Season 4 Roundup: What It Takes To Make AI Work in a Real Company

In this Season 4 roundup of the AI Radicals podcast, host Satyen Sangani of Alation and the show’s producer set up the season’s central question: what it takes to make AI work inside a real company.

Producer: Every conversation this season has started somewhere different, but they kept arriving at the same place. The hard part of AI isn’t the model, it’s everything underneath it and around it. In today’s roundup episode, we focused on one thing, what does it actually take to make AI work inside a real company?

Over the next half hour, you’ll hear from a venture investor who used to run Tableau, a Columbia professor building agent infrastructure, the CTO of IBM Software, two innovation leaders from ServiceNow, a New York Times bestselling author, data leaders from the BBC, CNA Insurance, and LVMH Beauty, and one voice out of the archive who predicted most of this three years ago.

All of them in conversation with Alation CEO Satyen Sangani. They disagree more than you’d expect, and where they do agree, it’s worth writing down. These are our core learnings from season four, and this is AI Radicals. This show is brought to you by Alation. Only 5% of enterprise AI pilots from 2025 delivered ROI in production.

Alation’s free agentic AI opportunity discovery guide shows you exactly how to find the right processes to automate and build AI that actually sticks. Get it at alation.com/ai-guide.

Why do enterprise AI pilots never reach production?

Charlene Li, co-author of Winning with AI, explains why executives approve AI pilots but not production, and what changes when AI becomes a strategic initiative. Teams building that roadmap can start with Alation’s AI operating system.

Producer: If you work in data or AI, you probably have a graveyard of projects that never made it out of the sandbox. Charlene Li, co-author of Winning with AI, thinks the problem starts with the word we use for them.

Charlene Li, co-author of Winning with AI: Same thing with pilots. Don’t do any more pilots. And people are like, “What do you mean? I have to test it to see if it works.” I’m like, “If it’s strategic, you’re going to find a way to make it work.”

And if it’s a pilot, everyone has the option to say, “Well, it’s a pilot. When it’s a production, when it’s a project, then I’ll pay attention to it. Then I’ll give you the data. Then I’ll give you the people. Then I’ll connect you to the customers. But while it’s a pilot, go off and do it yourself. I don’t really need to support you and pay attention to it.” But when it becomes a strategic initiative, something that is core important, and your executives are saying, “We need to do this,” everyone goes, “Okay, clear the decks. We’re gonna get this working and going.”

And that’s the key to success.

Producer: Li’s diagnosis isn’t that pilots fail technically. It’s that executives green-light experiments precisely because they haven’t decided anything yet.

Li: When I talk to vendors, they’re pulling their hair out because, especially as I talk to executives all the time, they go, “Can you please talk to this executive?” ’Cause they don’t know what they’re doing, and so they’re just green-lighting these pilots ’cause they feel they need to be doing something. But they won’t green-light into production because they don’t have an overarching AI roadmap that supports their strategy. They haven’t put the time and attention to it.

They’ve given it to IT. They have not really thought about, what are we trying to do with this? The way I think about it is that the business leaders have abdicated their responsibility for AI because they themselves don’t understand it. They’re not prepared. They don’t know. They’re not equipped to be able to have a conversation, a strategic conversation with it, ’cause they don’t, again, understand that it is a strategic issue.

Satyen Sangani, host: Yeah.

Li: So that’s the gap. It’s a leadership gap that we have right now.

Where do companies waste time and money on AI?

Erin McIntosh, who runs global data operations at CNA Insurance, one of the largest commercial property and casualty insurers in the US, names what companies should stop doing in AI deployments.

Producer: That’s the view from outside the building. Erin McIntosh runs global data operations at CNA Insurance, one of the largest commercial property and casualty insurers in the US. Ask her what companies should stop immediately, and she answers in one breath.

Sangani: What should companies stop doing immediately?

Erin McIntosh, global data operations at CNA Insurance: Focusing on a pilot as opposed to thinking about how they could actually scale a solution.

Sangani: Ah, nice. Where is the biggest waste of time and money in AI deployments right now? Follows from the last one probably.

McIntosh: I think seeking perfection as opposed to progress.

So especially as we think about all the things that could go wrong with this and the fear and trepidation, I think we need to focus really on how do we establish the right use cases to show functionality and get us to move forward.

How should you structure a team to scale AI?

Erin McIntosh of CNA Insurance describes embedding technical and business people in pods, building prototypes designed to be thrown away, and handing promising ones to an implementation team.

