v. Anthropic PBC, in the U.S. District Court for the Northern District of California. The complaint accuses Anthropic, OpenAI, SpaceXAI, and Google of an illegal agreement to slow AI development in violation of Section 1 of the Sherman Act, citing executives’ public agreement with Dario Amodei’s September 12 essay, “We Must Pace the Frontier.” The plaintiffs argue that coordinated limits on AI progress reduce the value of paid subscriptions, and the claims have not been tested in court.

What does human in the loop mean in AI, and why isn’t it enough?

Human in the loop means a person reviews or approves an AI system’s output before it takes effect. It becomes a problem when teams use it in place of process design, because it usually means nobody has decided what the person is checking or why. When a system makes thousands of decisions a day, the reviewer either becomes a bottleneck or approves outputs without reading them. A designed process names the specific decision the person makes, at a specific point, with the evidence in front of them.

Is an AI hallucination a bug?

For enterprise purposes, treat an AI hallucination as a defect. Calling a wrong output a hallucination makes it sound like an unavoidable quirk of the model, and teams tend to respond by rewording the prompt. Treating it as a defect gives the error an owner, a root cause, and a regression test, the same as any other software bug. That discipline also builds the audit trail that compliance teams need for AI systems that get reviewed.

Can AI agents go rogue?

AI agents don’t go rogue in the sense of deciding to break the rules, because that would require intent. As Eric Kavanagh put it on DM Radio’s Inside Analysis, “To go rogue, you have to be sentient and decide to go against the rules.” The incidents reported as rogue AI agents usually involve an agent deployed in an environment with open ports or weak containment, where it pursued its objective function as designed. Those incidents call for a security review of the deployment environment.

How accurate does AI need to be for enterprise workflows?

The required accuracy depends on what a wrong answer costs. Even at 95% accuracy, a large language model (LLM) is wrong one time in 20, which is unacceptable for fulfillment, accounting, and most other core processes, though tolerable for low-stakes tasks like marketing coupons. Enterprise AI systems also need traceability, since an answer that nobody can explain won’t survive an audit. Many teams handle both problems by calling a model for one narrow step and wrapping it in deterministic checks and logging.

Where should enterprise leaders start with a stalled AI program?

Start by designing the process around the model before waiting for a better model. Write down the specific decision behind every human in the loop in your AI plans, and mark the single step in each workflow where a probabilistic model does work that ordinary code can’t, with deterministic checks on either side of it. Then log model errors as defects, with an owner and a root cause, so the AI program builds an audit trail as it goes.

About David Sweenor

David Sweenor is the founder and host of the Data Faces podcast, where he talks with the people who are making data, analytics, AI, and marketing work in the real world. He is also the founder of TinyTechGuides and a recognized top 25 analytics thought leader and international speaker who specializes in practical business applications of artificial intelligence and advanced analytics.

With over 25 years of hands-on experience implementing AI and analytics solutions, David has supported organizations including Alation, Alteryx, TIBCO, SAS, IBM, Dell, and Quest. His work spans marketing leadership, analytics implementation, and specialized expertise in AI, machine learning, data science, IoT, and business intelligence. David holds several patents and consistently delivers insights that bridge technical capabilities with business value.

Books

- Artificial Intelligence: An Executive Guide to Make AI Work for Your Business

- Generative AI Business Applications: An Executive Guide with Real-Life Examples and Case Studies

- The Generative AI Practitioner’s Guide: How to Apply LLM Patterns for Enterprise Applications

- The CIO’s Guide to Adopting Generative AI: Five Keys to Success

- Modern B2B Marketing: A Practitioner’s Guide to Marketing Excellence

- The PMM’s Prompt Playbook: Mastering Generative AI for B2B Marketing Success

Follow David on Twitter @DavidSweenor and connect with him on LinkedIn.

Footnotes

^[1]Quinn Wilson. “OpenAI, Anthropic, Google, SpaceXAI Hit With Antitrust Lawsuit.” Bloomberg Law, September 18, 2026. https://news.bloomberglaw.com/litigation/openai-anthropic-google-spacexai-hit-with-antitrust-lawsuit. Case: Buist v. Anthropic PBC, No. 3:26-cv-10693 (N.D. Cal.).

^[2]Kaitlyn Huamani. “Lawsuit says Anthropic, OpenAI, SpaceXAI and Google made illegal agreement on AI slowdown.” Associated Press, via OPB, September 20, 2026. https://www.opb.org/article/2026/09/20/lawsuit-says-anthropic-openai-spacexai-and-google-made-illegal-agreement-on-ai-slowdown/.

^[3]PwC. “PwC’s 29th Global CEO Survey: Leading Through Uncertainty in the Age of AI.” PwC, January 19, 2026. https://www.pwc.com/gx/en/issues/c-suite-insights/ceo-survey.html.

Over the past couple of years, I’ve built more or less the same thing for five different clients, and not one of them called it AI GTM engineering. As each company got further along with AI, the work still showed up on the invoice as product marketing. The work was building AI systems into how a marketing team operates, which I have been writing about for a year.

So I went looking for what the rest of the market calls this. I found three perspectives on the same question, published within seven months of each other, pointing in different directions. Each of them had counted something real.

In September 2025, Kyle Poyar counted job postings for his Growth Unhinged newsletter and found 45 GTM engineer listings in a month, 128 over a quarter, and one for every 92 openings for a sales development rep.^[1] He noted the title barely registers in Google Trends. Poyar spent eight years at OpenView, so when he called the buzz well ahead of the hiring, he was right for the moment.

A month later, Henley Wing Chiu reached a blunter conclusion from a bigger sample. He co-founded BuzzSumo and now runs a job-postings index called Bloomberry, and he pushed a thousand GTM engineer postings through it. Nine out of ten responsibilities also turn up in RevOps postings, which led him to write that “GTM Engineering and RevOps jobs are essentially the same.”^[2]

Then in April 2026, Brendan Short came at it from the company side. He writes The Signal, and rather than counting job ads, he went through 63 of the fastest-growing SaaS companies. He found 54% had somebody doing the job, against 3% to 7% of private B2B SaaS overall.^[3]

So the category barely exists, already exists under an older name, or is standard equipment at the fastest-growing companies. They disagree because each is measured against a definition somebody else wrote, and nearly every definition in circulation comes from a company that sells the tools the job operates. Poyar and Short even drew on the same job-postings index and still reached opposite conclusions.

“AI GTM engineering builds AI systems into how a go-to-market team works. Write down the data and the voice first, then build workflows that run without anyone starting them, then hand the whole thing to your own team to operate.”

— David Sweenor, Founder/CEO, TinyTechGuides

That is the definition I have worked from on those five engagements, and the order is the point. Skip the first step and the other two inherit its absence, which is why so much AI output reads as though it came from a company with no particular opinion.

Every definition on offer was written by a vendor

Every GTM engineer definition I could find was published by a company that sells the tools the role operates. The cleanest comes from ZoomInfo, which describes a GTM engineer as someone who designs, builds, and operates the data pipelines, enrichment workflows, and automated plays behind how a company sells and markets.^[4] Clear thinking, from a company whose product is data pipelines and enrichment.

The job postings tell the same story. In Chiu’s sample, HubSpot appears in 52% of them, Outreach in 49%, and Salesforce in 45%, with Clay leading the list. Every tool named there either sells to a prospect or moves data between the systems that do.

Agencies sell the same picture. Poyar counted more than 120 agencies in Clay’s partner directory alone, and he passes along an estimate, his word rather than mine, that 45% of the people carrying a GTM engineer title work at agencies or consultancies. Most of those shops are built on the same enrichment and outbound tooling.

So the companies selling the software describe the role, and every description ends at outbound prospecting. That work is real, and somebody should own it. What none of it describes is marketing, by which I mean positioning and launches, competitive analysis and the content a buying committee reads, and the judgment about which of those is worth automating at all.

The three components, and why the order matters

AI GTM engineering starts underneath the marketing work, in what a model knows before it writes anything. The first of its three components covers the environment, the data, and your voice and editorial standards, written down where an AI model can read them. Teams skip this because it feels like preparation rather than progress, and everything built on top of it inherits that shortcut.

Workflows come second, and my test for whether one is finished is whether it runs without anyone starting it. A prompt somebody keeps in a document and pastes into a chat window on Tuesdays is a habit with extra steps. Something that wakes on its own, runs the competitive scan, and files the result where the team already looks still works the week its builder is on holiday.

Enablement comes third, and most engagements drop it. When you hand over a system nobody understands, you create a dependency rather than a capability. Ask who fixes it at 4 pm on a Friday, and if the answer is the consultant, you bought a subscription to a human being.

None of this belongs to any vendor. When an outage hit me mid-deadline earlier this month, switching models cost almost nothing, because the rules, the workflows, and the accumulated context were all files on a disk. Build this way, and you can change models or consultants without starting over.

The scarce skill is not the engineering

Here is where I part company with almost everyone. People hear the word engineering and assume the hard part is building the system, so they hire somebody who can wire tools together and write a little Python. That person is easy to find now, and the tools have gotten good enough that building became the cheap part.

Choosing what to build stayed hard. Knowing what a battlecard has to say is a different skill from knowing how to generate one. Understanding how a buying committee reaches a decision, which competitive claims a seller will use on a call, and which content exists only because somebody asked for it once, none of that comes from automation experience. It comes from more than two decades of doing the work and watching which parts of it mattered.

“Knowing which hours of a marketing week are worth taking back, and which ones would break something if you automated them, is the part that stayed expensive.”

— David Sweenor, Founder/CEO, TinyTechGuides

This explains why the two obvious places to buy each solve one half. AI implementation consultants integrate anything you point them at and hold no view on which marketing work is worth the effort, so they will cheerfully automate a process that should have been deleted. Marketing advisors have that judgment and cannot configure an environment or leave behind something that runs on its own. You buy twice and integrate the purchases yourself.

The expensive route is automating inside your tools

Under pressure, every team automates inside the tool it already controls. Marketing builds in the automation platform, sales in the CRM, and support in the ticketing system, because that is the cheapest move available without asking permission. A year of that produces visible progress and a real problem, since nobody owns what happens between those systems.

Other people are already measuring the cost. Forrester projects that 75% of technology decision-makers will see their technical debt reach moderate or high severity by 2026, driven by how fast AI development is adding complexity.^[5] Every tool-by-tool automation built this year is an entry on that ledger.

Reaching the next rung means pulling apart what each team built alone. The three components map onto the ladder below. Standards and unattended workflows get a team to the delegated stage, and enablement is what makes the connected stage possible.

Here’s a maturity model:

1. Personal

• What it looks like: People use AI on their own work, on whatever they paste in

• Who owns it: Nobody, since a habit cannot be handed over

• What it’s worth: Hours back for one person

2. Delegated

• What it looks like: A step runs without a person starting it, inside one tool

• Who owns it: Each team owns its tools, nobody owns what sits between them

• What it’s worth: A process metric moves

3. Connected

• What it looks like: Marketing, sales, and CS work from the same context

• Who owns it: A named owner of that shared context

• What it’s worth: Unit economics move

4. Self-correcting

• What it looks like: The system improves from outcomes without being rebuilt

• Who owns it: A dedicated role

• What it’s worth: Unit economics compound

Three questions move you between the rungs. Does anything run when nobody starts it? Do marketing, sales, and customer success work from the same context? Does it improve next quarter without somebody rebuilding it? Score an average week rather than the best example anyone can point to, because a company sits at its lowest rung, no matter how good that example is.

Where you probably score, and what to do about it

Almost every team I talk to sits at the personal stage, where AI is somebody’s habit, with one delegated-stage anecdote attached, and saying so out loud is usually the most useful ten minutes of the conversation. Between 3% and 7% of private B2B SaaS companies have anyone doing this job, so being behind is the ordinary condition. The 54% figure at the fastest-growing companies is the part worth staring at, because those companies did not grow faster by accident and staffed this before anyone had agreed what to call it.

So start with a question, and leave the hire for later. Picture your primary model going dark tomorrow, then picture your best prompter taking a two-week holiday, and ask what would still run. Write down the standards first, automate one thing that runs unattended second, and hand it to your team third.

If you want a second opinion, I am happy to spend thirty minutes walking the ladder against your situation and naming the one thing worth building first. No slides, and it is useful whether or not you hire anybody.

Vendors define this work as plumbing because plumbing is what they sell. What they left out is the part that was always hard: knowing what deserves to be built at all.

Interested in learning more? Book a meeting.

Share

Leave a comment

Frequently asked questions

What is AI GTM engineering?

AI GTM engineering builds AI systems into how a go-to-market team works, rather than handing individuals a chat window and hoping the way work moves changes on its own. It covers three components in order. First comes the environment, data, and voice standards written down where a model can read them. Second come workflows that run without a person starting them. Third comes the enablement that leaves your own team operating, extending, and repairing what got built.

How is AI GTM engineering different from RevOps?

An analysis of a thousand GTM engineer job postings found that nine out of ten listed responsibilities also appear in RevOps postings, so as those roles are currently written, the difference is thin. The distinction worth keeping is scope. RevOps maintains and optimizes the systems a company already runs, mostly inside sales and prospecting tools. AI GTM engineering covers the marketing work those definitions leave out, including positioning and launches, competitive analysis, and the content a buying committee reads.

Should we hire a GTM engineer or use a consultant?

Most teams are not ready to hire. Under one percent of seed and Series A B2B SaaS companies have anyone in this role, and a common recommendation is to prove the value with a consultant or agency on a specific use case before opening a headcount. An estimated 45% of the people carrying the title already work as agencies or consultants. If you decide the capability is permanent, hire, and let what the consultant built become the job description.

How do I know whether our AI work is a system or a habit?

Ask whether anything runs when nobody starts it. A prompt somebody keeps in a document and pastes into a chat window on Tuesdays is a habit, and it disappears the week that person gets busy. Something that wakes on its own, does the work, and files the result where the team already looks is a system. Score what most of your team does in an average week rather than the best example anyone can point to.

What should a small marketing team automate first?

Write down your standards before automating anything. The environment, the data, and your voice and editorial rules have to live in files a model can read, because everything built on top of them inherits that judgment about what good looks like. After that, pick one workflow that currently eats a repeating block of somebody’s week, build it so it runs unattended, and hand it to the person whose time it gives back.

Does this require a specific AI vendor?

No, and building it that way is a mistake. When your rules, workflows, and accumulated context live as files on a disk rather than inside one vendor’s product, any capable model can read them, and switching costs almost nothing. That property is what lets a system survive an outage, a price increase, or a procurement decision made above your head.

About David Sweenor

David Sweenor is the founder and host of the Data Faces podcast, where he talks with the people who are making data, analytics, AI, and marketing work in the real world. He is also the founder of TinyTechGuides and a recognized top 10 AI thought leader and international speaker who specializes in practical business applications of artificial intelligence and advanced analytics.

With over 25 years of hands-on experience implementing AI and analytics solutions, David has supported organizations including Alation, Alteryx, TIBCO, SAS, IBM, Dell, and Quest. His work spans marketing leadership, analytics implementation, and specialized expertise in AI, machine learning, data science, IoT, and business intelligence. David holds several patents and consistently delivers insights that bridge technical capabilities with business value.

