Why Model Choice Is the Wrong AI Obsession

[00:00:00] A huge thank you to Donodo for supporting the Tech Talks network and helping power more than 60 interviews every month. And if your business is already invested in Data Lakehouse, this next message is especially for you. Because if you've already invested in a Data Lakehouse like Snowflake or Databricks, Donodo will help you get more value from it. Faster AI, smarter analytics and trusted self-service data access.

[00:00:28] So turn your Lakehouse investment into real business outcomes with Donodo. So please visit Donodo.com. What if the model your company has spent months selecting matters far less than whether sales and finance agree on exactly what revenue actually means?

[00:00:53] Well, this is a wonderfully unglamorous problem at the heart of enterprise AI. And my guest on AI at Work today is Emma McGrann. She's CTO at Action. She's a self-described data obsessive who builds Lego and judges an old fashioned by the quality of its eyes. I've got a feeling we're going to have a fun conversation today.

[00:01:18] But on a serious side, Emma will argue today that competitive advantage actually comes from enterprise data, shared business meaning, clear ownership and employees that are willing to question those confident AI answers. Dare I say overconfident sometimes. And we'll also explore the gap between boardroom confidence and the reality seen by data engineers in their day-to-day work.

[00:01:45] And why agents expose weak governance and how one data product and one data contract can move a useful AI project toward production. So if your pilots look impressive but collapse when connected to real data, hopefully this conversation will help explain why. But enough for me. Let me introduce you to my guest right now. So thank you for joining me on the podcast today, Emma.

[00:02:13] Can you tell everyone listening a little about who you are and what you do? Yeah, sure. So my name is Emma McGratton and I'm the chief technology officer at Action. And I'm obsessed with Lego. That's really what I do. I make Lego and I drink old fashions. And I have a job that funds all of that. So my job at Action is to talk to people about data, right? And how we're... At the moment, data is like the sexiest it's ever been, right? But I started my career in data over 30 years ago.

[00:02:42] And back then, data was like plumbing, right? It was kind of in the background. You kind of assumed it was always going to work. And when it didn't work, things got really ugly. But now data is like a conversation piece, right? And I can talk to an Uber driver about the fact that I'm helping customers get their data ready for AI or to identify where they may have some problems. So a big part of what I do is speaking publicly. And there's three topics I'm really passionate about.

[00:03:10] One is data prep for AI, right? And the second is around the ethics of AI, right? And what data we should actually be exposing to AI. And making sure that as we're training AI, that we're doing it with data that's appropriate. And that we're not having bias built into some of the models and so on. And the third area I'm really passionate about right now is data sovereignty, right? And how do we unhook ourselves from our complete dependency on American tech, right? Because there's some challenges there, right?

[00:03:39] You see some of the cloud providers, some of the frontier model providers have you over a barrel on some of these things, right? So we're trying to figure out how do we help companies get unhooked from tech? So data AI sovereignty is the third topic that I'm quite passionate about at the moment. So doing a lot of public speaking. I'll be in the UK for Big Data London next month. So quite looking forward to that. Always a great event. And this year, Louis Thoreau is doing the key notes. So quite cool. Yeah, that'd be cool.

[00:04:08] I must admit you had me at building Lego and drinking old fashions as well. And you mentioned data sovereignty there, which is a huge topic over here in Europe and something I hear more and more about. And when I'm doing a little research on you, I was also reading that you argued that enterprises are possibly spending too much time debating on, hey, which AI model is the fastest, the cheapest, or even the smartest. Because if model... So on that side of things, if model choice isn't what separates the winners from everything else,

[00:04:38] the question I've got to ask is what is? What should they be debating? Right. So to me, yeah, the models are almost being commoditized at this point, right? It's like a horse race, right? It's like, oh, one is slightly ahead of it. Oh, the other one is slightly ahead, right? So to me, you know, which is the smartest, which is the cheapest, which gives you the best value for money? Definitely important to consider as you're rolling out your AI projects into production. But for me, the biggest differentiator that any enterprise has is its data, right?

[00:05:07] So your data is the most. So understanding your data, its readiness for AI, right? Right. Identifying the use cases where you're ready to rock and roll and you should start moving forward and build confidence in your teams that you know how to deliver on AI production. But, you know, the really important thing is understanding semantics. And organizations are very much like families. They can be very dysfunctional, right? And nobody knows how everything works.