Producer: Building those use cases and testing them effectively starts with the right team. For McIntosh, that team is comprised of technical and business people, and it demands an environment in which they’re free to experiment and even fail.

McIntosh: Some of the work that we’re doing is actually embedding individuals in a pod structure with the actual business partner. So this is not a technology going off, and using the new tools and techniques to come back to the business. It’s actually bringing them along in that journey. So what we’re doing is actually taking both technical and non-technical individuals as well as the business folks, determining what is the problem, having those individuals sit alongside the business folks to understand what are they actually trying to do and why, and then developing very quickly solutions that we anticipate we could throw away, and that would be absolutely okay, and then we’re also supplementing those teams if we do see some richness in those ideas with an implementation team.

How do you harden it? And I think that’s one thing that we’re really trying to do, and I think organizations really need to think through, is testing and pilots are fantastic, and you do really wanna explore these techniques and technologies. You’ve got to get to a point where you’re thinking, “Could this scale? Would this scale? How would I scale it?”

How do you win support for a data program?

Nathalie Berdat, data director of product at the British Broadcasting Corporation (BBC), explains why multi-year foundation programs fail to win support and where a data leader should start. Her work is also covered in Alation’s BBC data product transformation story.

Producer: Nathalie Berdat, data director of product at the BBC, has learned the same lesson. When you’re building with AI, you must bring along the business.

Nathalie Berdat, data director of product at the BBC: Because let’s make no mistake, I’ve been in an organization that I’ve come in and launched big data governance programs, all sounded amazing on papers, rebuild the foundation, start from scratch, and you go, “Well, you won’t see very much for the next two years, but when it’s done, it’s gonna be amazing.” That doesn’t sell. You have to sell a vision that direct the answer to a business priority, and it has to be a combination of what we can solve now and what’s the vision in the future. So I would say to a leader, either go back or go start work on Monday. For example, pick one metric that... We talked about one metric that has more than one version that’s causing such a headache for your leaders, that is talked about, that create a lot of overhead trying to reconcile and agree.

Producer: Data leaders, listen up. To win support, find the one number that’s causing an argument and fix it.

Should you automate a workflow or reinvent it with AI?

Brian Solis, head of global innovation at ServiceNow, and Dave Wright, ServiceNow’s chief innovation officer, argue that using AI to scale yesterday’s work misses the chance to reinvent how work gets done.

Producer: So if the answer is not building an AI pilot, then what is it? Brian Solis is head of global innovation at ServiceNow and co-author of Infinite. He thinks most companies are using AI to solve the wrong problems when they should be thinking bigger.

Brian Solis, head of global innovation at ServiceNow and co-author of Infinite: They are thinking about how they could be a better version of themselves tomorrow. So essentially, they’re scaling yesterday with artificial intelligence and looking at how we can optimize that work, how we can make it more efficient, more cost-effective. And those conversations are good, I guess, in the short term. They’re necessary, but you have to balance them with competitiveness, and that’s true without AI. I think the conversation that more and more companies, more and more leaders, I should say, should be having is, “What can we do with AI that we couldn’t do yesterday to create new value that keeps us competitive?”

Producer: His co-author, Dave Wright, ServiceNow’s chief innovation officer, turns that into a question you can use in a planning meeting on Monday morning.

Dave Wright, chief innovation officer at ServiceNow: As you’re debasing applying AI in this mode one fashion to optimize things, the real benefit you’re gonna get is not asking, “How do I automate this workflow?” It’s asking, “Why should I automate this workflow?” Because most of these workflows are being designed based on a constraint around the number of people you have to do the job. When you’re putting AI into something, you’re going from being constrained by not having enough resource to having an abundance, potentially almost infinite resource.

So the question you should ask is, “Is this workflow designed to run in the AI age, or is this not just an opportunity to automate a workflow? Is this an opportunity to reinvent that workflow?”

How IKEA turned an AI customer service agent into a billion-euro business

Brian Solis and Dave Wright of ServiceNow walk through IKEA’s AI agents deployment, a customer service agent named Billy that handled roughly 47 percent of inquiries, and what IKEA did next with its human agents.

Producer: The best illustration of that difference this season came from a bookcase. IKEA put an AI agent into customer service and named it Billy after the bookcase. Billy was good. It handled close to half of all inquiries. IKEA could have cut half of its customer service workforce. Instead, it retrained them.