Books

- Artificial Intelligence: An Executive Guide to Make AI Work for Your Business

- Generative AI Business Applications: An Executive Guide with Real-Life Examples and Case Studies

- The Generative AI Practitioner’s Guide: How to Apply LLM Patterns for Enterprise Applications

- The CIO’s Guide to Adopting Generative AI: Five Keys to Success

- Modern B2B Marketing: A Practitioner’s Guide to Marketing Excellence

- The PMM’s Prompt Playbook: Mastering Generative AI for B2B Marketing Success

Follow David on Twitter @DavidSweenor and connect with him on LinkedIn.

Footnotes

^[1]Poyar, Kyle. “Do you need a GTM engineer?” Growth Unhinged, September 10, 2025.

https://www.growthunhinged.com/p/do-you-need-a-gtm-engineer

^[2]Chiu, Henley Wing. “I analyzed 1000 GTM Engineering jobs - here is what I learned.” Bloomberry, October 3, 2025. https://bloomberry.com/blog/i-analyzed-1000-gtm-engineering-jobs-here-is-what-i-learned/

^[3]Short, Brendan J. “54% of the Fastest Growing B2B SaaS Companies have a GTM Engineer.” The Signal, April 23, 2026.

The Signal

54% of the Fastest Growing B2B SaaS Companies have a GTM Engineer

If you were forwarded this newsletter, join 7,783 weekly readers—some of the smartest GTM founders and operators—by subscribing here…

Read more

6 months ago · 13 likes · 2 comments · Brendan J Short

^[4]Woodward, Curt. “GTM Engineers: The Real Trends Behind the Hype.” ZoomInfo Pipeline, updated July 7, 2026. https://pipeline.zoominfo.com/sales/gtm-engineer-hype

^[5]Forrester Research, Predictions 2025, reported in Alexis, Alexei. “AI rush is fueling tech debt ‘tsunami’: Forrester.” CFO Dive, November 26, 2024. https://www.cfodive.com/news/tech-debt-tsunami-building-amid-ai-craze-forrester/733984/

▶ Watch the full episode and read the transcript: Jim Jinright on the AI job market and judgment capital

When I started at IBM, the site had about 13,000 people. By the time I left, it was closer to 3,000, and somewhere in between I absorbed the lesson that no company was going to keep me around any longer than tomorrow. A few RIFs, acquisitions, divestitures, etc. proved to me that this is how the business world behaves — it’s cruel, unapologetic, and without remorse. So when the early-career people I mentor tell me that AI took the jobs they were promised, I understand the feeling, and I also suspect the story is more complicated than the headline.

I wanted a second opinion from someone who reads the market for a living, so I asked Jim Jinright to join the Data Faces Podcast. Jim runs an executive search firm, and he spends his days talking to the CEOs, CIOs, and CTOs who decide whether a requisition opens or stays closed. His answer to “is it AI?” is mostly no. He sees a perfect storm that arrived a couple of years before the AI headlines did, and AI just happened to be standing there when the blame got assigned.

About Jim Jinright

Jim Jinright is the managing partner of JRJ Search Group, an executive search firm serving the Dallas-Fort Worth technology market. He has been in search for about twelve years and took over JRJ, the firm his father founded, after running CIO, CTO, and chief data officer searches for a regional boutique. JRJ places strategic and technical people in mid-market companies.

In this episode, Jim and I discuss:

• Why the layoffs blamed on AI trace back to pandemic overhiring, the 2022 interest rate increases, and labor arbitrage

• Where the junior work went, and why senior people with judgment are more in demand than before

• What a recruiter does when 500 applications arrive within an hour, and how a candidate gets from 500 to the final five

• Ghost jobs, age screens, and AI interviews, and what a candidate can do about each

• Jim’s advice for early-career engineers: move up the stack from writing code to running the software factory

Watch the full conversation here:

A perfect storm with AI’s name on it

Jim does not rule out layoffs caused by AI, and he thinks AI could create more jobs than it removes once the dust settles. I vaguely recall something about Schumpeterian destruction, but I digress. I made a version of that argument in January 2024 in Will AI take my job? Maybe.: every technology wave so far has created more jobs than it destroyed, just not for the same people. Jim simply doesn’t believe AI is what’s been happening in the searches he runs. His account starts with the pandemic, when companies hired far more people than they needed. Around 2022, interest rates rose, capital got expensive, and the startups and mid-market companies that had been expanding pulled back. Cost cutting followed, and much of the work that stayed went offshore.

AI arrived on top of all that, bringing a seductive idea with it. Maybe you do not need 50 engineers. Maybe you need 10, or none. Doug Laney told me on this show that he expects the first billion-dollar business run by a handful of people and a swarm of AI agents. Jim thinks the notion is correcting as companies realize they still need people to make decisions and build things, and he is getting more calls for talent than a year ago.

He calls the period we are leaving the no-hire, no-fire economy and the one we are entering the low-hire, no-fire economy. That is a change in direction rather than a boom, but for anyone who has been sending applications into a void, it matters that the void is starting to answer.

“I wouldn’t attribute all the layoffs to AI per se. I think a lot of people call this the no-hire, no-fire economy, and now we’re moving into the low-hire, no-fire economy.”

— Jim Jinright, Managing Partner, JRJ Search Group

Where the junior work went

Jim’s explanation for who has been affected starts with a simple observation: AI is powerful, but it has no context. Senior people are more in demand than before, since companies want someone who can review everything the model produces and say which facts matter and whether the work is right. That person knows what good looks like, and Jim says that is not an easy find.

Junior people have been hit harder, and the mechanism is specific. Jim would never hand a new recruiter his biggest client on day one, but over time he would train that person up to it, and the small problems were the training. Now much of that work goes to AI, and the ladder loses its bottom rungs.

On the engineering side, it takes real engineers to build AI systems, and anyone who has not built agents or put one into production is, in Jim’s words, probably talking a big game. Knowing how to use ChatGPT is not the bar. The requirement is someone who can own the outcome and produce a return, and Jim called that, when we spoke before the recording, judgment capital. Human capital used to sit above financial capital in what a company needs most. With AI doing the work at the bottom, everyone moves up the stack to judgment.

Every resume is perfect now

A client of Jim’s posted a position and received more than 500 applications within an hour. The company couldn’t get through them, and that is the state of hiring right now. Candidates use AI to apply, often to the same job more than once, and AI sorts the applications. In some companies, AI runs the first interview. Every resume has the right words, and every LinkedIn profile is polished, so the signals a recruiter used to rely on have stopped meaning anything.

Jim gets through the noise with conversations, a lot of them. AI can surface 500 profiles, but getting from 500 to five, and then to the one person a client should hire, means talking to 30, 40, or 50 people. He describes a hiring manager as a human being crying out for help, and the way through is to show that person you can solve their specific problem.

“That’s how you differentiate yourself, is just being able to bring the receipts with what you’ve done. Problem solved, value created.”

— Jim Jinright, Managing Partner, JRJ Search Group

Bringing the receipts means problems you owned and outcomes you delivered, with enough detail that the claim can be checked. Emotional intelligence and extreme ownership are the traits Jim’s clients ask for, and neither survives being written on a resume. Jim can tell within minutes whether someone is transactional or someone who cares about the problem. He wants to hire the person who would show up to open the store after a snowstorm. I added my own advice: be yourself in that conversation, because landing in the wrong role by pretending is bad news for everyone involved.

Ghost jobs, age screens, and the interview bot

People I mentor keep asking whether the listings they see are real, the same job posted in fifteen cities that never seems to close. Jim has seen them too, and executives he knows have confirmed the practice: some companies advertise roles that do not exist to build a talent pipeline for later. He calls it a horrible thing to put a person through when their back is against the wall. His larger point is one I wish someone had told me during my first layoff. When responses aren’t coming, it is not always you. Sometimes it is just the moment.

I also asked him about age, since more than one experienced person has told me about being waved off by a five-year-old company that only hires 20-year-olds. Show the last ten years rather than a snapshot of thirty, and lead with the value. People with experience, judgment, and a track record are in demand right now, and smart companies understand that hiring across the spectrum benefits them. Job hopping got a similar treatment. Most careers now run two years here and three there, and that reads fine as long as the resume explains the moves rather than leaving a hiring manager to guess.

The AI interview drew Jim’s sharpest reaction of the conversation. A company that cannot spare 30 minutes for a screening call is not going to land a rock star who already has a job, and putting a professional into a funnel to talk to a bot is, in his words, a slap in the face. The candidate is evaluating you too, and none of his clients hire that way. Leigh Felton made a related point on this show. Had AI been screening candidates when she started out, it would have rejected the person who went on to run Microsoft’s first ethical AI enablement program.

Move up the stack

When Jim started recruiting, every college graduate had something like ten open positions to choose from, and today the ratio has inverted. He feels for the engineers who finished a computer science degree and cannot find work. Coding is becoming a commodity because the AI can code, so his advice to a junior person is to stop competing on React or Kotlin and move into the judgment sphere. Be the person who can build the software factory: set the guardrails, decide what should and should not be built, and learn to manage an AI system once it is running. Strong technical skills still matter, and strong communication has to travel with them.

He also has a specific move for anyone whose applications are going nowhere. Stop asking people for a job. Identify leaders in your industry, tell them you researched their background, and ask for advice on which user groups and masterminds to join. Some of those leaders open doors later. Fractional and contract work fits the same pattern. Companies like bringing in an expert for a season while they figure out what to invest in. It brings in revenue, expands your network, and keeps you sane, which he says is the real job during a search.

So here is what I took away. AI is not hiring or firing anyone. Interest rates, overhiring, and cost-cutting did most of the damage, and the correction has started. What AI has changed is where the value sits, and it sits in judgment. Build the receipts that prove yours.

Listen to the full conversation with Jim Jinright on his Data Faces Podcast episode page.

Based on insights from Jim Jinright, Managing Partner at JRJ Search Group, featured on the Data Faces Podcast.

Share

Leave a comment

Podcast highlights

• [0:06] Introduction and welcome to the Data Faces Podcast

• [1:11] JRJ Search Group, twelve years in search, and the Dallas Fort Worth market

• [2:16] A would-be pilot, a demo tape, and the uncle who ran A&M Records in Canada

• [3:33] Is it AI? The perfect storm behind the layoffs

• [4:10] Pandemic overhiring, 2022 interest rates, labor arbitrage, and the “10 or none” idea

• [5:45] From no-hire, no-fire to low-hire, no-fire

• [6:16] AI has no context, and why senior judgment is in demand

• [7:30] The small problems used to go to the junior person

• [8:23] Every resume is perfect. How does anyone stand out?

• [9:04] 500 applicants in an hour, and getting from 500 to five

• [10:44] LinkedIn as the largest talent database, and why the network still wins

• [11:19] Ghost jobs, and why it is not always you

• [13:35] Dave’s own layoffs, and IBM going from 13,000 to 3,000

• [14:38] Staying sane during a search

• [16:03] Age discrimination, and showing the last ten years

• [18:24] What “AI literate” means, and the skills clients ask for

• [21:31] How do you demonstrate emotional intelligence on paper?

• [23:04] Behind every job posting is a human being asking for help

• [24:43] Be yourself, and the rise of fractional work

• [26:31] The AI interview, and why Jim calls it a slap in the face

• [28:44] Job hopping, acquisitions, and the resume that explains itself

• [32:57] Advice for early-career engineers: move up the stack

• [35:20] Ask leaders for advice, not for a job

• [36:57] Landman, boat adventures on YouTube, and being an information junkie

• [38:11] Where to find Jim

Frequently asked questions

Is AI causing the tech layoffs?

Mostly no, according to executive recruiter Jim Jinright of JRJ Search Group. The tech job market slowdown traces to a combination of forces that arrived before the AI headlines: overhiring during the pandemic, the interest rate increases that began in 2022, and the cost cutting and labor arbitrage that followed. AI added the idea that a company might need ten engineers rather than fifty, but Jim sees that notion correcting as companies realize they still need people to make decisions and build things.

What is a “low-hire, no-fire” economy?

A low-hire, no-fire economy is a job market in which companies are not laying people off in large numbers but are opening only a small number of new positions. Jim Jinright uses it to describe the market in late 2026, as it moves out of the no-hire, no-fire period that followed the 2022 rate increases. He reports more calls for talent than a year earlier, which he reads as an early sign of a turn.

Which jobs are most affected by AI?

Junior roles are the most affected, because AI now does much of the smaller, well-defined work that junior staff used to learn on. AI is powerful but has no context, so companies are hiring senior people who can judge which of its outputs matter and whether the work is correct. On the engineering side, demand is strongest for people who have built and run AI systems, rather than people who know how to use a chatbot.

How do you stand out when every resume is written by AI?

Through conversations, and by bringing the receipts. When a posting draws 500 applications in an hour, AI can sort the pile, but a recruiter still has to talk to dozens of people to find the one worth hiring. Candidates who differentiate themselves show specific problems they owned and outcomes they produced, in enough detail to be verified. Traits like emotional intelligence and ownership only show up in that conversation, never on the page.

What is judgment capital?

Judgment capital is Jim Jinright’s term for the value a person adds by knowing what good looks like and which facts matter, and by taking responsibility for the outcome. Human capital used to sit above financial capital as the scarce resource in a company. With AI absorbing the routine work, the scarce resource moves up the stack to judgment, and companies pay for people who have it.

What should an early-career engineer do about AI?

Move up the stack. Jim’s advice is to stop competing on a specific language and become the person who can build the software factory, which means setting the guardrails and deciding what should and should not be built, and then managing the AI system once it runs. Pair that with strong communication, and replace cold job applications with requests to industry leaders for advice on which user groups and masterminds to join.

About David Sweenor

David Sweenor is the founder and host of the Data Faces podcast, where he talks with the people who are making data, analytics, AI, and marketing work in the real world. He is also the founder of TinyTechGuides and a recognized top 10 AI thought leader and international speaker who specializes in practical business applications of artificial intelligence and advanced analytics.

With over 25 years of hands-on experience implementing AI and analytics solutions, David has supported organizations including Alation, Alteryx, TIBCO, SAS, IBM, Dell, and Quest. His work spans marketing leadership, analytics implementation, and specialized expertise in AI, machine learning, data science, IoT, and business intelligence. David holds several patents and consistently delivers insights that bridge technical capabilities with business value.

Books

• Artificial Intelligence: An Executive Guide to Make AI Work for Your Business

• Generative AI Business Applications: An Executive Guide with Real-Life Examples and Case Studies

• The Generative AI Practitioner’s Guide: How to Apply LLM Patterns for Enterprise Applications

• The CIO’s Guide to Adopting Generative AI: Five Keys to Success

• Modern B2B Marketing: A Practitioner’s Guide to Marketing Excellence

• The PMM’s Prompt Playbook: Mastering Generative AI for B2B Marketing Success

Follow David on Twitter @DavidSweenor and connect with him on LinkedIn.

Agentic AI drew the crowds and the CDO role got personal, but data quality was the theme nobody could avoid. How can you trust what AI tells you when you cannot trust the data underneath it? Over and over, leaders told me the AI conversation keeps dragging everyone back to fundamentals they hoped were solved 20 years ago, and they are not.

Tom Redman, the Data Doc and president of Data Quality Solutions, revisited the study that made his name, the one where only 3 percent of companies tested met basic data quality standards. His unpublished follow-up work suggests things have gotten no better since. Tom’s real point is about people, since the regular folks without data in their titles do most of the data work and hold the key to fixing it, and misaligned incentives manufacture bad data on purpose, like the badge scans that measure marketers on volume while salespeople drown in unqualified leads. If the data quality pitch worked, why are we still having this conversation 30 years later?