[00:05:32] So as we try to help customers deliver on AI projects, understanding the semantics that the business uses can be a very interesting and fun exercise, right? Because you use something as simple as revenue, right? So revenue to the sales team is very different to the finance team and to the tax team. So when you're asking, you know, show us our top 10 customers by revenue, right? What exactly do you mean by revenue in this context?

[00:05:59] So understanding the context in which that data is going to be used is incredibly important. So when it comes to, you know, that differentiation, it's so much less about the models. It's so much more about getting your data ready, getting your teams ready, right? The organization also has to be ready for AI. Like understanding, like, what are the implications of getting it wrong, right? And how do we build trust? How do we build a data curiosity that, to me, in every function across the organization,

[00:06:28] you need to have, you need to encourage people to be curious about the data, right? To ask questions, to have this kind of gut feel as to what looks right and what might be slightly off. Because we need those humans in the loop as we're deploying AI. We need people to actually understand what should the answer look like? Because AI is always going to be very confident in how it delivers an answer. And it's been at that same confidence with a completely wrong answer and a completely right answer.

[00:06:54] And humans don't do a great job of differentiating when the AI gets it right or wrong. That's why I really encourage people to, as well as looking at their data and semantics as a differentiator, also really focus on their people and building out a data culture across the organization. And elsewhere, I also came across Actium's research that found that 83% of organizations out there are facing data governance and compliance challenges.

[00:07:21] But many businesses still rate their governance maturity remarkably high. So why is there such a gap between that perceived readiness and what is actually happening inside enterprise data environments? Is this something you see a lot too? We do see a lot of it, right? Because we work a lot with heavily regulated industries, right? So financial services and healthcare. And when you talk to them, they'll say, we've got governance mails, right? We've got a governance council.

[00:07:49] We've got data and AI policies. We've got data catalogs. We've got all of this and we're well governed, right? But ask them a question like, can I take this particular piece of data? Can I give this particular AI model for this particular purpose within this jurisdiction? And can I do that right now? And they're like, oh, well, hang on a second. That's not something that I haven't answered to right now, right? So I think that paradox that 83% just having governance challenges while the organizations

[00:08:18] think that they have good governance practices. I don't think they've thought about removing the human from the equation, right? Because the human can deal with ambiguity. The human can deal with answers that don't look quite right. And they'll dig in or they'll make a phone call and they'll figure out what the right answer should be. But once the agents take over and we allow for autonomous decision making, that data has to be 100%.

[00:08:44] So governance for AI is something that a lot of organizations are concerned that they need to take a lot of time to get it just right. And I think they're right to be concerned and to invest the time now. Because once a stream is out of the bottle, there's no flicking it back in. And when I dug a little bit deeper on some of the findings in the report, another stat that stood out was executives rate their data maturity 12% higher than operational managers.

[00:09:11] So are AI strategies being approved in boardrooms based on assumptions? Maybe assumptions that their engineers and data teams closest to the technology simply don't recognize. What's going on here? Yeah. So we've seen this. And it's something we've seen for a couple of years where the executive team is saying, look, we're spending all this money on governance, right? We've just been through our compliance reviews and we've passed the audits and so on.

[00:09:40] So they're like, we've got good governance practices in place and we're ready to roll out AI. And then you've got like the data engineer that's actually looking in the data and they're like, oh God, this is all like nothing but duplicate customer records. We've got undocumented data pipelines. We've got spreadsheets that now all of a sudden are the mission critical in the organization. I've got a field that this whole business is kind of pinning their decision making on

[00:10:08] that's called revenue final version seven verified and revised, right? And it's like the people that see just how messy the enterprise data is are a lot more gun shy, right? They're like, yeah, maybe we're not quite ready yet. Whereas the at the boardroom level or in the executive level, they're like, well, we've invested all of this money in building out governance frameworks and delivering upon the audited

[00:10:35] regulatory compliance requirements for our business. So we've got to be ready. So I think both have a, both are correct in their thinking. But I would spend a lot of time talking to the data engineers and really understand what are some of the challenges that we have. And because any problem that you have in your data, right? AI is going to just shine this massive magnifying glass at a spotlight on it, right?

[00:11:04] And that little problem in your data all of a sudden could cause some real challenges for the business. So I think taking the time to get it right is something that I always encourage. And that's something that leaders need to make sure that their data teams are comfortable with the rollout and include the data teams in conversations about, you know, what are we going to deliver on these particular AI initiatives into production?