Solis: But with IKEA, this is a popular story, but I don’t know that it’s popular for all of the reasons it should be. And, and Dave, please, you tell a better business version of this story, so I’ll start it.

IKEA might have started this conversation by the way that most companies say, “What’s our use case? Let’s look at call deflection. Let’s put an AI agent in customer service.” Maybe they had a reason to do it, maybe they didn’t. But in most of those conversations, it starts with, we wanna put AI to deflect calls, as Dave shared with his government story.

The AI was so successful, its name was Billy, named after its most popular bookcase. I would’ve named it Allen Wrench. That’s just an IKEA joke. For those that get it, you know, you know. But Billy was so successful at navigating customer service engagements that I think it handled about forty-seven percent is the number, which is pretty substantial. It’s almost half. I think most companies might then look at taking the ROI of AI conversation and realizing that to pay for Billy, we’re going to cut heads to offset that cost and go with an automation story. But what IKEA also did was they analyzed the inbounds, and what they had learned was that a significant volume of inbounds were related to interior design, whether that’s for company, whether that’s for residential, commercial, et cetera. And so what they did was they reskilled, I think it was about eight thousand human agents.

Wright: Yeah, eight thousand two.

Solis: Yeah, eight thousand two hundred with AI to then spin them out as a pilot business of interior design services, consulting services. And as a result of that work in its first year, they generated about one billion euro with an impact of, what was that? Four percent on—

Wright: Yeah, four percent increase in top-line revenue.

When AI automation backfires: the Ford cautionary tale

Brian Solis of ServiceNow cites Ford’s reported decision to rehire 350 quality assurance engineers after letting them go in the name of AI automation.

Producer: That’s the version where it works. Here’s the version where it doesn’t.

Solis: I think we all saw the headlines recently that Ford had to rehire three hundred and fifty quality assurance engineers because they had been let go in the name of AI automation. And as a result of that AI automation, quality actually declined substantially, increased costs because of recalls, et cetera. And upon rehiring them, they put a human in the loop scenario or a human above the loop where AI is now working for these people in the greater purpose of increasing quality like they’re supposed to do and—

What’s the difference between automating and improving a decision?

Erin McIntosh of CNA Insurance separates automating a decision, which gains speed, from improving it with better information, which she says is what changes outcomes.

Producer: Erin McIntosh draws the same line in different language, and hers is the version to keep.

McIntosh: So I think automating a decision is really just, how do we get as fast through that decision process as possible, right? What we’re really trying to do is improving the decision. So can I bring additional information? Can I bring additional insights? Can I arm the individual with better information to make those decisions at scale? That is gonna lead to actually better outcomes. So if I can actually improve the outcome and improve not just the pace of the decision, but the quality of that decision, then I’m gonna actually win in the marketplace, and this is something that we’re really trying to drive. And automation just gets you efficiency. That’s all it’s going to do for you. It doesn’t actually improve the financial components to your organization.

Should AI be used to cut headcount or raise quality?

Charlene Li describes Konecta, a business process outsourcer that used AI to raise quality rather than cut staff, and a bank that retrained about half of the people whose work AI absorbed.

Producer: Charlene Li found a company that made that choice deliberately in January 2023, weeks after ChatGPT launched. Konecta is a large business process outsourcer, call centers, customer experience, exactly the industry everyone expects AI to hollow out.

Li: So we talked to Konecta, K-O-N-E-C-T-A, very large business processing outsourcing, call centers, customer experience. And they said from the very beginning, “We don’t have a headcount problem. We have a quality problem. We don’t have too many people. We don’t have enough of the right people.” So they used AI to increase the level of proficiency, decrease the error rates, so they reduced errors by eighty-five to ninety percent with AI. They gave people co-pilots, so they were able to engage with their customers more efficiently. They cut down the time because they had call summarization.

Producer: Li doesn’t pretend it always goes that way.

Li: No, I talked to a bank that said, “Our AI is so good now. We used to have ninety people reading checks, now we only need five.” That’s not one of those cases where you just don’t need the other eighty-five people anymore. So they said, we could retrain and reskill and reposition about half of them, and the other half just didn’t wanna go through it. It just wasn’t a fit for what they wanted to do. And this is a reality.

Producer: Same technology, two companies. The difference wasn’t the model, it was what the leadership decided the freed-up capacity was for. Every one of those stories assumes something that turns out to be rare, that people involved agree on the numbers. Two of our guests took new jobs about a year apart in completely unrelated industries and diagnosed the same problem within weeks.

What are the signs that people don’t trust your data?