Fern Halper, founder of the AI Foundations Group and VP of research at TDWI, arrived with a warning for every executive chasing the AI budget. AI will expose every weakness your company has. She separates data governance from AI governance, notes that the median score on TDWI’s data governance maturity model still sits at 58 out of 100, and points out that trust in unstructured data trails structured data by roughly 20 percentage points, right when generative and agentic AI make unstructured data essential. Companies hit a value ceiling, she told me, the moment they run generative AI without their own data underneath it.

Stacie Christensen, a senior data leader at H-E-B with 22 years in the industry, grades our collective data quality somewhere between a D plus and a C minus, and she is the one who framed the whole event’s fear for me. AI is a multiplier and a scaler with nothing positive to scale when the foundation is missing. She makes the case that governance belongs at the moment of data creation, because bad data gets exponentially more expensive the further down the pipeline it travels, and she uses circular data flows as her favorite red flag for hidden complexity.

Leticia Naqvi, a people analytics research manager at Apple, brought more than a year of research on organizational AI readiness, built on four pillars of leadership alignment, data maturity, innovation culture, and change management. She warns that the pillar leaders forget is data maturity, and it sits underneath every AI output they will ever trust. Her advice for CDOs wondering where to start is an honest current-state assessment and one function at a time rather than a big-bang rollout, since she watched four external partners each hold a different meaning for a single data definition.

The foundation theme ran through the largest group of conversations, so here is where the rest of them landed.

• Danette McGilvray, Granite Falls Consulting: She compares most data cleanup to a murder scene where someone hauls the body away without investigating, which is why the same crime scenes keep happening. Her Ten Steps methodology starts with the question too many teams skip, whether anyone actually cares about the problem they are excited to fix.

• Jonathan Agee, Validatar: The question that started his company is the one every leader is asking now, how do I know this data is right? He grades the state of data quality a C and argues for treating quality like software QA, with prevention that starts in development rather than firefighting in production.

• Kelley Kassa, BARC US: Her research found data management is the number one priority for corporate performance management teams, ahead of AI, because finance teams know their foundation is not ready. Only 9 percent of finance AI in North America is in production, and her line on accountability sticks, the agent does not go to jail when the number is wrong.

• Gwen Thomas, The Data Governance Institute: When she registered datagovernance.com in 2003, a search on the term returned 52 hits, and 39 of them were IBM. She explains how governance pulls AI projects out of POC purgatory by handling the edge cases, and why organizations govern for the same reason cars have brakes, so they dare to go fast.

• Dan Everett, Insightful Research: After a path into brain science that began with his son’s autism diagnosis, he makes the case that AI adoption is really change management. Recent studies, he notes, show claimed productivity gains shrink once domain experts have to validate the output on the back end.

• Intuit Credit Karma (Veenit Shah and Puneet Singh): Their team keeps credit scores current for about 140 million members across more than 100 tables and 40,000 columns, monitored on five data quality pillars. An AI remediation agent now compresses root cause investigations that once took 30 to 40 minutes into just a few, with guardrails and security review gating every step up the maturity curve.

Where this leaves you

Every one of these leaders came at data quality from a different angle, and they landed in the same place. AI scales a weak foundation rather than fixing it, and it does so with confidence that makes the errors harder to catch. The teams pulling ahead governed data at the moment of creation and measured quality where the work happens, and they refused to trust an output they could not trace back to something real.

This is one theme of three. Head back to the full CDOIQ 2026 field notes for the complete roster, or read the other two themes on agentic AI and the context problem and the chief data officer role at 20 years. All 24 interviews are on the Data Faces Podcast on-location hub.

Leave a comment

Share

Frequently asked questions

Why is data quality such a big theme for AI?

AI systems act on the data they are given, so poor data quality produces poor and often confidently wrong outputs. Tom Redman’s research found only 3 percent of companies met basic data quality standards, and BARC found 70 percent of companies say less than half of their unstructured data is usable for AI. As Stacie Christensen of H-E-B put it, AI is a multiplier with nothing positive to scale when the foundation is missing, so weak data becomes more expensive and more visible once AI acts on it.

Where should a company start with data quality for AI?

Leaders at CDOIQ 2026 recommended governing data at the moment of creation rather than downstream, because bad data gets exponentially more expensive the further it travels. Danette McGilvray advised starting with whether anyone actually cares about the problem, and Leticia Naqvi of Apple recommended an honest current-state assessment and starting with one function rather than a big-bang rollout. Every leader landed on the same first step, fixing the foundation before scaling AI on top of it.

What is the difference between data governance and AI governance?

Fern Halper of TDWI drew the line clearly. Data governance covers the quality, ownership, and control of the data itself, while AI governance adds oversight of the models, agents, and automated decisions built on that data. Confusing the two leads to mistakes, because governing the AI without governing the underlying data leaves the foundation untrusted. Both matter more as generative and agentic AI make unstructured data essential.

About David Sweenor

David Sweenor is the founder and host of the Data Faces podcast, where he talks with the people who are making data, analytics, AI, and marketing work in the real world. He is also the founder of TinyTechGuides and a recognized top 25 analytics thought leader and international speaker who specializes in practical business applications of artificial intelligence and advanced analytics.

With over 25 years of hands-on experience implementing AI and analytics solutions, David has supported organizations including Alation, Alteryx, TIBCO, SAS, IBM, Dell, and Quest. His work spans marketing leadership, analytics implementation, and specialized expertise in AI, machine learning, data science, IoT, and business intelligence. David holds several patents and consistently delivers insights that bridge technical capabilities with business value.

Books

• Artificial Intelligence: An Executive Guide to Make AI Work for Your Business

• Generative AI Business Applications: An Executive Guide with Real-Life Examples and Case Studies

• The Generative AI Practitioner’s Guide: How to Apply LLM Patterns for Enterprise Applications

• The CIO’s Guide to Adopting Generative AI: Five Keys to Success

• Modern B2B Marketing: A Practitioner’s Guide to Marketing Excellence

• The PMM’s Prompt Playbook: Mastering Generative AI for B2B Marketing Success

Follow David on Twitter @DavidSweenor and connect with him on LinkedIn.

▶ Watch the full episode and read the transcript: Data Quality for 140 Million Members | Intuit Credit Karma

What happens when a data quality alert fires at two in the morning and says that a few percent of last night’s data didn’t load properly? For many data teams, that’s a ticket to deal with later. Someone picks it up after their coffee or tea, reruns the pipeline, and all is copacetic. But what happens if you change your frame of reference and read the same alert as a count of people? At Intuit Credit Karma, Veenit Shah’s team treats that alert as a few percent of 140 million people who will open an app today and see a credit score that’s out of date. That habit of counting people rather than rows is the whole story of how the company got out of the firefighting years.

Part of the reason data quality programs fizzle is that nobody quite agrees on what quality data means anymore, and AI has made the definitions slipperier. Malcolm Hawker argued on this show that data is only AI-ready when it supports the use case in front of it, and Brendan Grady said that bad data didn’t matter much until AI started acting on it. Both of those conversations were about definitions. Veenit and Puneet Singh’s is about what a team does with an alert once the definitions run out.

I caught Veenit, Senior Manager of Data Engineering, and his colleague Puneet Singh, Senior Data Engineer, at the CDOIQ Symposium in Cambridge, Massachusetts, the day after they presented Credit Karma’s data quality journey to the conference.^[1] The talk wasn’t a victory lap – it’s still a work in progress. Veenit described a program that started in a reactive state a couple of years ago and picked up what he calls “scars along the way,” and the two of them came to share their learnings rather than the trophy.

“Regardless of what level you’re at, whether you’re just starting as a junior engineer or whether you’re a senior manager as well, it is everybody’s job.”

— Veenit Shah, Senior Manager of Data Engineering, Intuit Credit Karma

So, who’s job is data quality? Everybody’s job is a line I’ve heard in quite a few governance decks, and when you hear this, it’s usually a signal that it’s nobody’s job. What makes it true at Credit Karma is a translation the team does on every alert, and I think that translation explains the pillars, the measurement discipline, and the AI agent that now does the investigating.

About Veenit Shah and Puneet Singh

Veenit Shah has spent close to five years at Credit Karma and leads the data engineering team that keeps credit scores current for the company’s members. Credit Karma became part of Intuit in December 2020, when it had more than 110 million members, and it now reports more than 140 million.^[2],[3] Every one of those members expects an accurate score when they open the app, which is the number Veenit’s team measures itself against.

Puneet Singh is a Senior Data Engineer on Veenit’s team and has been at Credit Karma for four and a half years. His twelve years in data engineering started as a Hadoop developer, then moved to the cloud, and today he works mostly on Google Cloud, building microservices that ingest data and check its quality. He demonstrated the team’s agents writing production pipeline code on stage at the symposium.

In this episode, Veenit, Puneet, and I discuss:

- What startup-era data engineering looks like when there are no pillars to rely on

- Why five data quality pillars beat nineteen, and how a small proof of concept earned the whole program

- How every alert across 40,000 columns gets logged, measured, and reviewed each week

- Why a single ingestion alert has to be read as a count of members

- What an AI remediation agent does, what it is allowed to touch, and why it only works because the discipline came first

Watch the full conversation here:

Nightly calls and no pillars to stand on

When Puneet joined Credit Karma, the company still had what he calls the start-up feeling, even after Intuit acquired it. Move as fast as you can and put out as many fires as you can. There were no set data quality checks and no patterns to build on, so the job ran on long hours and nightly calls, and the same problems came back the following week because nothing had been fixed at the root.

“There were no set data quality checks, there were no patterns or pillars that we could rely on, and it was a mess, like long hours and nightly calls.”

— Puneet Singh, Senior Data Engineer, Intuit Credit Karma

It’s easy to stay in reactionary mode when there’s always another problem to fix, so how does a team ever win the mandate to stop and fix the root? Veenit says that a few incidents made it clear that the team’s data quality posture wasn’t where it needed to be, and the leaders above them let the engineers trust their instincts about what to double down on. Both things had to be true at once, because incidents without trust produce a blame cycle, and trust without incidents produces a roadmap that nobody funds.

What they had in front of them, in Veenit’s words, was a blank board, and they had to decide how to start drawing. The program is now at least three years old, and he was careful to say that it’s never complete. Generative AI and agentic applications are changing the end state faster than anyone can describe, so the team focuses on the next step rather than the destination.

Five pillars, not nineteen

A speaker in an earlier session at CDOIQ had offered nineteen different ways to do data quality. Puneet’s team went the other way and defined the pillars that fit its own data, including timeliness, completeness, accuracy, and governance. That sounds like a smaller ambition, doesn’t it? In practice, it’s the more demanding one, because five pillars that are measured on every table beat nineteen dimensions on a slide that nobody instruments.

Puneet was equally practical about how the program earned its funding. He pitched a data quality vendor for one narrow use case, showed a measurable benefit, and only then came back for the whole thing. Leadership trusting the team’s instincts didn’t mean signing a blank check. It was the the small proof of concept that translated it to a budget item.

“Each and every alert is logged, measured, and we do analysis on a weekly basis to make sure where the fires are, where we need to get better, and where we need brutal prioritization to remove them or fine-tune those.”

— Puneet Singh, Senior Data Engineer, Intuit Credit Karma

The team runs data quality checks on more than 100 tables and roughly 40,000 columns, and logs and measures every alert those checks produce. A weekly review then decides where the fires are and which alerts deserve brutal prioritization, either fixing the underlying issue or tuning the check so it stops crying wolf. Most data quality programs I’ve seen skip that review, and they end up with a dashboard full of red that everyone has learned to ignore. How many of those dashboards are in your own organization right now?

Translate the alert into people

I asked the question I ask every data leader: whose job is data quality? The business? The data engineering team? An agent? Veenit’s answer was the everybody’s-job line, and normally I would have moved on, because everybody’s job is what people say when they haven’t decided. Then he explained why it holds on his team, and the reason turned out to be a translation habit rather than a value statement.

“A simple alert could look like, ‘ Hey, we have not been able to ingest X percent of data. But what that actually means is X percent of members out of the 140 million are going to be impacted.”

— Veenit Shah, Senior Manager of Data Engineering, Intuit Credit Karma

Once an alert is read as a count of members rather than a percentage of rows, ownership stops being a governance question. A junior engineer who sees millions of people behind a failed ingestion doesn’t need a RACI chart to know it matters, and a senior manager can’t delegate it into a backlog. Veenit calls this a fundamental aspect of the team’s culture, and it’s also the pitch that I wish more data leaders would make to their own organizations. Asking people to care about the members, customers, or patients on the other side of the pipeline is a much easier sell than asking them to care about data quality.

Veenit also drew a useful line between the two sources of trouble. External issues arrive from partners and financial institutions whose data the team doesn’t control, and those are the cards you are dealt, so the only question is how to play the hand. Internal issues show up when the team onboards a new partner or builds a new feature, and better engineering practice can prevent those. Separating the two keeps the weekly review from treating an upstream mess and a homegrown bug as the same kind of fire.

The agent came last

Every conversation about data operations right now turns to agents within ten minutes, and this one was no different. Credit Karma built an AI remediation agent that picks up a data quality error, runs the diagnostic query, refines it, and keeps running queries and analysis until it can state the root cause and the blast radius. Puneet said that investigation used to take an engineer 30 to 40 minutes and now takes a few minutes. On the build side, the team has given agents the context to write production-ready pipeline code, including the data quality and observability pieces, in five to ten minutes.

Those results get headlines. So why does the agent work here when so many bolted-on agents don’t? The agent runs against alerts that are already logged and measured, on checks that three years of weekly reviews have tuned, with runbooks the team wrote for humans and then handed to the agent as explicit instructions. It operates in a protected environment, never touches personally identifiable data, and works only with anonymized data that was already approved for analytics. You can’t automate a mess, and Credit Karma’s agent finds the root cause in minutes because the discipline it runs on came first.

“The agent goes there, looks into the data quality error, runs the query, as well as refines those queries, and keeps running the queries and analysis till it comes to the conclusion with the blast radius, saving our time from 30 to 40 minutes to just a few minutes.”

— Puneet Singh, Senior Data Engineer, Intuit Credit Karma

I raised the obvious worry: agents sometimes go off the rails. The team places itself at level two on its own agentic maturity model, and everything planned for levels three and four is still under security review. Selective access, which tools an agent can use, and what its role should be will get decided slowly, and the team is comfortable saying it doesn’t yet know what levels four or five look like. Veenit put it plainly: the game itself is changing, and they have to keep playing it.

The benefit compounds

Veenit described the payoff as multifold, and the first part is the one everyone expects. Investigations that took 30 minutes to an hour are compressed into minutes, giving the team back time he described as insane. The second part is what I think matters more. With firefighting handled, the team can think about the next frontier, find gaps it didn’t know existed, and stop every data quality issue the program catches from compounding on the last one.

If your own data quality program is stuck in the firefighting years, the place to start isn’t a framework or an agent. Start by naming who is on the other side of the next alert.

Listen to the full conversation with Veenit Shah and Puneet Singh on their Data Faces Podcast episode page.