[00:11:31] And another phrase we've heard a lot over the last few years is AI ready. Beyond that, it's AI native. That seems to be the latest one. But of course, enterprise data was largely designed for reporting dashboards and human interpretation, a very different era. So what does data need to look like when the consumer is an AI system or autonomous agent rather than a person? I would imagine it needs to be a little bit different. Yeah, very different, right?

[00:12:00] Because, you know, as I said earlier, humans are great at dealing with ambiguity, right? And they fill in the gaps that agents can't fill in, right? You've got to give the agent all the pieces that they need. So the way I often talk about this is, for the longest time, we've treated data as a byproduct of different applications, right? The applications are out there. They're collecting and updating data that we then store in a repository for a bit of cleansing, a bit of analytics later on.

[00:12:28] But what we're really encouraging enterprises to think about is, think of your data as the product, right? So bring in this concept of data products, right? So a data product, treat it like you would any software product that you're delivering in the organization. So it has to have an owner, right? It has to have a set of requirements that are articulated and well understood across the business. The quality of the data product has to be measurable, right?

[00:12:54] So if you're dealing with something like financial data, it has to be 100% accurate 100% of the time. If you're dealing with marketing data and you're about to win a campaign and you discover that, you know, 10% of the email addresses are empty, right? It doesn't really matter. You can win your marketing campaign on 90% of the data that you've captured and that's totally fine. So understanding that quality for that particular use case is important, right?

[00:13:19] And then making these data products discoverable by both humans and by machines, right? Is the first step. And the second step, and I really strongly believe in data contracts, right? So a data contract is an agreement between the data producers, so the data team and the data consumers, which is the business team, on exactly what's going to be delivered, right? So what's the format of this data? What's the frequency at which it's going to be refreshed?

[00:13:46] What are the quality signals that it's going to conform to? Who owns it, right? And typically we recommend two owners. One will be on the data side. Somebody you can call if, you know, oh, I noticed that this column's missing from this latest drop of the data product, right? So you can call that person, you know, who the owner is. But you can also need an owner on the business side, right? Somebody you can say, I don't understand why, you know, the revenue has dropped by 23%. Can you explain what's gone on with the pipeline here?

[00:14:15] So having that owner is important. Understanding who can use this data product and for what purpose, right? So it may be like certain data sets, HR data, for instance, right? The HR team has access to all of that data. If you go beyond that, I lead our CTO office, right? So I have access to my CTO office employee data, but not all of it, right? I can't see their social security numbers in it. I don't need to know that, right?

[00:14:43] But being able to understand who can see what from that data product is important, too. And then being able to enforce these contracts, right? So this isn't something that's sitting on a wiki, right? It's a contract that's wrapped around your data product. And then you can have an observability engine that makes sure that you're conforming to the terms of the contract. So let's say you say this is trade data, right? And it needs to be updated every one minute, right?

[00:15:12] And all of a sudden you see that that data is not being updated. So you're dealing with something like foreign exchange trading, right? That can change in seconds, right? The rates at which currencies are being exchanged. So you don't want to make decisions on stale data. So you could actually have a circuit breaker in place and say, hit the circuit breaker. I don't want us to transact any business on stale data, right? So let's figure out what went wrong in the data pipeline and get that data corrected. And then we can resume business.

[00:15:42] So that data conflict is the second thing. And then the third is semantics, right? And we use, we throw around terms like customer and revenue and so on. And depending upon where you are in the business, those things can mean very different things, right? So for me, in my role, to me, the customer is the person within an organization that we have a contract with who's feeding me ideas that will help us innovate and deliver on like the next generation of our products and really understand where the market's going and so on.

[00:16:11] In our finance department, the customer might be the accounts payable department in that account, right? To our sales team, it's the person who actually engages in the contract negotiation and signs a contract. So when we say, show me my top 10 customers, for me, it's the 10 most innovative customers that we're working with, right? That's who I want to talk to. If I'm doing a customer advisory board or I'm visiting a city and I want to get some of our customers together

[00:16:38] for an evening of old fashions and mega, I can say, show me my top 10 customers in this city. Whereas for the finance team, they typically want the customers that are spending the most with us, right? So they may have a campaign where they want to go and talk to those customers. So giving the AI the context is hugely important, right? So for every data product, we should have the contract wrapped around it and the semantics embedded within it, right?