Erin McIntosh of CNA Insurance and Nathalie Berdat of the BBC describe how conflicting numbers for the same metric reveal low trust in data. Teams working on the same problem often pair certified metrics with data quality software.

Producer: Let’s start with Erin McIntosh sitting down with claims adjusters.

McIntosh: Absolutely. So, where I’ll start with is kinda how I entered into data and AI and kind of found my way into this space, and I really wanna call out, after twenty-one years in the industry, there’s a moment that really sticks with me, and I was sitting down with a variety of adjusters in the claim organization, and we were talking about processes that was going on, and they were talking about their different reports. And what they found was that everybody kinda had some different interpretations, slightly different numbers, and it was that inconsistency that really struck. Each person had to learn which version that they wanted to trust and which one they wanted to use based on their own experience. And that became the system of finding the right pieces of information that helped their story.

And that’s really when it clicked for me that this isn’t just a data problem, and it wasn’t just a reporting problem, and it wasn’t just a technology problem. It was a trust problem. And what I took away from that has really shaped my career now here at CNA, is that once trust is gone, people don’t stop or start working. They really just kind of build their own version of reality.

Producer: Now the BBC. Nathalie Berdat joined into a newly created data and AI governance role and watched for the same tell on the number that matters most there, weekly active accounts, meaning people actually logged in to BBC platforms.

Berdat: That number was very dependent on where it was produced and people asking... stopped asking whether the number is right and started asking, “So who produced that number?” As in which team produced it, because then it really allowed them to go, “Ooh, I can trust that number because that team produced it. I’m not sure I can trust that number because it’s used for a different purpose.” So when you see that, it is a symptom, when two teams show up with two different charts for the same metric and no one questioned that, it’s a symptom that data, the trust in data is very low in an organization.

The BBC’s trade-off on a 0.1% data discrepancy

Nathalie Berdat of the BBC explains how her team decided how long to keep reconciling a certified source of truth against the older tables.

Producer: So you build one certified source of truth, then you discover it doesn’t quite match the old tables, and you have to decide how much that matters.

Berdat: What took a long time to get to is we still had discrepancy with the other tables, and there was a trade-off decision to go, “Are we going to invest or spend much longer addressing those small differences?” We’re talking about 0.1%. Every number matters. Every active user matters. But then there is diminishing returns. You could spend a long time trying to figure out where are the differences, and we knew sort of where the differences were from, and it wasn’t something systematically that we would worry about.

Can AI agents work with messy data?

Anant Jhingran, chief technology officer of IBM Software, explains how AI tolerates input errors that would break a traditional API, and the data fidelity floor he says remains.

Producer: Anant Jhingran, CTO of IBM Software, pushes in the opposite direction. He spent a career in databases and APIs, and he thinks the machines are stranger than we give them credit for. They’re oddly forgiving of messy input.

Anant Jhingran, chief technology officer of IBM Software: It is actually quite remarkable how the AI that uses the data is actually fairly tolerant of bad stuff. Okay? So previously, if my API had a /member/ID and I accidentally typed /member/ID as opposed to /member/ID, then it’ll just kind of return whatever it is. I don’t even know what the code will be, probably some 400s or something like that. But today, AI can say, “Okay, I’ll type /member/ID,” and it’ll come back with that code, and it’ll kind of reason among itself and say, “Oh. Uh-oh, I got it wrong.” Okay? But they kind of self-correct around stuff that very bad. So in some ways, if you believe that the users of all of this stuff are shifting towards AI, you have to be less prescriptive with it being exactly correct.

Producer: But even Jhingran draws a floor, and it’s the floor the whole season rests on.

Jhingran: But exactly. After that, obviously without that kind of fidelity, everything else is basically garbage. And as much as the agents and AIs kind of change, they’re improving on their hallucination. Right? Hallucination is a big problem two years back. It’s a smaller problem now because there’s enough RLing that’s happened and everything else. But the fact of the matter is, if the data doesn’t have high fidelity, any amount of lack of hallucination is not gonna solve it.

Producer: We’re brought to you by Alation. Healthcare organizations are sitting on some of the most valuable and most fragmented data in the world. How are leading innovators getting that data ready for AI to better serve patients? Alation partnered with teams at Children’s Hospital of Philadelphia and Rady Children’s Health to document how leading health systems are actually doing this, building trusted data foundations, creating usable data products, and deploying AI that clinicians will actually act on. If you’re working at the intersection of data and healthcare, this guide is worth your time. Grab it free at alation.com/ai-healthcare.