Based on insights from Veenit Shah, Senior Manager of Data Engineering, and Puneet Singh, Senior Data Engineer, at Intuit Credit Karma, featured on the Data Faces Podcast.

Share

Leave a comment

Podcast highlights

- [0:03] David introduces Veenit Shah and Puneet Singh on location at the CDOIQ Symposium in Cambridge, Massachusetts

- [0:27] Veenit on keeping credit scores current for about 140 million members

- [1:49] Veenit on the session, a data quality adoption journey from reactive to proactive, and the scars along the way

- [2:23] Puneet on the starter feeling, extinguishing fires, and nightly calls with no pillars to rely on

- [3:25] Veenit on the twofold mandate, incidents plus leaders who trusted the team’s instincts

- [5:08] Puneet on choosing five pillars over nineteen, and the small proof of concept that earned the program

- [5:46] Puneet on 100-plus tables, 40,000 columns, and the weekly review with brutal prioritization

- [7:07] Veenit on why data quality is everybody’s job when an alert means members

- [8:13] Puneet on the AI remediation agent and 30 to 40 minutes becoming a few

- [10:09] Veenit on level two of the agentic maturity model and the runbooks handed to the agent

- [12:21] Puneet on guardrails, anonymized data, and what stays under security review

- [13:22] Veenit on the compounding benefit and the gaps they did not know existed

About David Sweenor

David Sweenor is the founder of TinyTechGuides and host of the Data Faces Podcast. He is an international speaker, advisor, and the author of eleven books on artificial intelligence, analytics, and B2B marketing, including Generative AI Business Applications, The CIO’s Guide to Adopting Generative AI, and Modern B2B Marketing. With more than twenty-five years in analytics and AI at companies including Alteryx, Tableau, TIBCO, SAS, IBM, and Dell, David advises technology companies on product marketing, content strategy, and go-to-market execution. He holds several patents and has been named a top influencer in data and analytics by Onalytica, Thinkers360, and Analytics Insight.

Connect with David on LinkedIn and subscribe to the Data Faces Podcast for conversations with the people shaping enterprise data and AI.

Footnotes

^[1]CDOIQ Symposium. “The 20th Annual CDOIQ Symposium.” July 21–23, 2026, Hyatt Regency Cambridge, Massachusetts.

https://2026cdoiq.org/

^[2]Intuit Inc. “Intuit Completes Acquisition of Credit Karma.” Intuit Press Room, December 3, 2020. https://www.intuit.com/company/press-room/press-releases/2020/intuit-completes-acquisition-of-credit-karma/

^[3]Intuit Credit Karma. “About Intuit Credit Karma.” https://www.creditkarma.com/about

CDOIQ is where the chief data officer role was born, so the 20th anniversary put the job itself on the table. If the job description changes every single year, how do you know when you have the right person in the seat? The people who invented the role and the people living it now agreed on one thing, that it is being rewritten in real time, with the pressure coming straight from AI.

Randy Bean, founder and CEO of the Data & AI Leadership Exchange, has organized the CDOIQ chief data officer panel for 12 years, and his test for data leaders is blunt. If you are not getting measurable business value from your data and AI investments, or a clear path to it, go back to the office this afternoon and shut them down. He has watched the role move from defensive risk and compliance work to an offensive, business-focused mandate, and he told me every chief data officer from last year’s panel will be out of that seat by the end of this year. His read on the hype is the line worth taping to your monitor, AI is overestimated in the short term and underestimated in the long term.

Richard Wang, co-founder of the MIT CDOIQ Symposium, gave one of the conversations I will remember longest, fresh off a standing ovation and a Lifetime Achievement Award. He traced the arc from a two-year assignment at the Pentagon that convinced him every major organization needs a chief data officer with budget and authority, to a world with more than 2,000 CDOs today. He also coined a term every data leader will recognize the second they hear it, data archaeology, the cycle where organizations lose the people who hold the knowledge and end up digging up and repeating the same work.

Peter Aiken, a professor at Virginia Commonwealth University, wrote the first book on the chief data officer role 20 years ago, and he still argues data needs one throat to choke. He is blunt about the odds, the average CDO lasts about a year and a half, and the best ones have been fired three times, because doing the job right means telling people things they do not want to hear. His button carries the rest of the message, grow tomatoes, not data centers, and his Anything Awesome premise is the warning underneath all the AI hype, since bad data plus anything awesome still produces bad results.

Mark Ramsey, managing partner at Ramsey International, took up professional motorsports in his 50s and turned it into Data at Speed, a book that maps the five crashes every chief data officer will hit. The first one he named stuck with the room, the shadow pit crew, what forms when the CDO does not drive access to data and the business quietly builds its own workaround. He also put a number on the stakes, half of CDOs do not survive three years, and he argued that blocking generative AI only pushes employees toward free tools that expose company data.

The rest of the CDO conversations circled the same question from different chairs.

• Stuart Madnick, MIT: A Lifetime Achievement honoree who helped launch the data quality work that became CDOIQ, he traced 40 years from the Total Quality Management movement to the insight that data is a manufactured product. His warning for the AI era is direct, by every measure he has seen, attackers are using AI more aggressively and more efficiently than defenders.

• John Ladley, author and advisor: When he asked ChatGPT for a data model and got the rudiments back in 38 seconds, he called it career altering and postponed retirement to fold AI into a third edition of his governance book. His take on the profession is humbling, a field that is only 40 years old still has not figured out its 1.0.

• Leigh Felton, AI for Job Security Foundation: She spent 20 years at Microsoft and built its first ethical AI enablement program, and she argues you cannot de-bias a system built to recognize historical patterns. She makes the case for governing AI like an environment that shapes how people think, closer to OSHA than to a tool you pick up and put down.

Where this leaves you

Twenty years ago, the chief data officer job meant defense, risk, and cleaning up messes. These leaders described a job that is now about proving business value fast enough to keep the seat, with AI raising the bar every quarter. A CDO keeps that chair by holding a real mandate, the authority to drive data access and kill what does not deliver, and that matters more than whether the role reports to technology or the business.

This is one theme of three. Head back to the full CDOIQ 2026 field notes for the complete roster, or read the other two themes on agentic AI and the context problem and data quality as the AI foundation. All 24 interviews are on the Data Faces Podcast on-location hub.

Share

Leave a comment

Frequently asked questions

How has the chief data officer role changed in 20 years?

Leaders at CDOIQ 2026 described the CDO role moving from defensive risk and compliance work toward an offensive, business-value mandate. Randy Bean of the Data & AI Leadership Exchange traced that arc directly and noted survey data showing 42 percent of CDOs report under technology and 33 percent under business leadership. AI is accelerating the change by forcing data leaders to prove measurable value or shut initiatives down.

Why do chief data officers have such short tenures?

Multiple leaders cited high turnover, with Mark Ramsey noting that roughly half of CDOs do not survive three years and Randy Bean predicting every member of last year’s CDO panel would leave that role within a year. Most of them traced it to a mandate problem. When the CDO does not own data access and cannot show business value, the role becomes a target, and AI has raised the stakes on proving that value quickly.

Where should the chief data officer report?

Opinions varied, but the leaders agreed the mandate outweighs the reporting line. Peter Aiken argued data belongs in the business rather than in IT, because every business problem has a data component, and that power and influence matter more than the org chart. Survey figures shared at the event put 42 percent of CDOs under technology and 33 percent under business leadership, with the rest split elsewhere.

About David Sweenor

David Sweenor is the founder and host of the Data Faces podcast, where he talks with the people who are making data, analytics, AI, and marketing work in the real world. He is also the founder of TinyTechGuides and a recognized top 25 analytics thought leader and international speaker who specializes in practical business applications of artificial intelligence and advanced analytics.

With over 25 years of hands-on experience implementing AI and analytics solutions, David has supported organizations including Alation, Alteryx, TIBCO, SAS, IBM, Dell, and Quest. His work spans marketing leadership, analytics implementation, and specialized expertise in AI, machine learning, data science, IoT, and business intelligence. David holds several patents and consistently delivers insights that bridge technical capabilities with business value.

Books

• Artificial Intelligence: An Executive Guide to Make AI Work for Your Business

• Generative AI Business Applications: An Executive Guide with Real-Life Examples and Case Studies

• The Generative AI Practitioner’s Guide: How to Apply LLM Patterns for Enterprise Applications

• The CIO’s Guide to Adopting Generative AI: Five Keys to Success

• Modern B2B Marketing: A Practitioner’s Guide to Marketing Excellence

• The PMM’s Prompt Playbook: Mastering Generative AI for B2B Marketing Success

Follow David on Twitter @DavidSweenor and connect with him on LinkedIn.

▶ Watch the full episode and read the transcript: Malcolm Hawker on AI-ready data and identity resolution

When I built data warehouses at IBM, I used to argue with our chief architect about master data management (MDM). My view was that if the manufacturing data was wrong, we should fix it in the source systems where it was created. His view was that we would build the mapping tables, reconcile the identifiers downstream, and leave the source systems alone. He won those arguments, and for years afterward I understood MDM mechanically without ever being convinced it was solving the right problem.

Then I spent the second half of my career marketing data, analytics, and AI platforms at SAS, TIBCO, Alteryx, and Alation, and I still write that copy for clients. Today it says “AI-ready.” Much as it pains me, the phrase is on nearly every vendor site, it draws a great deal of budget, and I have yet to see two companies define it the same way. So when Malcolm Hawker came on the Data Faces Podcast, I asked him what it means, and whether the companies claiming their data is AI-ready are fooling themselves.

Malcolm’s not really a fan of the term “AI-ready.” Of course, data quality still matters, but readiness was never a property your data has or lacks, and treating it as a yes-or-no measurement is why the phrase is ambiguous and nearly meaningless. The industry, he adds, has strong incentives to keep producing phrases exactly like it.

About Malcolm Hawker

Malcolm Hawker is the Chief Data Officer at Profisee, a master data management vendor, where most of his job is external thought leadership. He spent three years as a Gartner analyst covering MDM and governance, and before that held IT leadership and product roles including Distinguished Architect at Dun & Bradstreet. He hosts the CDO Matters podcast and wrote The Data Hero Playbook.^[1]

In this episode, Malcolm and I discuss:

- Why “AI-ready data” is an overloaded term that collapses a spectrum into a binary

- Why the cost of being wrong decides whether an AI use case can go into production

- The difference between defining what a customer means and knowing which customer record is real

- Why telling your CEO “garbage in, garbage out” is a career-limiting move

- The semantic pedantic feedback loop, and how data literacy went from nonexistent to a top-three problem in a single year

Watch the full conversation here:

Ready for what, exactly

Malcolm’s definition of AI-ready data is a practical one. Data is AI-ready when it supports the use case in front of it. It maps onto the oldest working definition of data quality, which is fitness for purpose, and it means the same customer table can be perfectly ready for one job and nowhere near ready for the one running beside it.

Large language models (LLMs) were trained on the open internet, which he describes as a complete cesspool when it comes to data quality, and they work well enough that the entire market reorganized itself around them. At the other end sit healthcare, legal, finance, and anything carrying compliance or audit obligations, where accuracy and consistency are not negotiable. Both of those are AI use cases, and no single quality standard covers both.

Malcolm reframes readiness as a question of consequences rather than a question about data. Send a digital coupon to the wrong person, and it costs you close to nothing, which means your data is probably ready today. Model a specific customer’s behavior, or decide what offer that person should see next, and you need a far higher standard. In Malcolm’s read, that mismatch is also why so many proofs of concept die: organizations take a probabilistic system and drop it into a business process that has always run on deterministic rules, where the output was predictable, auditable, and safely inside a governance policy.

AI models do not behave that way, and the reason they misbehave is rarely recoverable after the fact. Malcolm calls this the attribution problem, and no amount of retrieval patterns, grounding, or knowledge graph scaffolding fully removes it. You can reduce how often the system surprises you, but you can’t get to a place where you know why it did what it did every time.

“AI-ready data is data that supports a given use case. I mean, literally, that’s it.”

— Malcolm Hawker, Chief Data Officer, Profisee

What the semantic layer cannot fix

Semantic layers exist for a good reason, and Malcolm is careful not to argue against them. An LLM cannot look at your raw tables and say anything useful about your customers, because it does not know what a customer is in your business or how your customer table joins to your product table. What it wants is text. Josh Howard put the same point more bluntly on this show: your AI is dumb without your data. Layering definitions over the tables, which is roughly what the market means by a semantic layer, is how the tabular data becomes legible to it.

Knowledge engineers have carried a name for this split for decades. The T-box, short for terminology box, holds the definitions: what counts as a customer, how net revenue is calculated, and what a location is.^[2] Nearly every conversation in the market right now, whether it is labeled ontology, knowledge graph, or semantic layer, is a conversation about the T-box. Malcolm argues that the other half of the structure has gone mum. The A-box, the assertional box, holds the actual values, and it answers a different kind of question: is Malcolm Hawker a customer, and if five Malcolm Hawker records exist in the system, which one is him?

Identity resolution, master data management, and old-fashioned data quality all live in the A-box, and no quantity of definitions reaches them. A perfectly specified ontology will tell an AI agent exactly what a customer is while remaining completely silent about which of your customer records is the real one. On this show, Matt Hayes made a related argument that AI-ready data clears a much higher bar than analytics-ready data, and Sam Pierson described the industry work to standardize semantic definitions so meaning travels with the data. Malcolm is adding the layer underneath both, where meaning gets you part of the way, and identity gets you the rest.

“If you’ve got 15 different versions of David Sweenor, all the context in the world isn’t going to solve that problem.”

— Malcolm Hawker, Chief Data Officer, Profisee

Identity problems are also harder to catch than the failures data teams are trained to look for. People can see a broken dashboard because they notice a number is off. Duplicate customer records produce answers that look entirely reasonable while resting on the wrong person, so the system returns them with total confidence, and nobody has a reason to check.

Never say “garbage in, garbage out” to your CEO

Garbage in, garbage out is probably the most repeated line in data management and analytics. The phrase flattens a spectrum into a binary, the same way “AI-ready” does, and it writes off data that might be entirely sufficient for the job at hand. That alone would make it sloppy, and Brendan Grady made the companion argument on this show about why bad data didn’t matter until now, which is that analytics absorbed errors AI will not. What makes the phrase dangerous is what happens when a data leader says it out loud in front of the business.

“First of all, the whole concept of garbage in, garbage out gives me hives.”

— Malcolm Hawker, Chief Data Officer, Profisee

Malcolm puts it in a boardroom. You are the CDO, and your CEO is holding a report with two David Sweenors on it and wants to know why there are two of a customer you both know is one person. No data leader answers that question with “sorry, boss, garbage in, garbage out” and keeps their standing in the room. I mentioned to Malcolm that this is career-limiting, because the phrase disowns the entire investment. In that one sentence, the warehouse, the pipelines, and every governance and quality platform on the books are declared powerless against whatever arrived upstream. It hands away the argument that data work changes outcomes, which is the entire argument a CDO is employed to make.

The semantic pedantic feedback loop

Malcolm first posted the most useful idea in this conversation on LinkedIn as a joke, then could not stop finding evidence for it. He calls it the semantic pedantic feedback loop, and it explains why our field keeps inventing new names for things that already had names. I knew to ask about it because Scott Taylor, who came on this show to argue for putting truth before meaning, told me I had to.