[00:17:05] So saying for this particular data product, customer means this, right? Churn means this, revenue means this, and have very specific definitions so that when the AI is making decisions, that it's making the right decisions and it's got everything that it needs to actually make those decisions. So in the research that we recently did with Bark, what we saw was that those organizations that embrace data contracts and data products

[00:17:31] were 3.4 times more likely to see success with scaled AI across the enterprise. And scaled, it's not like a massive thing at the moment. To scale your AI, it's just seen as being successful with three or more AI projects in production, right? So what we're seeing is that you've 340% more chance of being successful if you're embracing data contracts and products than if you're not, right?

[00:17:59] That if you just see them as a bit of a science experiment within the data team. So really believe in that. And it's borne out by the Bark study that we've shared on the Acton.com website. Wow, so many great points there. And if we spend any time scrolling down our news feeds on LinkedIn or company pages on LinkedIn or indeed keynotes at tech conferences, it can feel like every company out there is living the AI dream.

[00:18:26] But a quick peek behind the curtain reveals that your research actually found that challenges are almost everywhere, whether that be scale and complexity to security, data quality, skills, and culture, which is often underestimated. So for any leader listening who knows their foundations aren't ready or there are a few problems, where should they begin rather than attempting another enormous data transformation program? Yeah, absolutely Kevin, another transformation project, right?

[00:18:55] We do that for as long as I've been in this industry, which is over 30 years, right? So what I always caution is, when you look at your data infrastructure, it can be quite overwhelming, right? You've got the average enterprise, according to Gartner, has 400 different data sources, right? So how do you make sense of all of that, right? And how do you get that all governed and to the quality you need and so on? So what I always recommend is find one AI use case that matters to the business, right?

[00:19:25] So find just one, right, where the business is actually going to see value in what you deliver. Define a data product that's going to satisfy the needs of that one use case. So put that data product in place, put a contract around it, and deliver on the quality for that, right? How fresh is that data? How are nulls acceptable? Are duplicate records acceptable and so on?

[00:19:52] So putting all of those quality signals in place, understanding what's going to be delivered, deliver it to the business. And you'll build confidence then in your data team that they know how to do this. You'll build confidence in the business team that the data team knows how to deliver on AI. And then the whole enterprise becomes much more confident that we know what we're doing. But I would start with one data contract, one data product, right? Deliver on that, show success.

[00:20:21] And then when you've got that confidence, you've got that experience, then build it out. So you could handle your entire data estate in the order in which the business is consuming things, right? So start small, show success, and build out from there. But going back to that 340% improved chance of delivering on success is with data products. I think it makes sense to start with small, single data product, show success, and grow out from there.

[00:20:50] And I also love your argument around data culture because I think it's something that moves way beyond just architecture and governance. But for people listening, what does a strong data culture, what does that actually look like inside an organization? And how do they know whether their employees genuinely trust and understand the data that they're giving to AI? I suspect it's a conversation you're having quite a lot at the moment. But tell me more about that. Yeah, sure. So I'll start with a little anecdote there, right?

[00:21:19] So when I first got a GPS, right, I was so excited by a little garment that I had on the dashboard of my car, right? And I used the GPS to go anywhere, grocery store, one mile, straight shot up the road, put the GPS on, right? And there was one evening I was coming home from work and I had the GPS on and it said, take the next ride, right? So we started taking the next ride and then realized it's the exit ramp from a freeway, right?

[00:21:45] So this car's flying down this thing and I'm about to head, you know, swimming against the tide here, right? So it's like, wow, it just like disengaged my brain and was following this thing mindlessly, even though this was the same ride that I'd taken home from work for years, right? I think with AI it's the same thing, right? AI gives you an answer and you're like, oh, a muscular right AI just gave me that answer, right? And instead of what we need to do is to treat it with a heavy dose of, or healthy dose, I should say, of skepticism, right? Is it the right answer?

[00:22:15] And the only way you can know whether or not it's the right answer is if you really have a sense for the data, right? And what the data should look like, what should the answer look like, right? And then you can kind of build confidence that AI is, when it makes mistakes, that you've got people within the organization that are going to recognize those mistakes and that are going to be able to say, oh, hang on a second. I think that, you know, we should dig in a little bit here. Where did this answer come from?

[00:22:44] How did we wind up with this answer? Question it, find out, you know, through the lineage where it came from and where the problems might lie. So I really believe that in every function across the organization, if you don't have somebody that's a data nerd, that loves like just digging into data and gets excited when they find anomalies in it, and they can tell you a story about the data, right? If you don't have that within the function, you should hire for it. I think it becomes even more important.