Is data governance really slow and bureaucratic?

Erin McIntosh of CNA Insurance says data governance has earned its reputation for being slow, and explains why she treats effective governance and burdensome governance as different things. For evaluation, see Alation’s data governance platform.

Producer: If trusted data is the precondition, governance is the work that delivers it, and governance has a reputation problem that every guest this season acknowledged. Erin McIntosh was asked what people get wrong about it.

Sangani: What’s the one belief about data governance that’s just wrong?

McIntosh: That it is slow, and I will say good data governance is actually effective, bad data governance is actually slow and burdensome.

Producer: Which is easy to say and hard to prove, and McIntosh knows the challenge, she says. Her answer is to take her own medicine.

McIntosh: The biggest question that we’re getting is, “Okay, you’ve talked about governance for a long time now. Nothing’s really changed in the governance space. How’s it gonna actually work, and how’s it gonna scale, and how’s it gonna improve?” Because governance organizations have failed along the way, right? We’ve not seen them scale well. We all have those stories from our prior experience. And that adage that governance is slow and bureaucratic, I will say, yes, it has been. The difference now, and this is the direction that CNA is taking, is we know we need governance, so how can we actually take our own medicine, and can we actually apply agentic capabilities to deliver governance at scale, minimizing or guiding those stewards, those business owners, those quality assessments?

How can AI agents help scale data governance?

Erin McIntosh of CNA Insurance names three places to apply agents in governance: data stewardship and ownership, regulatory compliance, and the data catalog.

Producer: For McIntosh, agentic data management offers enormous promise.

McIntosh: Well, then let’s throw out the playbook, and let’s rewrite it, and let’s do this agentically led. And that’s the direction that we’re headed, and that’s actually what I’m most excited about is not having to hire consultants and stewards and having them be, 20% of their capacity, but really giving them a cockpit, if you will, on how we can drive governance and improve this.

Producer: She shares three key places to start.

McIntosh: So a couple places where we’re going to be focusing, one is around stewardship and data ownership. In governance, getting stewards to buy in, getting them want to stay in the roles, getting them want to actually develop rules and definitions, it’s a hard job. And after two years, they typically move on to other things, so that sustainability. So we’re gonna be focusing on how do we create agents to help those stewards and if not actually become the stewards. So how can we do that? That’s a major piece. Another piece that we feel is ripe for agentic capabilities, I will tell you, is compliance. So we’ve all seen the various legislation documents, GDPR, secrecy acts. How can we actually create agents built to interpret and help guide us in that? And then the last piece that we’ll be doing is how do we actually engage with our catalog and serve up those insights and quality together to actually guide those stewardships?

Should agent rules live in the prompt or the system?

Eugene Wu, associate professor at Columbia and co-director of the Data Agents and Processes Lab, explains why many rules people write into agent prompts can be enforced by the system itself. Host Satyen Sangani adds a stop-sign analogy.

Producer: That’s a practitioner betting that governance is process shaped enough to hand to agents. Eugene Wu has the research that explains why the instinct is right and why most teams are doing it in the wrong place. Wu is an associate professor at Columbia and co-directs the Data Agents and Processes Lab. His students went looking at what people actually type into agent prompts at large organizations.

Eugene Wu, associate professor at Columbia: So with coding agents or any of these automated agents, what you actually see in these prompts is a huge portion of it is actually trying to tell the agent what to do and what not to do, right? Obey this policy. Don’t write to this file. Hey, before you commit, you should always do X, Y, and Z and so on. Or don’t touch the database. Don’t touch the database, right? And maybe that’s copy and pasted like five times. And so there’s all these things that people put in, right? And our HSI students found kind of like hundreds or like thousands of these rules. But the thing is we looked at these, and we said, “Oh, actually, the majority of this stuff we can push into the system,” right? And so it doesn’t have to be probabilistically enforced, right? We’ll just enforce it.

Producer: Satyen shared an analogy describing how we should rethink prompting agents.

Sangani: On some level, the policy lives with the object that it’s trying to govern, not necessarily with the agent that’s trying to leverage it. It’s like, you put a stop sign at a road. You don’t put the stop sign in the head of the user so they have to memorize that stop sign every single time they are getting to the intersection because they wouldn’t know how to disambiguate when to stop and when not to stop.

How should you govern autonomous AI agents?