It runs in a circle. Analysts, consultants, and thought leaders need to be seen saying something new, and in the analyst business there is a subscription renewal attached to that need, because nobody renews to hear that best practice has not changed much. So new vocabulary gets manufactured. Vendors pick it up and build campaigns around it. Buyers start hearing it from every vendor and at every conference, decide it must be important, and call their analyst to ask what it is. The analyst hears their own coinage echoed back by the whole market and confirms that it is real. Around it goes.

Malcolm’s evidence is a receipt from inside Gartner itself. By his account, Gartner’s annual CDO survey asks respondents to name their top roadblocks, and in 2017 no respondent named data literacy, because it was not among the options. Gartner added it the following year, and it immediately ranked third. He is pointed about why that particular option travels so well: it locates the obstacle in everyone else’s knowledge. The impediment sits with the people being served rather than with the tools, the dashboards, or the data team itself. As he puts it, the thing had a name before it got a new one, and the name was training.

Let me say where I am standing while I write this. I host a podcast, I have written books, and a decade of my career went into product marketing for data platforms. That puts me inside the loop being described, along with the phrase that opened this article. “AI-ready” is the loop’s most recent output. Recognizing that buys you one useful habit: every time a new term arrives, ask what it was called last time and whether anything underneath it changed.

What the good ones do differently

Malcolm has an unusually large sample to draw on when he describes what separates the data leaders who succeed. Gartner analysts take client calls they call inquiries, and he did roughly 1,500 of them in three years with CDOs, CIOs, and VPs of data and analytics. In struggling organizations, he heard the same attribution pattern every time. They blamed data literacy, or culture, or the fact that nobody showed up to the governance committee meeting. The blocker was always somewhere else.

Leaders getting lasting results in governance, data quality, and MDM sounded different. They expected to be wrong the first time and to learn from it. They measured themselves on whether the people consuming their data products were successful, and they asked for feedback often enough to find out when they were not. That argument is what The Data Hero Playbook is built on, and the book’s premise is that the limiting factor in most data organizations is a set of beliefs rather than a missing capability.^[3]

So the next time someone asks whether your data is AI-ready, the useful response is that the question is incomplete. Ask what the use case is, because readiness is meaningless without one. Ask what being wrong costs, because that number decides whether the use case can go into production. Then ask whether you can tell which record is the real one, because if you cannot, no amount of context is going to rescue the answer.

Listen to the full conversation with Malcolm Hawker on his Data Faces Podcast episode page.

Based on insights from Malcolm Hawker, Chief Data Officer at Profisee, featured on the Data Faces Podcast.

Share

Leave a comment

Podcast highlights

- [0:06] Introduction and welcome to the Data Faces Podcast

- [1:18] From IT operations to Distinguished Architect at Dun & Bradstreet

- [2:20] Three years at Gartner, and what an externally facing CDO actually does

- [4:22] The record label he never started

- [5:09] A graduate thesis on Napster, and the case for buying a turntable again

- [7:02] What does “AI-ready data” even mean?

- [7:46] Fitness for purpose, and why the term is overloaded

- [8:05] ChatGPT was trained on the internet and works anyway

- [11:56] The cost of being wrong, and why a coupon is not a claims decision

- [13:10] Semantic layers, the T-box, and the fifteen David Sweenors

- [14:51] Does “garbage in, garbage out” mean anything for unstructured data?

- [15:44] Why the phrase gives Malcolm hives

- [15:56] The boardroom scene, and why the answer is career-limiting

- [17:02] Chunking text destroys the context that made it useful

- [19:26] An underserved market, and back to the semantic pedantic cycle

- [19:57] The feedback loop: pundits, vendors, buyers, analysts

- [23:36] How data literacy became a top-three CDO roadblock in one year

- [24:09] “There was a word for it, and it was called training”

- [25:17] Does master data management apply to unstructured data?

- [25:59] Master data as the connective tissue across business processes

- [28:03] Profiling, entity linkage, and the fox watching the henhouse

- [29:32] Contracts and forms are tractable; email is the hard problem

- [29:54] The Data Hero Playbook, and whether CDOs play defense

- [31:21] Growth mindset versus fixed mindset, drawn from 1,500 Gartner inquiries

- [35:18] What Malcolm is reading

- [37:29] Where to find Malcolm

Frequently asked questions

What does “AI-ready data” mean?

Data is AI-ready when it supports the specific use case you intend to run on it. There is no universal threshold, because fitness depends entirely on the job. Malcolm Hawker, Chief Data Officer at Profisee, argues the term has become overloaded precisely because the market treats it as a binary state your data either has or lacks. The same customer data can be sufficient for a marketing recommendation and unacceptable for a clinical or compliance decision, so always ask “ready for what.”

How is AI-ready data different from analytics-ready data?

Analytics-ready data supports human analysts reading dashboards and reports, where a person applies judgment and notices when a number looks wrong. AI-ready data feeds systems that act on the data without that check, so errors propagate at machine speed with nobody in the loop to catch them. The accuracy, consistency, and identity requirements are correspondingly higher, and the bar rises further as the cost of being wrong rises.

Does a semantic layer or knowledge graph solve AI data quality?

Only partly. Semantic layers, ontologies, and knowledge graphs define what your terms mean, which knowledge engineers call the T-box. They do not establish which specific records are true, which is the A-box. If your systems hold fifteen versions of the same customer, a perfect set of definitions will still not tell an AI agent which one is the real person. Identity resolution and master data management address that half of the problem.

Why do so many AI proofs of concept fail?

A recurring reason is that organizations deploy probabilistic systems into business processes that historically ran on deterministic rules. Those processes were built on outputs that were predictable, auditable, and safely inside a governance policy. Language models do not offer that guarantee, and what practitioners call the attribution problem means you often cannot determine why the system produced a given answer, which makes the failure difficult to diagnose or defend.

How do you apply data quality to unstructured data?

Nobody has solved this well yet. Traditional quality rules need small, structured units, so the standard move is to break text into chunks, but chunking strips away the context that made the text useful to a language model. Practical progress starts with profiling, identifying which documents reference your master data objects such as customers, suppliers, or products, and then determining whether the claims in those documents are true. That last step remains the hardest.

What should a data leader do instead of asking “is our data AI-ready?”

Replace it with three questions. What is the use case, since readiness has no meaning without one. What does being wrong cost, since that determines whether the use case can go into production at all. And can you identify which record is the real one, since context and definitions cannot compensate for unresolved identity. Answering those tells you far more than any readiness score.

About David Sweenor

David Sweenor is the founder and host of the Data Faces podcast, where he talks with the people who are making data, analytics, AI, and marketing work in the real world. He is also the founder of TinyTechGuides and a recognized top 10 AI thought leader and international speaker who specializes in practical business applications of artificial intelligence and advanced analytics.

With over 25 years of hands-on experience implementing AI and analytics solutions, David has supported organizations including Alation, Alteryx, TIBCO, SAS, IBM, Dell, and Quest. His work spans marketing leadership, analytics implementation, and specialized expertise in AI, machine learning, data science, IoT, and business intelligence. David holds several patents and consistently delivers insights that bridge technical capabilities with business value.

Books

- Artificial Intelligence: An Executive Guide to Make AI Work for Your Business

- Generative AI Business Applications: An Executive Guide with Real-Life Examples and Case Studies

- The Generative AI Practitioner’s Guide: How to Apply LLM Patterns for Enterprise Applications

- The CIO’s Guide to Adopting Generative AI: Five Keys to Success

- Modern B2B Marketing: A Practitioner’s Guide to Marketing Excellence

- The PMM’s Prompt Playbook: Mastering Generative AI for B2B Marketing Success

Follow David on Twitter @DavidSweenor and connect with him on LinkedIn.

Footnotes

^[1]Hawker, Malcolm. The Data Hero Playbook: Developing Your Data Leadership Superpowers. Hoboken, NJ: Wiley, 2026. https://www.wiley.com/en-us/The+Data+Hero+Playbook%3A+Developing+Your+Data+Leadership+Superpowers-p-9781394310647.

^[2]Baader, Franz, and Werner Nutt. “Basic Description Logics.” In The Description Logic Handbook: Theory, Implementation and Applications. Cambridge University Press, 2003. https://www.inf.unibz.it/~franconi/dl/course/dlhb/dlhb-02.pdf.

^[3]Hawker, Malcolm. The Data Hero Playbook: Developing Your Data Leadership Superpowers. Hoboken, NJ: Wiley, 2026. https://www.wiley.com/en-us/The+Data+Hero+Playbook%3A+Developing+Your+Data+Leadership+Superpowers-p-9781394310647.

Agentic AI was the pervasive topic on the floor this year. Everyone wants agents, and almost nobody has the data plumbing to feed their voracious appetites. So what exactly are all these agents supposed to run on? That question, the context problem, showed up in nearly every conversation about where AI is headed.

Stewart Bond, VP of Data Intelligence Software at IDC, came off his session with a question that pokes at a decade of investment. Is the data lake already obsolete? He argues that agentic AI favors federated architectures over centralized ones, because agents need data in real time where it lives, and latency is the enemy of agentic AI. He walked me through IDC’s model lake convergence idea, the forecast that the model itself becomes the analytical layer by 2029, which pushes data quality, security, and governance closer to the source and turns policy into code that runs on every access rather than a committee that approves it later.

Kevin Petrie, VP of Research at BARC US, returned to the show with a number that should stop any agent project cold. In BARC’s survey, 70 percent of companies said less than half of their unstructured data is currently discoverable and usable for AI. He compares enterprise unstructured data to his kids’ Lego bin, real value buried in a free-for-all, and he makes the case that five to fifteen well-governed agents beat 70,000 ungoverned ones. He also gave the fear a name that landed with the whole room, vibe slop, the mess you get when teams deploy agents on shaky foundations, and he expects agent cleanup and consolidation to become next year’s project.

Amy Lenander and Christina Egea of Capital One showed what solving the context problem actually looks like at scale. Amy is Capital One’s Chief Data Officer and Christina is SVP of Enterprise Data, and their answer is to treat data like a product. A usage analysis revealed that nine categories of data covered most of the company’s needs, and every data product gets a single accountable owner. Those products now launch use cases three times faster and cut the cost of maintaining standardized data by 30 percent. Christina was refreshingly honest about the hardest part, which is getting started, and the constant tension between building fast for one use case and building something that scales to a hundred.

Amin Venjara, chief data and product officer at ADP, gave the framework that made the context problem concrete. Value equals data plus capabilities, and his team runs ADP’s internal data platform like a product that builders across the company choose to use rather than a service they are forced to accept. He described the annual Data and AI Day that drew 2,100 people and the hackathon that feeds it, and a data product that hands chat and agent experiences full customer context, so no team rebuilds the same stitching work twice.

Three more conversations rounded out the picture on agents and context.

• Douglas Laney, Infonomics: The founder of infonomics predicts we will see the first billion dollar business run by a handful of people and a swarm of AI agents. He also warned that traditional data quality dimensions do not translate to unstructured data and AI slop, and that token costs are already making some organizations rethink swapping cheap labor for expensive inference.

• Terry Dorsey, Denodo: She draws a deliberate line between delivering data and delivering information, and argues AI succeeds when you treat it as one capability among many rather than one big encapsulated project. Data quality, in her words, sits in the eye of the consumer.

• Steven Moskowitz, Industry Forward: He works with manufacturers who own the AI, own the technology, and still get no value, because manufacturing runs on wide data, a richer mix of timestamps, images, CAD drawings, and equipment signals than most industries manage. His cognitive KPIs, like mean time to root cause, measure how AI changes thinking rather than how often people click it.

Where this leaves you

The teams making agents work had all done the same unglamorous work first. They moved data quality, governance, and context to where the agents actually reach for data before chasing the newest model. Everything an agent does downstream depends on that groundwork, and no amount of orchestration makes up for a foundation the agent cannot trust.

This is one theme of three. Head back to the full CDOIQ 2026 field notes for the complete roster, or read the other two themes on the chief data officer role at 20 years and data quality as the AI foundation. All 24 interviews are on the Data Faces Podcast on-location hub.

Share

Leave a comment

Frequently asked questions

What is the context problem in agentic AI?

The context problem is the mismatch between what AI agents need and what most data environments provide. Agents act in real time and need governed, discoverable data where it lives, but leaders at CDOIQ 2026 described most enterprise data as centralized, slow, and largely unstructured. BARC found that 70 percent of companies say less than half of their unstructured data is usable for AI, which means agents often have nothing trustworthy to act on.

Why do agents favor federated data architectures over data lakes?

According to IDC’s Stewart Bond, agents cannot wait for batch processes to move data into a central lake, because latency is the enemy of agentic AI. Federated architectures let agents reach data in real time where it already lives, which pushes data quality, security, and governance closer to the source. Bond expects the model itself to become the analytical layer, what IDC calls model lake convergence, by around 2029.

How should enterprises prepare their data for AI agents?

Leaders at CDOIQ 2026 recommended treating data like a product with clear ownership, moving governance to the point of data creation, and favoring a small number of well-governed agents over thousands of ungoverned ones. Capital One and ADP both described running internal data platforms as products, which gave chat and agent experiences the full context they need. Agents need trustworthy, real-time data they can act on without amplifying errors.

About David Sweenor

David Sweenor is the founder and host of the Data Faces podcast, where he talks with the people who are making data, analytics, AI, and marketing work in the real world. He is also the founder of TinyTechGuides and a recognized top 25 analytics thought leader and international speaker who specializes in practical business applications of artificial intelligence and advanced analytics.

With over 25 years of hands-on experience implementing AI and analytics solutions, David has supported organizations including Alation, Alteryx, TIBCO, SAS, IBM, Dell, and Quest. His work spans marketing leadership, analytics implementation, and specialized expertise in AI, machine learning, data science, IoT, and business intelligence. David holds several patents and consistently delivers insights that bridge technical capabilities with business value.

Books

• Artificial Intelligence: An Executive Guide to Make AI Work for Your Business

• Generative AI Business Applications: An Executive Guide with Real-Life Examples and Case Studies

• The Generative AI Practitioner’s Guide: How to Apply LLM Patterns for Enterprise Applications

• The CIO’s Guide to Adopting Generative AI: Five Keys to Success

• Modern B2B Marketing: A Practitioner’s Guide to Marketing Excellence

• The PMM’s Prompt Playbook: Mastering Generative AI for B2B Marketing Success

Follow David on Twitter @DavidSweenor and connect with him on LinkedIn.

Earlier this month, I was heads down working on a Gartner Magic Quadrant (MQ) submission for a client. Anyone who’s been through one knows what a pain in the butt they are. There’s an unreasonable deadline that doesn’t move, a long list of questions, and a pile of evidence that has to be assembled and written so that it survives a fast read from an analyst (or their LLM). I was running my Claude sidekick to help put the submission package together.

Then all of a sudden, wouldn’t you know, Claude stopped and handed me an overloaded message. I refreshed, waited a minute, tried again, and got the same thing, so I went to status.claude.com and found a system outage posted there. Well, that was a bummer, and it wasn’t the first time I’d seen the overloaded error.

So what do you do? Do you call it a day and grab a tasty beverage, email the client, or ask a stubborn analyst firm to move a date? This raises an interesting point because most of my clients and colleagues now heavily rely on Claude, ChatGPT, or Gemini for a significant part of their work.