[00:23:11] So right now with a lot of companies very nervous about getting AI right and wrong, they're putting a human in the loop, right? They're saying, we're not going to trust agents to make decisions autonomously right now. Let's have a human just do that double check and verification. And it's like, yeah, okay, you can go ahead and take the next step here. And, yeah, a lot of organizations have those humans in the loop. But those humans, right, need to really be able to identify when things are a little bit off

[00:23:37] and to be able to give them the tools where they can kind of dig into that and see where things might have gone wrong. So to me, a strong data culture is one in which you encourage people to be data curious. You encourage people to ask questions and to doubt the data and to have the ability to look backwards and say, okay, where did this data point come from, right? It was data taken from these two tables that got joined at this point.

[00:24:02] And here's maybe a wrong decision that was made or a mismatch on data mapping that happened. But having those people that you encourage and give access to the data, that to me is like a really strong data culture. And that's what builds trust, right? So it's like, okay, well, Emma looked at this data and she thinks it's pretty good, right? And then other people are like, oh, well, she knows what she's talking about. So, you know, we can trust this data. But really important that those humans that we put in the loop are enabled with all of the tools

[00:24:32] to dig in when things don't look right and to either to confirm that the answer isn't that correct or if things are wrong, right, to be able to dig into why it went wrong and prevent that happening again. Yeah, such important points. I'd encourage everyone listening to be data curious, but question the output and develop skills to question it rather than just believe everything that it throws at you because as we all know, it can lie pretty convincingly to your face.

[00:25:00] And if we dig a little bit deeper on that, obviously the next step, agentic AI is what everyone's talking about this year. And that's raising the stakes because we're moving from systems that just generate answers and feeds them to the system to what could potentially make decisions and actions on its own. So what weaknesses in enterprise data foundations do you think agents will expose that companies may previously have not been able to tolerate or even hide?

[00:25:29] Yeah, so, you know, as we remove the humans from every step in here, right, to me, the human has always been that final governance layer, right, that gets removed now as you're dealing with agents. So to me, the most important thing that we need to do is to make sure that governance becomes executable, right, that you can say I have all of these policies in place, and the agent can understand all of your policies, but that doesn't mean that it's always going to do the right thing, right?

[00:25:57] So making sure that you've got a data architecture in place that prevents the agents doing the wrong thing by having access to the wrong data is incredibly important. That, to me, is the most important thing that we need to do with agentic AI, is making sure that we have a data architecture where governance is executable and that we're protecting the agent from doing the wrong thing, essentially, right?

[00:26:26] So data contracts become incredibly powerful here, right? Because data contract is how you make the governance executable, right? That the data contract says who can use the data for what purpose, what parts of this data can be exposed. We give different output ports on a data product. So you could say, like if I go back to my HR example, right? When I say something like, show me the total salary spend we've got, that's only from my organization, right? They'll see it holistically across all of Actium.

[00:26:55] And so making sure that the agents are not given access to data that they shouldn't have access to needs to be something that's enforced at runtime. And that's a really important role of data contracts is making sure that we're enforcing governance, that it's not a set of policies and documents that are sitting on a website that the agent's read and kind of understands them, right? But it's always going to do the right thing.

[00:27:20] I had an interesting conversation at a conference I was at last week, and it was with a CIO who'd had a $300,000 bill from one of the AI providers that his expense agent had signed off on. I'm like, don't you have limits in place where the agent can't approve something more than five grand? He's like, oh yeah, yeah, I do, right?

[00:27:47] But this was a $1,300 purchases, right? And the agent approved $300 because it's allowed, but it did it a thousand times. Now all of a sudden, he's approved at $300,000 spend. It's like, oh, that wasn't budgeted for this month, right? So really, you know, putting that executable governance in place is incredibly important, and data contracts are the way that you do that. Wow, that's an incredible story.

[00:28:16] Word of caution there for people listening. A few light bulb moments going off, and I do always try and give people listening a few valuable takeaways. So if you were sitting with a leadership team that maybe are a little frustrated that its AI initiatives are still stuck in pilot phase, any kind of questions that you would ask them to determine whether they genuinely have an AI problem or actually have a data governance and organizational readiness problem?

[00:28:45] Because there's a lot that could go wrong there, and I'm curious on how you would identify and help them move forward from that. Yeah, so this is most companies, right? That they're stuck at that pilot phase, and they're trying to figure out how do they go from pilot to production. And it's a frustration for many organizations because they've made commitments to roll out their AI projects into production.