Dave Wright of ServiceNow explains how governance maps help spot misbehaving agents, then Nathalie Berdat of the BBC and Erin McIntosh of CNA Insurance disagree on how much autonomy an agent should get. Both want AI governance that machines can read.

Producer: ServiceNow’s Dave Wright adds a layer above that, what you watch once the rules are in place.

Wright: Once you start to get those governance maps in place, that allows you to spot the exceptions. So once you can start to understand the trend of how often an agent runs, if it normally runs 50 times an hour and suddenly it’s running 8,000 times an hour, you’ve got a problem. If you start to see a prompt length vary that never normally varied, then you might have a problem.

Producer: Which brings us to the sharpest disagreement of the season. Both of these leaders want governance that machines can read. They differ completely on when an agent gets to act on it. Nathalie Berdat at an institution built on public trust goes first.

Berdat: So before we would hand an agent autonomy, and even to a small degree, we absolutely need clear ownership of the data it would touch. We need that documented lineage so we can trace a bad decision back to its source. We need machine-readable governance rules, and not just a policy. It has to be embedded.

Producer: And Erin McIntosh at a nationwide insurer whose entire business is pricing risk argues for innovation.

McIntosh: So the risk of doing something is better than not doing anything in this space. And I really hate saying that, but that’s really where we’re at. So it’s a conversation that we’ll have to have. It is internally focused. These are for our governance channels. We know we’re not pushing this out, so we feel that it’s a relatively safe place to start. We will have more of a human direction in there to start with, and then we’ll be able to toggle this off.

Producer: A broadcaster built on public trust holding back, an insurance company whose whole business is pricing risk moving first. Both of them are being careful and innovating at a price that is right for them.

How does poor metadata quality affect AI applications?

Eugene Wu of Columbia explains how weak metadata undermines search results and pushes the burden of judgment onto the end user. Alation covers the link between metadata and AI accuracy in more detail.

Producer: Eugene Wu has the version of that failure that starts much further down. If the metadata underneath is bad, the damage surfaces somewhere you’d never look for it.

Wu: If the quality of the metadata is not very good then you can’t really trust the search results. And if you can’t trust the search results, then any application you build on top of it needs to expose this information to the user, right? And the user has to make a judgment call. And that places a huge cognitive load on the end user trying to use this tool. Something very, very, you know, low in kind of like the stack, just indexing, can permeate all the way up to the end user.

What does it take to keep AI-generated output reliable?

Francois Ajenstat, former chief product officer at Tableau and now building Golden Analytics, describes why AI-generated work needs ongoing supervision and what happened when his team connected AI to all of its tools.

Producer: Francois Ajenstat spent a decade as chief product officer at Tableau, and is now building Golden Analytics, where as he puts it, 99% of the code, the documentation, and the internal knowledge is AI-generated. He’s clear about which day matters.

Francois Ajenstat, Golden Analytics and former chief product officer at Tableau: I think the same thing will happen is the AI is only as good as the supervision you give it, right? If you just let it go do its thing without ever tweaking it, changing things, adapting things, I don’t think it’ll ever be good. So the fact that it’s AI-generated is good today, but it requires care and feeding for it to be amazing.

This magical thing is you get this sugar pill on day one, like, “Look at what it did so quickly.” But what about day two, day three, day four? That’s actually even more important because there is drift, there are challenges that happen, right? Inconsistencies that you have to maintain.

Producer: He also ran the experiment everyone assumes will work. Connect the AI to everything you’ve got and let it figure things out.

Ajenstat: We did a test where, with Golden, we just connected to everything and said, “Great, now enrich the answer.” It couldn’t conclusively connect the dots. It tried. It invented possible correlations. But it wasn’t quite there because there wasn’t the work done ahead of time. The fact that it could read my Notion doc and Slack, it found things that happened at the same time, but not necessarily the real answer to the question.

So our view of it is if you have no context, you should be able to get value, and if you have a lot of it, you should be able to leverage it, but just giving more context actually doesn’t give you a better answer.

Why are AI agent systems never really finished?

Erin McIntosh of CNA Insurance flags the cost of AI and of maintaining it, and Satyen Sangani and Columbia professor Eugene Wu describe agent systems as a continuous improvement loop with permeable boundaries.

Producer: None of which is free. Erin McIntosh names the line item most AI business cases leave out.

McIntosh: What’s the cost of AI, and is it gonna be cheaper and more efficient and more effective to do it this way, even with that throughput and quality, than having a human doing it, and kind of making those trade-off conversations. We’re having a lot of those discussions. And also the cost of maintenance.