While we’re here, if you want to know what I really think about the Gartner process itself, I wrote that up last year while helping a different client through their submission. They move their dates all the time, but we don’t have that luxury.

Can you survive an outage, or did you just run out of tokens?

Depending on how heavy a user you are, you may not have encountered an outage, and that’s fine, because the ordinary version has the same symptom and the same remedy. You blew through all of your tokens and hit a usage limit partway through your work, and the session tells you to come back in a few hours. For most, they’ll be just as stuck as I was. Whatever you were going to finish today, it’s not going to happen.

I’m a huge fan and an advocate of Claude, and anyone who’s been following this series already knows it. I moved my marketing work to Claude Code more than a year ago. I’ve written a stack of articles about how to run a marketing practice on it, and I still use it most days for most things. Nothing here changes any of that, and none of this is a slam against Anthropic.

This may surprise you, but the specific LLM matters a lot less than you’d think for most tasks. Sure, they each have their own quirks, and they’re better and worse at different things, but with the same rules and workflows, any of them will get the job done. Your work either survives the LLM going away for an afternoon or it doesn’t, and you find out which one is true on the day you can least afford to find out.

What broke, and what didn’t

Despite wanting that tasty beverage, I didn’t have to wait. I simply switched over to Codex, which is OpenAI’s command-line tool, and carried on in the same folder, on the same submission, in the same terminal window. It could just as easily have been Gemini CLI. I delivered the submission on time.

The switch was relatively painless. The one thing that changed was how I start a job: Codex doesn’t take a slash command, so instead of typing /edit-video, I asked Codex to use /edit-video. Same instructions, same result.

Everything else sat exactly where I left it. The project rulebook was a file on disk. The 55 workflow files in my TinyTechGuides folder, and another 31 in my consulting folder, were files on disk. The 124 memory facts were files on disk. So were the 1,056 tracked files of actual client work, research, transcripts, and meeting notes. The connections into Gmail, Sheets, Analytics, and Buffer were still running and still authenticated. None of it lived in a vendor’s database, and none of it needed anybody’s permission to open.

That’s when something I’ve been advocating for six months finally clicked. I’ve written before that the memory is the asset, which is true and still undersells it. The asset is the whole system you set up, and the memory is only one piece of it. Your rules, your workflows, your memory, and your live connections to other tools are all files, which is exactly why none of them went anywhere. What I’ve been calling my AI system (or Marketing OS) is mostly a folder of markdown files in a git repository, and the AI is the part I rent monthly.

“Your rules, your workflows, your memory, and your connections are the asset, and every one of them is a file.”

So take a minute and count your own. Prompts that live in a chat history, brand rules that live in a conversation you had back in March, and a process that lives in your head all leave with the session. Files stay.

Running the same system on a different engine

My setup is deliberately boring, and the picture below is the whole of it. I keep the file tree on the left, a markdown preview on the top right, and a terminal along the bottom where the agent actually runs. The editor isn’t doing anything clever. The only thing it buys me is being able to watch my files while something else works on them, which matters more than it sounds like it should.

Adding a second engine takes two commands.

npm install -g @openai/codex

npm install -g @google/gemini-cli

That gives you Codex from OpenAI and the Gemini command-line tool from Google. Both run in the same terminal, in the same folder, against the same files. From there, you have three things to point at them, and only three.

• Context. Codex looks for a file called AGENTS.md instead of CLAUDE.md. Don’t copy the file over, because then you have two rulebooks drifting apart, which is the same one-source-of-truth mistake I made with my content calendar. Ask Codex to make AGENTS.md a symlink to CLAUDE.md, which is a pointer to the original rather than a copy of it, and it will set that up for you. You end up with one file that answers to two names. Gemini is easier still, since you can point it straight at CLAUDE.md in its settings file.

• Workflows. Your workflow files are markdown, which is the part most people miss. Codex and Gemini don’t take /edit-video as a command, but ask either one to use /edit-video and it goes and finds .claude/commands/edit-video.md and follows what it says, because that file is a set of instructions and nothing more. You lose the keyboard shortcut, and you keep the work.

• Connections. This was the only piece that needed real work. Your live connections into Gmail, Sheets, and the rest have to be registered once per tool, and the formats differ, since Codex keeps them in one global config while Gemini keeps them per project. I didn’t write any of that by hand. I asked Codex to read the configuration I already had and write its own version, and it did.

Get one thing right while you’re in there. Keep your tokens in environment variables rather than pasting them into a config file that ends up in your repository. Codex enforces this and refuses a literal token, which is the right call. Gemini will happily let you do the sloppy version, so don’t take it up on the offer.

All in, this is about five minutes, once. The models will do most of it for you if you tell them what you want.

I made this argument to CIOs two years ago

Back in January 2024, I wrote the CIO’s guide to generative AI vendor selection, and the second question I told technology leaders to ask a vendor was who owns and controls the data. Those readers had procurement teams and security reviews standing behind them. Marketing has to answer the same question for itself now, because nobody in IT is answering it on our behalf.

If you’ve been reading this series, you’ve already done most of the work without my telling you why. Over six articles, I’ve argued for putting your rules in a project file, turning your prompt workflows into skills, keeping a real memory, and wiring live connections into the tools you already use. Not one of those is a Claude feature. Each one’s a file, and that’s exactly why they were still there on an afternoon when the model wasn’t.

Where to start

None of this is a prediction about Claude. Outages happen to every provider, and usage limits catch everybody on a heavy week. That is the deal when you rent your intelligence by the month, and it isn’t going to change. Everything comes down to one question. Does your working life have a single point of failure sitting in the middle of it? Most marketing teams have never stopped to ask.

So ask it. Picture your primary model going dark for six hours tomorrow, right in the middle of the one thing you can’t move. What would you still be able to do? If the answer is not much, a different vendor won’t save you. Move your rules, your workflows, and your accumulated context out of a chat window and into files you own, then spend the five minutes it takes to show a second engine where to find them.

Own the work. Create the process. Set up the system. Rent the model.

Share

Leave a comment

Frequently asked questions

What does it mean to future-proof a marketing workflow?

Future-proofing a marketing workflow means keeping the parts of your work that last in files you own rather than inside one vendor’s product. Your project rules, workflow instructions, accumulated memory, and client work all sit as plain markdown in a folder on your disk. An AI model reads those files and acts on them, so when one model is unavailable, you point a different one at the same folder and carry on. The model becomes the part you rent, and the work stays the part you own.

Do I have to stop using Claude to do this?

No, and that is the point. Setting up a second engine changes nothing about how you work day to day. I still use Claude Code most days for most things, and Codex and Gemini sit there unused until I need them. Installing them costs about five minutes once, and the files they read are the same files Claude already reads. It is a spare tire, not a new car.

How is this different from paying for a second AI subscription?

A second subscription gives you a different chatbot and does nothing for the work you have already built up. When your prompts, brand rules, and processes live inside one vendor’s chat history, a second subscription drops you into a new empty window with none of it. Moving that accumulated context into files first is what makes any model useful to you. The subscription is the cheap part, and the portable context is what carries over.

Will my Claude skills run in Codex or Gemini?

Yes, with one difference in how you start them. Claude Code lets you type a slash command such as /edit-video. In Codex and Gemini, you ask the model to use /edit-video instead, and it goes and finds the markdown file behind that command and follows what it says—same instructions, same output. You give up the keyboard shortcut, and you keep the workflow.

How long does the setup take?

About five minutes, once. Two npm commands install Codex and the Gemini command-line tool. After that, you need to connect three things. A symlink lets Codex read the project rules you already wrote, your workflow folder needs no change at all, and your live tool connections get registered once per tool. Ask the model to handle the configuration, and it will read what you have and write its own version.

Where do I start if all my prompts live in a chat history?

Start by moving them out. Open a folder for the project, put your standing rules in one file, and save each repeated prompt as its own markdown file with a name that says what it does. That alone gets the work out of a vendor’s database and onto your disk, which is what matters. Adding a second model afterward takes minutes, because the hard work of writing things down is already behind you.

About David Sweenor

David Sweenor is the founder and host of the Data Faces podcast, where he talks with the people who are making data, analytics, AI, and marketing work in the real world. He is also the founder of TinyTechGuides and a recognized top 10 AI thought leader and international speaker who specializes in practical business applications of artificial intelligence and advanced analytics.

With over 25 years of hands-on experience implementing AI and analytics solutions, David has supported organizations including Alation, Alteryx, TIBCO, SAS, IBM, Dell, and Quest. His work spans marketing leadership, analytics implementation, and specialized expertise in AI, machine learning, data science, IoT, and business intelligence. David holds several patents and consistently delivers insights that bridge technical capabilities with business value.

Books

- Artificial Intelligence: An Executive Guide to Make AI Work for Your Business

- Generative AI Business Applications: An Executive Guide with Real-Life Examples and Case Studies

- The Generative AI Practitioner’s Guide: How to Apply LLM Patterns for Enterprise Applications

- The CIO’s Guide to Adopting Generative AI: Five Keys to Success

- Modern B2B Marketing: A Practitioner’s Guide to Marketing Excellence

- The PMM’s Prompt Playbook: Mastering Generative AI for B2B Marketing Success

Follow David on Twitter @DavidSweenor and connect with him on LinkedIn.

What decides whether anyone uses the data product your team just built? When a data team finishes one, the quality is usually better than the previous generation, but adoption stays flat anyway. What do you do? Typically, the team responds by evangelizing harder and escalating to whoever runs the reluctant business unit. Amy Lenander, Chief Data Officer at Capital One, thinks all of that arrives too late to help, because the decision that governs adoption got made long before the build ever started.

I caught Amy and her colleague Christina Egea, SVP of Enterprise Data, at the CDOIQ Symposium in Cambridge, Massachusetts, and pulled them in front of a Data Faces microphone.^[1] Their session, “Doubling down on data products to drive business value,” followed the talk Amy gave the previous year on Capital One’s broader data transformation. Data products drew more interest than anything else that year, so they double-clicked on the piece of their strategy that peers find least familiar: Capital One curates a deliberately small set of data products instead of letting every domain publish its own.

“One of the things that we do that I think is most unique among the companies I’ve talked to is being very intentional about curating the data we want to manage as data products.”

— Amy Lenander, Chief Data Officer, Capital One

That restraint is a design decision rather than a resourcing constraint, and it traces back to a question Christina’s team asks before building anything: will this still hold up at the hundredth use case?

About Amy Lenander and Christina Egea

Amy Lenander became Chief Data Officer at Capital One in February 2023, capping two decades at a company she joined as an analyst who made decisions from however much data would fit in a spreadsheet.^[2] She has led loyalty and rewards, served as CEO of Capital One’s UK business, and worked across credit card and lending. Because she came to the data seat from the business rather than from engineering, Amy has what most Chief Data Officers (CDOs) acquire secondhand: a working sense of where the real impact in each business sits.

Christina Egea is SVP of Enterprise Data on Amy’s team, and she has spent three years building Capital One’s data product strategy. Their data ecosystem keeps growing by acquisition — Capital One completed its purchase of Discover in May 2025 and its purchase of Brex in April 2026 — so any strategy worth adopting has to survive a footprint that expands faster than a roadmap.^[3][4]

In this episode, Amy, Christina, and I discuss:

- Why the fastest data product to build is usually the one that helps least

- How a study of company-wide usage collapsed a sprawling data landscape into nine categories

- The case for one accountable owner on every data product

- Why a data product covering 70% of what someone needs will never displace the feed covering all of it

- What three times faster time to market rests on

Watch the full conversation here:

The hundredth use case is the design decision

Christina Egea describes the tension at the center of Capital One’s data product strategy as pressure from two directions at once. One side of the company asks why the team cannot reach the market more quickly, while the other asks how anything built that fast could be correct, reusable, or safe for customers. Both questions are reasonable, and a data organization that answers only one of them ends up with either a backlog or a mess. Christina resolves it by being explicit about which use case her team is building for.

“The fastest thing to build is the thing for a single use case. The harder thing to build is the thing that will scale.”

— Christina Egea, SVP of Enterprise Data, Capital One

Capital One aims well past the first requester. Christina told me her team focuses on making sure a data product will scale well past a single use case, past the next five or twenty, out to something closer to the hundredth use case and the hundredth customer. A specific use case still gets most data products launched, because someone has to want the thing badly enough now to fund it, but the design target sits far beyond whoever asked first. A data product scoped to one requester can hard-code their assumptions and skip the modeling that would let another team use it next quarter.

Christina also names the hardest part of the whole endeavor, and it is not the part most people guess. Getting started is what hurts. The payoff arrives on the far side of that first set of data products, once teams have adopted the same product five or twenty times and the flywheel begins to turn on its own.

Nine categories covered most of the usage

Capital One did not begin its data product work by drawing a domain map or carving the catalog along org-chart lines. Christina’s team studied how data was already being used across the company, asking which core categories and domains that usage fell into. Despite enormous breadth and depth across the business, roughly nine data categories accounted for most of the consumption, and those nine became the first data products the company built.

Measured usage is what makes curation possible, and curation is where Amy locates Capital One’s real point of difference. Two kinds of data qualify in her model. The first is the most-used data in the company, which earns its place because so many lines of business reuse it. The second is data the company believes carries real value but nobody has drawn on much, because getting to it has been too hard. Both are chosen for leverage rather than volume.

“We’ve very intentionally started with the data products that have the most leverage to the company.”

— Amy Lenander, Chief Data Officer, Capital One

Amy’s read on where that impact concentrates comes from having run the businesses she now serves, and the endless list of things a data team could improve becomes tractable once you know which ones the business genuinely needs. Her background also helps with a less comfortable part of the job, because her team has to get business units to do things they would rather not, which she compares to eating your vegetables.

The top-down start was only the opening move. Each business unit has since built a bottom-up view of the data products it needs, while Christina’s central team maintains the framework that holds the picture together.

One owner, and no overlap

Every data product at Capital One has one accountable owner who makes the final decisions on it. Amy is direct about why that matters at her company’s scale: one owner per product, responsible for the decisions and for making sure the products work together well. Plenty of the building is distributed across business units. Without a single named owner, two teams publish their own version of the customer table, and nobody has the authority to say which one is right.

Christina’s central team spends a meaningful share of its time on the shape of the portfolio, working toward a universe of data products that is mutually exclusive, where the pieces do not overlap, and each product’s scope is defined.

“Really making sure that we’re investing in a universe of data products that’s mutually exclusive, that the things don’t overlap, that they clearly link together but have boundaries is a place where my team spends a bunch of our time.”

— Christina Egea, SVP of Enterprise Data, Capital One

From there, Christina’s team scopes each data product and names its owner, then models the data into an ontology before building the assets that bring it to life, including the application programming interfaces (APIs) and tables consumers touch. Settling semantics in that ontology first is what keeps the hundredth use case reachable, because the definitions get negotiated once rather than relitigated by every team that shows up later.

Adoption is a listening problem

Asked what to do when a data product is demonstrably better and nobody uses it, Amy does not reach for another enablement session. She puts the question to the people who have not switched, and then does the harder part, which is hearing what they say.

“The first thing is to listen to the answer when you actually ask that question, to stop pushing and start listening.”

— Amy Lenander, Chief Data Officer, Capital One

The answers are rarely mysterious once someone asks. Switching from a working data feed to a new data product is painful, and the person being asked to move already has something that produces numbers today. Winning a brand-new use case is far easier than displacing an incumbent feed, because displacement forces the consumer to absorb migration work on top of their actual job. They may agree the data product is better and still wonder whether it is worth the trouble.