[00:29:11] They've got some really impressive prototypes that everybody gets super excited by, and then they put it in production and things go wrong. So the first thing I would dig into is what happens when you connect that pilot to real enterprise data, right? So let's understand when you say things are going wrong, what is going wrong, right? So what are the symptoms that you're seeing in it going wrong? So that would be the first question. The second one would be, can you tell me who owns the data that the AI depends upon, right?

[00:29:41] Because if your organization doesn't have clear ownership of the data, you need that one throat to choke, right? You need that one person who's responsible for it. So understanding if the organization is set up in such a way where there's clear ownership of data products is the second question I'd ask. The third question I'd ask is, can you prove where the data came from, right? What that data means, whether it's fit for this particular use case. So understanding that is going to be the third question.

[00:30:11] The fourth, I would say, is are your governance policies executable, or do they sit in Word and Confluence pages, right? So it's really important that, as I mentioned earlier when you asked about the agentic, right? And data writing is for agentic. You have to have executable governance and contracts are the way to do that. And then the fifth is, if your AI was to make a consequential decision tomorrow, right,

[00:30:39] can you reconstruct why it made that decision, right? So looking at, can you get to the proof of why that AI made that decision, really important. So those are the five questions that I'd really challenge them with. And that will help expose where they have the weaknesses in the organization or within the data or governance that they've got. And then we can dig in and take it from there. But yeah, demoing flashy AI is super easy, right?

[00:31:08] Prototypes, you can build them half a day now with the tooling that's available. But getting that to production readiness is really hard. Because when we think about the enterprise, security is something that we've really got to consider. And when it comes to AI prototyping, these are things that they don't always think about for the prototype that when we try to roll that into production, we can have some very ugly consequences. 100% with you.

[00:31:38] And a great moment to end on there. Some real actionable takeaways. And for anybody listening who wants to dig a little bit deeper, where's the best place for them to find you, your team online, and indeed find more about anything that we talked about today, including that report we've referenced. Where should they go? So I would go to any really good cocktail lounge that tell me you. You'll likely find me surfing on an old fast. I'm on LinkedIn. I'm fairly active on LinkedIn.

[00:32:05] I post a lot of upcoming presentations that I'm doing, photographs of stuff that I've worked on. Actually, last week, I got stuck in LaGuardia for four hours just due to some weather challenges. And I wrote a LinkedIn piece on old fashions and the correlation between an old fashioned and a data contract and product. And I managed to test my theory multiple, multiple multiple times over the course of four hours. But I do have fun on LinkedIn. You'll find me there. Obviously,

[00:32:35] Actium.com has a wealth of resources, including a link to the Bark study that we referenced today. And really understanding that 3.4x figure, if you read that Bark survey, that all crystallizes it. And it becomes apparent that the only way to be successful with AI and enterprise scale is to embrace the use of data products and data contracts. Well, one thing I can promise you is as soon as this podcast ends, I'm going to be checking that article out

[00:33:04] on old fashions on your LinkedIn. And when I finish that, I will add links to everything that you mentioned there. So many big, big talking points today around data culture. The fact that nearly 90% of data professionals are struggling with scaling and complexity, but the bottleneck is not the model selection. It's enterprise data built for reporting, not the semantic richness and lineage and shared meaning AI actually requires. And we gave it a few,

[00:33:34] gave everyone listening a few actionable takeaways. And if that's not enough to sit back in my favorite chair and have a sip of an old fashioned, I don't know what is, but more than anything, thank you for coming on here, sharing your story today. Thank you. Really enjoyed it. Thanks for your time today. I think Emma's message is stop treating every AI problem as a reason to rebuild the entire data estate. Choose one business use case, create one own data product, wrap a contract around it,

[00:34:04] define quality and access rules and prove how it works. Then rinse, wash and repeat. And her GPS story is also worth remembering too, because a confident instruction can send you straight towards the exit ramp in the wrong direction, especially if you stop using your own judgment and just rely a little too heavily on the tech. And yes, AI can do the same at enterprise speed, which is why data curiosity,

[00:34:32] lineage and executable governance will always matter. And a big thank you to Emma for being a real breath of fresh air today and a conversation that somehow connected autonomous agents, revenue definitions, Lego and old fashions. And did it all without losing the plot. Kudos to my guests there. But remember, you can find Emma on LinkedIn and I'll leave links to the Actian website too. But over to you, does your organization have an AI problem

[00:35:01] or a data understanding problem that's just wearing an AI badge? As always, techtalksnetwork.com. Let me know. But that's it for now. Speak to you soon. Bye for now.