Producer: Which is the thing that separates an agent system from a piece of software. Satyen put it to Eugene Wu directly.

Sangani: And so you’re in this constantly iterating loop of improvement that’s needed because you can’t assure fidelity day one in the same way that when you build a software program you can. And so it’s just like you’re never done with the work on some level. But these things are just completely not... Like, they’re completely tied together. So where you wanna draw big lines with boxes, the lines are actually highly permeable. I mean, it’s a little bit more like biology than it is like physics on some level.

Wu: Yep. Yeah. This is kind of like that semantic coupling, right? Yeah. It’s like everything affects... Somewhere that you’ve never even heard of or thought about is gonna affect whether or not you get correct results or good results, and that’s very, very difficult to even kind of like measure and evaluate.

If code isn’t the bottleneck anymore, what are people for?

Mark Nelson, a venture investor at Madrona and former CEO of Tableau, uses an analogy from computer scientist Yejin Choi to explain where the bottleneck moved once AI can generate code.

Producer: Which leaves a question underneath all of it. If the machines are doing more, what are the people for? Mark Nelson spent three decades in enterprise software, including as CEO of Tableau, and now invests at Madrona. He borrows an analogy from the computer scientist Yejin Choi.

Mark Nelson, Madrona and former CEO of Tableau: And the analogy she used at the end is like, look, we just invented skyscrapers, and skyscrapers were amazing, and they changed human life, and they changed how cities were, and, like, they changed so much about how the human dynamic worked. But by the way, if you wanna go to the moon, like, you can’t build a taller skyscraper. Like, the taller skyscraper is not going to get you to the moon.

Producer: So the bottleneck moved. Nelson is precise about where it went.

Nelson: It used to be getting the code generated was the blocker, right? And now kind of overnight, generating code is not the blocker. Oh, but great software isn’t just flowing off of people’s fingers either. Software is flowing off of people’s fingers. Great software is not. But do you have the good taste to understand what to build? Can you figure out how to sell it? Can you put all the things around it to make it a useful product for your customer?

How do you keep human taste in AI-assisted work?

Julie De Moyer, who leads data and AI across 15 beauty brands at LVMH (Louis Vuitton Moët Hennessy), Ben Taylor, former chief AI strategist at Dataiku, and Anant Jhingran of IBM Software describe how AI can assist human craft at scale. Raising everyone’s level starts with data literacy.

Producer: Julie De Moyer runs data and AI across 15 beauty brands at LVMH, where the product is craftsmanship. Her AI is deliberately hard to see.

Julie De Moyer, data and AI leader at LVMH Beauty: Along the value chain, I’ll give you a couple of examples. If you look at the making of perfumes or the way we actually make the wines. In other industry, we would use the AI in order to help those, I would say, the scientists to go faster, to optimize their trials. It will never replace the final sense or the final product that is decided on, but it can help with the substitutions of products that might need to go out of, as a result of regulatory changes. It might also help with making sure that the quality of the products lasts as long as possible. So we really help those researcher scientists do their job better and easier.

Producer: Back in 2023, Ben Taylor had a phrase for the same instinct, borrowed from someone else and sharpened.

Sangani: There’s this saying that there’s no soul in the machine, only in front of it.

Ben Taylor, former chief AI strategist at Dataiku: We don’t celebrate the machine, we celebrate the user. We celebrate the soul, the human behind it, because the human’s gonna understand how to quantify the value. The human’s gonna come up with the best problem, and the human is ultimately the creator, the artist, the creative.

Producer: Anant Jhingran’s version of the taste problem is a scale problem. It’s easy to have a handful of master craftspeople. He’s responsible for thousands of engineers.

Jhingran: Within that context of taste, the thing that I’m focused on, again, coming back to the scale, is how do you raise everybody’s level, okay? So they’re not just slop producers, but are actually working towards and working through taste. And I read a lot about, look, Linus himself came out with how will he accept AI, et cetera, within his code base. And so there’s a lot of interesting work to be done out there where we go from, look, here are the master craftsmen or craftswomen, to how do you go to kind of mass thing?

What should you do with the time AI gives back?

Charlene Li, co-author of Winning with AI, and Brian Solis of ServiceNow ask what leaders will do with the capacity AI frees up, a question Solis says few executives can answer.

Producer: And Charlene Li’s answer to the human versus AI problem is the most ambitious one anyone gave us.