The coverage trap is the same principle seen from the customer’s side. A consumer might tell Amy’s team that the data product covers 70% of the data they need, and 70% will not displace a source that covers all of it, no matter how much better the newer thing is. Her response is to listen and go build the missing 30%, because a data product meant for broad reuse has to cover the ground its consumers stand on.

Amy frames the discipline in product management terms, which carries a warning for data teams that have never had to win a customer. Nobody churns in a way that shows up on a dashboard, so consumers keep using the old data feed and say nothing, and Amy reads that silence as information about the product.

“They don’t pay for your data, but they vote with their feet. So you need to really deeply understand them like a good product manager would.”

— Amy Lenander, Chief Data Officer, Capital One

The numbers stand on something

Christina is careful about how she presents Capital One’s results, and the caveat she leads with is more instructive than the figures that follow. The company’s data product investments stand on the shoulders of its foundational ones, she says — a core set of platforms, catalogs, and an integrated lake.

With that foundation in place, the returns are specific. Capital One sees roughly three times faster time to market for use cases that launch on data products, measured against how the same work went before, and about 30% lower cost to maintain data once it lives in a standardized data product held to a higher quality bar.

For anyone benchmarking against those numbers, the sequence matters more than the percentages. Capital One measured its usage before picking its first data products and curated for leverage instead of coverage, and only then named an owner for each one and settled the semantics before building anything on top. The three times and the 30% arrived after all of that, which makes them the result of the strategy rather than the argument for starting it.

Watch the full conversation and read the transcript on the Data Faces CDOIQ page for Capital One.

Based on insights from Amy Lenander, Chief Data Officer, and Christina Egea, SVP of Enterprise Data, at Capital One, featured on the Data Faces Podcast.

Share

Leave a comment

Podcast highlights

- [0:39] David introduces Amy Lenander and Christina Egea on location at the CDOIQ Symposium in Cambridge, Massachusetts

- [2:11] Amy on the session title, “Doubling down on data products to drive business value,” and why data products drew the most interest at the prior year’s talk

- [2:54] Christina on top-down buy-in, data maturity, and building data product frameworks with the right ownership

- [4:14] Amy on curating data products by leverage, and why it is the most unusual thing Capital One does

- [5:22] Christina on building for the hundredth use case rather than the first

- [6:26] Christina on the nine categories of data that covered most company usage

- [7:55] Amy on coming to the CDO seat from the business, and getting partners to eat their vegetables

- [9:51] Christina on the hardest part of scaling data products

- [11:18] Amy on adoption as a listening problem, the 70/30 coverage trap, and voting with their feet

- [13:13] Christina on three times faster time to market and 30% lower maintenance cost

About David Sweenor

David Sweenor is the founder of TinyTechGuides and host of the Data Faces Podcast. He is an international speaker, advisor, and the author of eleven books on artificial intelligence, analytics, and B2B marketing, including Generative AI Business Applications, The CIO’s Guide to Adopting Generative AI, and Modern B2B Marketing. With more than twenty-five years in analytics and AI at companies including Alteryx, Tableau, TIBCO, SAS, IBM, and Dell, David advises technology companies on product marketing, content strategy, and go-to-market execution. He holds several patents and has been named a top influencer in data and analytics by Onalytica, Thinkers360, and Analytics Insight.

Connect with David on LinkedIn and subscribe to the Data Faces Podcast for conversations with the people shaping enterprise data and AI.

Footnotes

^[1]CDOIQ Symposium. “The 20th Annual CDOIQ Symposium.” July 21–23, 2026, Hyatt Regency Cambridge, Massachusetts.

https://2026cdoiq.org/

^[2]Capital One. “Interview with Amy Lenander, Chief Data Officer.” Capital One Tech. https://www.capitalone.com/tech/culture/amy-lenander-chief-data-officer-interview/

^[3]Capital One Financial Corp. “Capital One Completes Acquisition of Discover.” May 18, 2025. https://investor.capitalone.com/news-releases/news-release-details/capital-one-completes-acquisition-discover

^[4]Capital One. “Capital One Completes Acquisition of Brex.” April 7, 2026. https://www.capitalone.com/about/newsroom/capital-one-completes-acquisition-of-brex/

Somewhere in almost every data organization, an eager executive has asked questions about the usability of the organization’s data. You’ve ingested the data, put it neatly all in one place, and built the lake. So, can we use it now? Amin Venjara, Chief Data and Product Officer at ADP, hears that question constantly, and he says data sitting in a data swamp was never the point. The value shows up later, in the work you do after capture.

I caught Amin’s session at the CDOIQ Symposium in Cambridge, Massachusetts, and then pulled him in front of a Data Faces Podcast mic. His talk was called “Value Equals Data and Capabilities: a field guide to product-led data platforms,” and the equation in that title is the whole argument. Data is one term. Capabilities are the other. Executives fund the first and forget the second, then wonder why the picture never comes together.

“Data plus capabilities equals the value.”

— Amin Venjara, Chief Data and Product Officer, ADP

About Amin Venjara

Amin Venjara is the Chief Data and Product Officer at ADP, where he oversees the product portfolio and the data that differentiates it, from the smallest nail salon and tire center up to Fortune 500 multinationals. Amin told me the company operates in 140 countries, pays close to 20% of the U.S. working population, more than 40 million people a month, serves over a million clients, and moves roughly $3 trillion a year once you count taxes and government filings.¹ He fell into data through an unlikely door. He grew up a New York Mets fan who dreamed of being a sports broadcaster, played trumpet, and loved math, and he found the thread that ties it all together: data work is technology, people, and storytelling at once.

In this episode, Amin and I discuss:

• Why “it’s all in one place” is not the same as value

• What a data capability is, from the semantic layer to entity resolution

• The contractor who talks about wood while the customer wants a backyard

• How ADP runs its data platform like a product nobody is forced to use

• The one metric that shows whether the data team is earning its keep

Watch the full conversation here:

What counts as a data capability

What’s a data capability? A data capability is the engineering work that turns co-located data into something a business can use, and Amin is precise about what that means. He points to a semantic layer, so that revenue means one thing everywhere and a headcount number resolves the same way in every report. He points to an entity resolution graph, so that the same customer lines up across sales, service, finance, and support. He points to the stitching that takes four systems full of data about a client and produces one coherent picture of that client. None of that comes free when the data lands in the same place.

“Just because it’s in the same place, can you extract the value? No. There’s a lot of work that has to happen, these capabilities.”

— Amin Venjara, Chief Data and Product Officer, ADP

That distinction is where Amin says his own organization turned a corner. Once his team could separate the data from the capabilities, they could finally have a direct conversation with the business about what it takes to create value, because the business could see that the work wasn’t finished the moment the ingestion pipeline ran. Owning the ingredients is not the same as cooking the meal. You can have sales data, service data, customer data, and finance data in the same warehouse and still not know how to get the right answer for revenue or customer count, because the semantic definition that produces that answer is itself a capability you have to build. Data plus capabilities is not a slogan; it forces you to budget for the half of the work that otherwise goes unfunded.

Why the capability half never gets funded

If capabilities are half the value, why do so few organizations pay for them? Amin answers that data teams describe the wrong thing. They walk into a business conversation and talk about sources, ingestion, ETL pipelines, data quality, and lineage, all the machinery under the floor, and the executive on the other side of the table glazes over. To help me understand, he shared an analogy about contractors.

A contractor shows up and says he has the best wood, the strongest concrete, and the finest nails, and he can tell you exactly how many pounds the concrete will hold. The homeowner just wants to entertain people in the backyard. Should we build a deck? A patio-scape? That is the conversation the homeowner wants to have, and the contractor keeps selling lumber.

“This guy’s like, I got the best wood. It’s amazing concrete. And the guy’s like, I just want people to have a good time.”

— Amin Venjara, Chief Data and Product Officer, ADP

The materials matter, but the value has to come first, and Amin’s team took the translation on as their own job rather than waiting for the business to learn to speak data. Business and data speak different languages, and someone needs to translate. When his organization decided that translating value was part of the data team’s mandate, the capabilities stopped being line items the business kept cutting and became things it asked for by name. An unfunded capability is almost always a capability nobody managed to explain.

What a capability buys you in the AI era

The capabilities argument gets sharper the moment you point it at AI, because the AI era is where the cost of skipping capabilities compounds. Every company is racing to put a chat experience or an agent in front of its customers. Amin walks through what that requires. A customer engages inside a SaaS application. That application knows something about the customer, but so does the CRM, the financial system, the sales system, the service system, the ticketing system, and the call transcripts. Each of them holds a fragment. If the chat application has to reach into all of them and resolve the customer, reconcile the definitions, and handle the latency every single time, you have handed an enormous amount of repetitive work to every application you build.

Now consider the alternative. One data product has already normalized the data across those systems. It has resolved the customer IDs, it understands latency, it keeps the definitions right, and maintains the lineage and refresh cadence underneath. Your chat application calls that one data product and gets the full customer context. So does your agent, so does the next application, and the one after that.

“Imagine that when that customer engages, that application could call a data product that had already normalized data across these different systems so that it could call the appropriate context for that customer.”

— Amin Venjara, Chief Data and Product Officer, ADP

The capability layer, not the model, decides what your AI can do. An agent pointed at raw systems reinvents the stitching on every call, and it inherits every inconsistency those systems carry. A data product built once as a real capability gives every agent and every application the same clean context to stand on. When I said back to Amin that he had just articulated why you build data products at all – to stop reinventing the wheel – he agreed without hesitation. Everyone can see the model. Far fewer people fund the client-360 data product underneath it, and that data product is what determines whether the model has anything trustworthy to say.

Run it like a product, and prove it

None of this works if people are forced to use it, which is the most counterintuitive thing Amin has built into ADP’s data organization. His team treats the data platform like a product, and they mean it. You do not have to use us, they tell the rest of the company. We are going to make you want to use us. Just as you choose a favorite provider or a hyperscaler, internal teams get to choose the platform, and the platform has to earn that choice.

“You don’t have to use us. We are going to make you want to use us. You’re going to choose to use us.”

— Amin Venjara, Chief Data and Product Officer, ADP

The philosophy sounds soft until you see how they measure it. A mandate produces compliance and quiet resentment, while chosen adoption produces the kind of pull Amin can point to, like the company’s annual Data and AI Day, which in its second year drew about 2,100 people in person and online. That number is a demand signal, and his team tracks demand the way any product organization would. They watch three things: the outcome of each use case, since value is specific and one use case might be revenue while another is a count of active users; the raw usage, meaning how many users and active users the platform has; and efficiency, which is where Amin offered the most useful number I heard all day. ADP runs a hub-and-spoke model, where a central hub builds the shared capabilities, and the spoke teams take them the last mile to a business outcome. To measure whether the hub is pulling its weight, they divide what the spokes spend on compute and storage by the total platform spend. A higher spoke share of that cost means the hub is more efficient, because the central team enables more value at the edges than it consumes in the middle. Most data leaders cannot tell you a number like that about their own organization.

Build the capability, then the AI has somewhere to stand

Bring it back to the equation Amin started with. When the business cannot use the data, more data rarely fixes it, because the thing standing in the way is a capability nobody named, funded, or built: a semantic layer, an entity resolution graph, or a client-360 data product an agent can call. The organizations getting value from AI right now are the ones that did the unglamorous capability work first, so the model has something solid to stand on.

If you lead a data organization, the Monday-morning version of Amin’s argument is a short audit. Walk your list of value claims and mark which ones are data with a real capability behind them, and which ones are data sitting in a lake with the capability still unbuilt. Then run the second group like a product, and measure whether your hub gets more efficient as the spokes grow. That is a harder conversation than buying another tool, and it is the one that actually moves the needle.

Listen to the full conversation with Amin Venjara on the Data Faces Podcast.

Based on insights from Amin Venjara, Chief Data and Product Officer at ADP, featured on the Data Faces Podcast.

Share

Leave a comment

Podcast highlights

• [0:24] What Amin wanted to be growing up: a Mets fan’s dream of sports broadcasting, trumpet, and a love of math

• [2:19] What a Chief Data and Product Officer at ADP does, and the scale of the company

• [3:43] The CDOIQ session, and the premise that value equals data and capabilities

• [4:25] The contractor and the backyard, and why data teams lose the room

• [5:21] The half executives miss: capabilities, not just data

• [6:07] Semantic layer, entity resolution, and the stitching behind a client 360

• [7:11] Treating the data platform like a product nobody has to use

• [11:40] Two languages, and why translation is the data team’s job

• [12:34] The annual Data and AI Day, and 2,100 people showing up

• [13:16] The three things ADP measures: outcomes, users, and efficiency

• [16:27] A data product that gives every chat app and agent full customer context

• [17:25] Stop reinventing the wheel

Frequently asked questions

What does “value equals data plus capabilities” mean?

Amin Venjara’s formula holds that data alone does not create value; capabilities do. A capability is the engineering work that turns co-located data into something a business can use, such as a semantic layer that makes revenue mean one thing everywhere, or an entity resolution graph that lines the same customer up across systems. Executives tend to fund the data and forget the capabilities, then wonder why the business still cannot use what was ingested. The value comes from both terms together, not from the data alone.

What is a data capability?

A data capability is the specific engineering work that converts raw, co-located data into usable value. Amin Venjara points to a semantic layer that gives revenue, headcount, and customer count a single consistent definition, and an entity resolution graph that identifies the same customer across sales, service, finance, and support systems. The stitching that turns several systems into one coherent client 360 is a capability. Putting data in a lake does not produce these; a team must build them intentionally.

How is a data product different from a data lake?

A data lake is a place where raw data from many systems is collected. A data product is a curated, reusable asset built on top of that data, with customer IDs resolved, definitions standardized, and lineage and refresh cadence maintained. In Amin Venjara’s client-360 example, one data product normalizes customer data across systems so a chat application or agent can call it for full context, instead of every application reinventing that work. The lake stores the data; the data product serves it.

What is a product-led data platform?

A product-led data platform is run like a commercial product, with internal teams as customers who choose to use it rather than being forced to. At ADP, Amin Venjara’s team tells the company, “You don’t have to use us; we’re going to make you want to use us.” Adoption is earned through usefulness, and the platform team tracks outcomes, active users, and efficiency the way any product organization tracks demand. The approach replaces mandated compliance with genuine pull from the business.

How do you measure a data platform’s efficiency?

Amin Venjara’s team uses a hub-and-spoke metric: divide what the spoke teams spend on compute and storage by the total platform spend. A higher spoke share means the central hub is enabling more value at the edges than it consumes in the middle, which signals an efficient hub. ADP tracks this alongside use-case outcomes and active users. The metric gives data leaders a concrete way to show whether their central platform is compounding value rather than absorbing it.

About David Sweenor

David Sweenor is the founder and host of the Data Faces podcast, where he talks with the people who are making data, analytics, AI, and marketing work in the real world. He is also the founder of TinyTechGuides and a recognized top 25 analytics thought leader and international speaker who specializes in practical business applications of artificial intelligence and advanced analytics.

With over 25 years of hands-on experience implementing AI and analytics solutions, David has supported organizations including Alation, Alteryx, TIBCO, SAS, IBM, Dell, and Quest. His work spans marketing leadership, analytics implementation, and specialized expertise in AI, machine learning, data science, IoT, and business intelligence. David holds several patents and consistently delivers insights that bridge technical capabilities with business value.