Li: I do believe it’s gonna make us better humans. What it does is that, again, there’s just ways to optimize your work and be more productive and everything, but if we only just use AI to do more work, that, I think, we would’ve missed the point. If instead we use AI to do all the work that we can supplement with, and then we use that extra time and capacity to develop what is uniquely human about us, our empathy, our self-reflection, our intuition, our judgment, our wisdom. If we were to elevate those things, those are the things that combine with a very strategic and deep use of AI allows us to become truly superhuman.

Producer: Which brings us back to Brian Solis and the question he says almost nobody in the room can answer.

Solis: Let’s say you use the metric that Dave shared that we save 10,000 hours, for example. The Venn diagram is what are you gonna do with that time? What are you gonna do with that freed up resource that can create new value? And that’s a conversation that most people aren’t gonna have today, just to be honest, because it’s not a discipline. And when, for example, Dave or I ask that question, I almost get that face. They might not say it, but I get that face or that look like, “I hadn’t thought about that,” or “nobody’s asked that before. We just thought we’d stop right after we saved so many hours.”

What does it take to make AI work in a company?

The AI Radicals producer closes Season 4 with three lessons: stop calling AI work a pilot, earn trust one reconciled metric at a time, and enforce rules in the system rather than the prompt. A sponsor message from Alation about its data products marketplace follows.

Producer: Across seven conversations, three things held. One, stop calling it a pilot. If it isn’t strategic enough to go into production, it isn’t strategic enough to do. Two, trust is earned one reconciled metric at a time. Both of the data leaders in this episode started in the same place, not with a problem, with a single number that two teams disagreed about. Three, the rules belong in the system, not in the prompt. A stop sign goes at the intersection. And underneath all three, something none of these guests set out to say.

For years, the industry built data stacks with clean layers and clean handoffs. What these leaders describe is coupled, permeable, and never quite finished, closer to biology than physics. That’s uncomfortable if the goal was ever to be finished. It’s good news for anyone who’d rather keep adapting. Brian Solis asks what you’ll do with the hours you just got back. That’s the question worth leaving with. Thanks for listening to AI Radicals. See you next time.

Before you go, a word from our sponsor, Alation. You’ve been listening to us talk about making AI work in the enterprise, and here’s one of the biggest unlocks. Your data has to be trustworthy before AI can be. That’s exactly what Alation’s data products marketplace is built for. Data teams use it to package raw data into certified, governed, reusable data products without writing a line of code. An AI-assisted builder does the heavy lifting, governance travels with every product automatically, and business users can chat with data in plain English and get accurate explainable answers.

Companies like Cisco, Pfizer, and Nasdaq are already doing this at scale. If you want your AI initiatives to actually land, this is where it starts. Download the free data sheet and see what AI-ready data actually looks like at alation.com/data-products.

Frequently asked questions

Who should own AI strategy, IT or business leaders?

Business leaders should own AI strategy, according to Charlene Li, co-author of Winning with AI. She says many executives hand AI to IT and abdicate responsibility because they don’t understand it, so no overarching roadmap tied to business strategy exists. She calls this a leadership gap, and says executives won’t green-light production without that roadmap.

What costs do AI business cases usually leave out?

The cost of maintenance is the line item many AI business cases leave out, according to Erin McIntosh of CNA Insurance. She also says teams should weigh the total cost of AI, including throughput and quality, against having a human do the work. Her point is that an AI system keeps incurring cost after it launches.

Does connecting AI to more company data improve its answers?

No, according to Francois Ajenstat, former chief product officer at Tableau, who tested this at Golden Analytics. His team connected the AI to everything, including Notion and Slack, and it invented possible correlations instead of finding the real answer because the groundwork was missing. He concludes that giving AI more context does not by itself produce a better answer.

Is AI hallucination still a problem for enterprise AI?

Hallucination is a smaller problem than it was two years ago, according to Anant Jhingran, chief technology officer of IBM Software, but it does not fix bad data. He says improved training has reduced it, yet if the underlying data lacks high fidelity, any reduction in hallucination will not solve the problem.

What are the warning signs that an AI agent is misbehaving?

Sudden changes in an agent’s normal behavior are the warning signs, according to Dave Wright, chief innovation officer at ServiceNow. An agent that normally runs 50 times an hour and suddenly runs 8,000 times an hour has a problem, as does one whose prompt length starts varying when it never did. Governance maps make these exceptions visible.

Want to see how Alation helps data teams get AI-ready? Book a demo or learn more.