Books

• Artificial Intelligence: An Executive Guide to Make AI Work for Your Business

• Generative AI Business Applications: An Executive Guide with Real-Life Examples and Case Studies

• The Generative AI Practitioner’s Guide: How to Apply LLM Patterns for Enterprise Applications

• The CIO’s Guide to Adopting Generative AI: Five Keys to Success

• Modern B2B Marketing: A Practitioner’s Guide to Marketing Excellence

• The PMM’s Prompt Playbook: Mastering Generative AI for B2B Marketing Success

Follow David on Twitter @DavidSweenor and connect with him on LinkedIn.

Footnotes

1. ADP. “ADP Corporate Overview.” Accessed August 2026. https://www.adp.com/-/media/corporate-overview/adp-corporate-overview.pdf. ADP publicly reports serving more than 1.1 million clients across 140+ countries and providing payroll to over 42 million workers, roughly one in six U.S. workers. The “$3 trillion a year” and “20% of the U.S. working population” figures are Amin Venjara’s characterizations on the podcast. ↩

▶ Watch the full episode and read the transcript: Sam Pierson on adaptable AI architecture

I have spent my career on both sides of the data platform business. In the first half, I built data warehouses and did data science and analytics at IBM. In the second half, I marketed data platforms at companies like SAS, TIBCO, Alteryx, and Alation. That means I have pitched freedom from vendor lock-in a number of times, and I have also worked with enough buyers to know how much that promise matters to them. So when a CTO tells me freedom is the way to win at AI, I want to know what is in the code and what is in the brochure.

Sam Pierson joined me on the Data Faces Podcast to discuss this notion. As Chief Technology Officer at Qlik, he owns the cloud and AI platforms as well as the development teams for both. As CTO of Talend, Sam joined Qlik as part of the acquisition, so he knows a thing or two about messy enterprise data. His argument runs counter to the instinct to pick a winner. The models will keep overtaking each other, the patterns will keep changing, and the AI conversation itself will look different in a year. What you can control is how quickly you can pivot, and most of that, along with your increasing AI bill, gets decided in the architecture before a model ever processes a token.

About Sam Pierson

Sam Pierson is the Chief Technology Officer at Qlik, where he leads engineering and spends much of his week talking with customers. Before Talend and Qlik, he held engineering leadership roles at Illuminate Education, Datica, SPS Commerce, Veritas, and Symantec. He holds an MBA and a computer science degree from the University of Minnesota.

In this episode, Sam and I discuss:

- Why AI is pushing enterprises to rebuild their data architectures on open formats like Apache Iceberg

- The model selector Qlik wired in early, and how it routes a task to a model that is good enough but 10 times cheaper

- Where AI token costs pile up before a model ever runs, and how a pre-calculated in-memory engine avoids them

- Whether freedom from vendor lock-in is a real architectural property or a story the industry tells itself

- Why 97 percent of enterprises have budgeted for agentic AI while only 18 percent have fully deployed it

Watch the full conversation here:

An architecture you can take apart

Sam has been around long enough to watch the architecture pendulum make several full swings. Huge monolithic databases gave way to open distributed systems in the Hadoop era, the market consolidated again around a handful of big cloud data platforms, and now the momentum is swinging back toward open. Apache Iceberg, an open table format that keeps data readable by any engine, sits at the center of the current swing, and Sam calls it absolutely huge right now.

In his telling, AI itself is driving the rebuild. Companies rethinking their data architectures for AI do not want to be locked into any one system, partly for vendor impartiality and partly because no one can say what next year’s patterns will demand. Nobody wants to re-architect everything and then, six to twelve months later, chalk it up to a sunk cost and then pursue the next new thing. So buyers are demanding modular designs at the storage and compute layers, and again at the metadata and model layers, where each piece can be decoupled and swapped without disturbing the rest.

Qlik has put its weight behind that direction. Its Open Lakehouse is built on Iceberg,^[1] and in January the company joined the Open Semantic Interchange, a vendor-neutral industry effort to standardize semantic definitions so business meaning travels with the data instead of living inside any single tool.^[2]

“We may have a completely different conversation about AI in 12 months... having the ability to migrate or swap out certain parts of the stack in the future is almost a must-have.”

— Sam Pierson, Chief Technology Officer, Qlik

Good enough and 10 times cheaper

The most concrete example of adaptability Sam described is the model selector his team wired into the platform early on. The lesson came from the past decade of software architecture: the move away from the monolith toward microservices, where strong interfaces abstract away details so that any one piece can change without breaking the rest. His team knew the frontier models would keep trading places on the leaderboards, so they refused to hard-wire any single one into the product.

“The router inside of our platform can make the decision of what’s the best model for this task, or what is the best cost-performance trade-off, and maybe route this to a model that’s good enough but it’s 10 times cheaper.”

— Sam Pierson, Chief Technology Officer, Qlik

I asked who makes that choice, the platform or the user, and Sam called it a combo platter. Inside the product, Qlik maintains a family of models that it benchmarks and evaluates constantly, and the user never sees any of it. When customers reach the platform from their own AI ecosystem, whether they are building an agent or connecting from a tool like Claude Desktop via the Model Context Protocol (MCP), they have an explicit choice of model and skills.

The tokens you never have to buy

Back in my semiconductor days at IBM, I thought of data as plumbing, and I said so out loud, which made my ETL developers furious. Sam was more diplomatic when I raised the cost question, but his answer vindicated my old developers. Where the data preparation work happens decides what everything downstream costs.

Point an LLM directly at structured data in a cloud data warehouse and watch the meter run. A single question can take ten minutes as the agent inspects the schema, filters it, and analyzes each round of results, generating tokens the entire time. Multiply that by every question an enterprise asks in a day, and you have found one of the places Sam says companies are burning money without realizing it.

Qlik is sitting on an unusual answer here, and Sam admits some luck is involved. It is better to be lucky than good, he joked. The Qlik engine, developed three decades ago, reads data into memory and precalculates multidimensional analyses and their relationships. The company built an AI-friendly interface to that engine, so once a question reaches it, the inference cost is zero. Compared with the warehouse pattern, Sam says customers get higher-quality answers at lower latency and a much lower token cost per question, an argument that echoes what Josh Howard told me about AI being only as good as the data underneath it.

Is freedom real, or a story vendors tell?

Having written my share of freedom-from-lock-in marketing, I put the uncomfortable question to Sam directly. Is this an architectural property you can point to in the code, or a story the industry tells itself? He gave me the honest version. Freedom starts as a philosophy. Not many vendors operating at Qlik’s scale remain independent, and even strong partners like Snowflake and Databricks would understandably prefer that your workloads run on their platforms. Qlik’s bet is that facilitating choice is worth more than capturing it.

Architects have long memories, though, and that history gives the philosophy teeth. The people building today remember being beholden to a small number of powerful vendors, and they refuse to go back, an instinct Sam now sees extending into model choice. For some buyers, the flexibility is peace of mind they will never cash in. He insists it is more than insurance because, as new technologies and patterns emerge, his customers are actually making the switches.

For a buyer trying to separate real openness from brochure openness, the episode suggests a practical test. Check whether your data sits in an open format you could walk away with. Read the terms for a plain statement that the vendor does not train models on your data, a commitment Qlik makes explicitly. Ask about bring-your-own-key encryption and sovereignty options in the regions where you operate. A vendor that passes is selling you an exit, which is the strangest and most reassuring thing a vendor can sell.

Matt Hayes made the data-side version of this argument a few weeks ago on this show, that owning your data keeps enterprise AI affordable. Sam runs the same principle one layer down, at the models and the compute. And if you are wondering whether the freedom people and the governance people at Qlik get into shouting matches, they do, but Sam swears the matches are usually about hockey or baseball.

The prompt-clone test

Two recent guests set up a question I could not resist putting to a CTO. Donald Farmer argued on this show that software features are dying because AI can clone them, and April Dunford called that claim absolute BS. Sam completed the trilogy without hesitating. “Largely it’s BS.”

“If you literally could just get copied in an afternoon by a prompt, it’s probably not all that valuable of a business.”

— Sam Pierson, Chief Technology Officer, Qlik

The line cuts both ways, and Sam knows it. His team sat down and did the work of articulating what is genuinely hard to copy, and the answer was the engine, the data fabric that feeds it, and the governance, security, and quality machinery wrapped around both. What changes is who shows up at the front door. Sam expects fewer people to log into a dashboard and perhaps a hundred or a thousand times as many to use Qlik in the background, through MCP, from inside tools like ServiceNow or Claude Desktop.

Still the first inning

Qlik’s own research shows how early all of this remains. In the company’s 2025 Agentic AI Study, 97 percent of large enterprises had committed budget to agentic AI, yet only 18 percent reported full deployment.^[3] When I asked Sam why the money runs so far ahead of the deployments, he did not blame the technology. Anyone deploying capital wants to know what downstream productivity comes back for a given token budget, and most organizations cannot answer that yet because the practices are not well understood. We are just in the first inning, he told me.

His prescription follows from that honesty. Start with narrow use cases where the value is provable, and pair outside expertise with the people who hold the business context, because that combination, in Sam’s words, is the magic spot. The models will keep improving, and none of it will rescue an organization that locked itself into last year’s answer. The winning AI stack is the one you can change your mind about, at a price you can defend.

Listen to the full conversation with Sam Pierson on the Data Faces Podcast.

Based on insights from Sam Pierson, Chief Technology Officer at Qlik, featured on the Data Faces Podcast.

Share

Leave a comment

Podcast highlights

- [0:06] Introduction and welcome to the Data Faces Podcast

- [1:16] Sam’s role as CTO at Qlik

- [2:00] First job: the teenage lawn-mowing business

- [2:34] What running Talend taught him about where companies get stuck

- [3:55] How AI changes the data integration job

- [7:18] Baking context and guardrails into the data itself

- [9:06] The role of unstructured data

- [10:13] Open formats, Apache Iceberg, and the architecture pendulum

- [13:07] How hard is it to swap out a component?

- [13:43] The model selector and routing to the 10x cheaper model

- [16:02] Token costs, ROI, sovereignty, and IP protection

- [18:43] Is freedom from lock-in a philosophy or an architecture?

- [21:44] Who picks the model, the platform or the user?

- [22:58] Token-maxxing and the zero-inference-cost engine

- [25:43] Freedom versus governance inside Qlik

- [28:05] The prompt-clone question: “largely it’s BS”

- [31:04] Using Qlik through MCP instead of a dashboard

- [32:33] 97 percent budgeted, 18 percent deployed

- [35:30] Close

Frequently asked questions

What is AI model routing?

AI model routing is the practice of directing each task to the AI model best suited for it, based on capability, latency, and cost, instead of hard-wiring one model into an application. Qlik built a model selector into its platform early, benchmarking a family of models and routing each job to the best cost-performance option. As CTO Sam Pierson describes it, the router can send a task to a frontier model when quality demands it or to a model that is good enough but 10 times cheaper.

How do open data formats like Apache Iceberg reduce vendor lock-in?

Apache Iceberg is an open table format that keeps enterprise data readable by any query engine, so the data layer is not welded to a single vendor’s system. Storing data in open formats lets an organization swap storage, compute, or analytics components as technology changes, without migrating the data itself. That flexibility is a leading reason enterprises are rebuilding their data architectures for AI, and it is why Qlik built its Open Lakehouse on Iceberg.

Where do enterprise AI token costs come from?

Much of an enterprise AI bill accrues before a model produces an answer. When an agent queries structured data in a cloud data warehouse, it inspects the schema, filters it, and analyzes intermediate results, generating tokens the entire time, and a single question can take minutes. Architecture decisions made long before deployment, such as pre-calculating analysis in an in-memory engine with an AI-friendly interface, can cut the token cost per question dramatically while improving answer quality and latency.

Can AI clone any software feature with a prompt?

Qlik CTO Sam Pierson calls that claim largely BS. Software that could genuinely be copied in an afternoon by a prompt was probably never a valuable business, and the market shows no wave of software companies losing their revenue to prompt-built clones. The durable advantage sits below the feature level, in assets like a pre-calculated analytics engine, the data fabric feeding it, and the governance and security machinery around both, which cannot be reproduced by generating code.

Why have so few enterprises fully deployed agentic AI?

Qlik’s 2025 Agentic AI Study found that 97 percent of large enterprises have committed budget to agentic AI, yet only 18 percent report full deployment. The blockers are data quality, integration with existing systems, and an inability to prove what downstream productivity comes back for a given token budget. Sam Pierson describes the industry as being in the first inning, where practices are not yet well understood and narrow, provable use cases are the sensible starting point.

How can a buyer tell whether a vendor’s openness is real?

Test for openness in the contract and the architecture rather than the marketing. Confirm the data sits in an open format you could take with you, such as Apache Iceberg. Read the terms for an explicit commitment that the vendor does not train models on your data, and ask about bring-your-own-key encryption and regional sovereignty options. A vendor that passes those checks is giving you a credible exit, which is the strongest evidence that its openness claims are real.

About David Sweenor

David Sweenor is the founder and host of the Data Faces podcast, where he talks with the people who are making data, analytics, AI, and marketing work in the real world. He is also the founder of TinyTechGuides and a recognized top 25 analytics thought leader and international speaker who specializes in practical business applications of artificial intelligence and advanced analytics.

With over 25 years of hands-on experience implementing AI and analytics solutions, David has supported organizations including Alation, Alteryx, TIBCO, SAS, IBM, Dell, and Quest. His work spans marketing leadership, analytics implementation, and specialized expertise in AI, machine learning, data science, IoT, and business intelligence. David holds several patents and consistently delivers insights that bridge technical capabilities with business value.

Books

- Artificial Intelligence: An Executive Guide to Make AI Work for Your Business

- Generative AI Business Applications: An Executive Guide with Real-Life Examples and Case Studies

- The Generative AI Practitioner’s Guide: How to Apply LLM Patterns for Enterprise Applications

- The CIO’s Guide to Adopting Generative AI: Five Keys to Success

- Modern B2B Marketing: A Practitioner’s Guide to Marketing Excellence

- The PMM’s Prompt Playbook: Mastering Generative AI for B2B Marketing Success

Follow David on Twitter @DavidSweenor and connect with him on LinkedIn.

^[1]Qlik. “Qlik Open Lakehouse Now Generally Available, Giving Enterprises Rapid, AI-Ready Data on Apache Iceberg.” Qlik Press Release, 2026. https://www.qlik.com/us/news/company/press-room/press-releases/qlik-open-lakehouse-now-generally-available.

^[2]Qlik. “Qlik Joins Snowflake & Industry Leaders to Support Data and AI Interoperability Across the Ecosystem Through the Open Semantic Interchange.” Qlik Press Release, January 27, 2026. https://www.qlik.com/us/news/company/press-room/press-releases/qlik-joins-snowflake-and-industry-leaders-to-support-data-and-ai-interoperability.

^[3]Qlik. “Qlik 2025 Agentic AI Study: Budgets Surge, but Data Readiness Delays Scale.” Qlik Press Release, October 16, 2025. https://www.qlik.com/us/news/company/press-room/press-releases/qlik-2025-agentic-ai-study-budgets-surge-but-data-readiness-delays-scale.