Who’s accountable when AI rewrites the rules

[00:00:00] Dr. Eva-Marie Muller-Stuler: The CEO comes back from a conference or something, they sit in the board meeting, and they’re like, oh, we really need to do something with AI. Let’s do something. And then they spin off what they normally do. They go like, okay, let’s create a roadmap. Let’s collect some use cases. Let’s talk to the vendors, build pilots, and then celebrate the demos. And that’s where they normally end. But this is not really progress. This is motion. And I actually often say, but now a proof of concept that is not designed with like the whole plan of scaling at the beginning is not just a waste of time and money. And the problem that companies are seeing is they see AI as another technology layer, but that’s not all it is. AI is something that changes how decisions are made. So, it changes the evidence they use, they change who gets influenced, who is accountable, and if you make a wrong decision, you can repeat it many, many times very quickly. So, instead, the companies really need to focus on the core questions. Don’t ask what model, ask which decision are we trying. Don’t ask how accurate is it, it’s, like, what are the consequences? And really start with the more strategic and the operational plan. So, yeah, if you cannot name the decision or the owner, the acceptable failure route, very important to stop authority. You’re not really scaling AI. You’re just scaling an unmanaged list.

[00:01:20] Jason Hemingway: Welcome to In Other Words, the podcast from Phrase, where we speak to leaders shaping how global businesses grow. I’m your host, Jason Hemingway, Chief Marketing Officer at Phrase. And today I’m joined by Dr Eva-Marie Muller-Stuler, Founder and Chief AI Officer at The Hummingbird Group. Eva is a globally recognized AI strategist and governance advisor and the author of Responsible AI Product Development. Over the last 25 years, she’s worked across IBM, KPMG, EY, all building AI capabilities and advising organizations from Fortune 500 companies through to the UN and the EU Parliament. Today, we’re going to explore what leaders get wrong when scaling AI across markets and why governance is becoming a real competitive advantage. And then, kind of what accountability looks like as AI systems become much more autonomous. Dr Eva-Marie, it’s a real pleasure to have you on the show. How are you doing today?

[00:02:14] Dr. Eva-Marie Muller-Stuler: I’m very well. Thank you for inviting me, Jason.

[00:02:17] Jason Hemingway: Absolute pleasure. So, let’s get underway. And Dr Eva, you’ve had an extraordinary career from mathematics and corporate restructuring through to being a chief technology officer for AI at IBM, a partner at EY. And now you’re actually a founder of your own firm. So, what really drew you from the world of numbers and performance optimization into AI? And how’s that foundation shaped how you approach it today?

[00:02:47] Dr. Eva-Marie Muller-Stuler: Oh, I love that question. So, I think one of the things that when I did mathematics, everyone was like, there’s no jobs for a mathematician. So, I decided to do economics and business as a minor subject to make sure that I won’t end up unemployed because everyone’s like, can become a teacher, but nothing else. And then I started my career in corporate restructuring, and I realized like, actually, everything that they’re doing is math in the background. Every decision you’re making in the company has a financial impact. And if you look at the assumptions, you can build a model that runs through and tells you the answer before you actually make the decision. At that time, it was called Financial Modeling and was, I think, the steps before. And I think that also came to a very clear understanding of how can we use information, data and everything we have to make better, faster and cheaper business decisions. We then, in 2013, set up at KPMG our first team that was called Decision Science. We were building, back then it was called Always On Machines to Support Business Decisions. And we used much smarter and much more advanced models than just Excel in the day to really help companies to optimize their decision on what are the perfect products? Can we forecast the risk of a person or of a new development? Can we take better insurance policies? Can we forecast what the policy and safety has to be in advance so that that attachment works? So, it was basically, we realized there is actually space for Decision Science as it was called back then, everywhere where we make a decision. And the other aspect that was coming with it was, looking at the data that people produce told us so much about the person itself. So, for me, it was an absolute bliss, and I’ve never had a boring day in my career using what I deeply love, the mathematics, to actually explain the world around me. And that could be people, that could be corporates, that could be countries and governments. So, that’s basically my logic behind all of that.

[00:05:14] Yeah, 

[00:05:22] Jason Hemingway: it’s fascinating. And I think there’s some other interesting parts to your backstory. And I heard you mention on a recent podcast that, you know, as a teenager, you were building electronics and playing bass in a rock band, which naturally isn’t the typical path for somebody who then goes on to advise the UN on AI governance. What did that version of you, that more creative side perhaps of you, help you understand about the problem-solving you’ve just talked about, that decision science and all of those kinds of things? And how does that show up today?

[00:05:57] Dr. Eva-Marie Muller-Stuler: Actually, I think it really does show up still a lot. It’s this questioning the current status quo of saying, “Come on, there must be a way to make it better.” And there’s this rebel teenager of me, I think, it’s still there when I’m sitting in business meetings, and I’m saying like, oh no, yes, you’ve always done it this way. But that doesn’t mean you cannot make it smarter. You cannot be more innovative about it. And I think there’s this, yeah, breaking the current status, breaking their, and even putting dots together. One of the things I realized like when I was writing my book. It’s so often about, we have these frameworks and these frameworks, but when you actually put them together, they make a lot of sense. I realized at The Responsible AI Conference a couple of months ago, the leaders were asked what they would do if they realized, like, the AI models they’re running in their company, they’re not audited, they’re not signed off, but some of them are revenue critical. And they were all governance leaders of top companies around the world, who are very forward-thinking or very AI-driven. Some of the big tech names we know. And most of the, I think over 70% of the audience said, “Oh, we would just wait and grandfather them because they’re revenue critical. We wouldn’t touch them.” And I was like, guys, that is not your job. Your job is to stop it. Because if your plane runs on something that is not audited, it might crash every second. If your AI model is signing off credits, you’ll find out five years later if they’re wrong or right. So, of course, if it’s revenue critical, it can go both ways. So, stop it now.

[00:07:51] Interesting.

[00:07:52] And I realized, like, from the mindset that we have in bribery, anti-money laundering and so on, where you also wouldn’t be like, yeah, we wait and see and tell them don’t do it again. We can’t do that. So, take that approach, that mindset, and be as clear in our industry because you have to.

[00:08:14] Jason Hemingway: Yeah, 

[00:08:15] Dr. Eva-Marie Muller-Stuler: I love that.

[00:08:17] Just bringing these completely different concepts and saying, no, actually, the underlying problem is the same. I think that’s something that the rebel inside me is still doing. It’s like, oh, I want to play in a rock band. I don’t have money for all the pedals that I need for my bass. I find a way to build them. I call my Uncle. He’ll have fun building it with me.

[00:08:40] Jason Hemingway: Love it. I mean, there is a serious undertone in what you’re saying there. Although you’re respectfully sort of challenged and that rebel teenager that you have in you. But I think we’ll come back to that kind of idea of accountability, which I think is quite important at the moment. But look, before we do that, let’s talk a bit more about kind of your career and what shaped you. And you designed IBM’s first Center of Excellence and scaled that globally. And that’s really interesting to me about how leaders should think about this idea of scaling a central model, perhaps versus what happens in those regions. And what challenges did you come up with or do organizations come up with when they try and scale, you know, AI globally and maybe outside of that Center of Excellence? Or how do you think about it today?

[00:09:30] Dr. Eva-Marie Muller-Stuler: I think in every large corporation, the people are afraid of change. And people in the most powerful positions are the ones most afraid of change because they have the most to lose. So, I was actually very surprised when I joined IBM because I was like, you’re the technology company. You should have never let any of the consulting companies of the Big Four into that market. Why are strategy houses doing AI modeling when it should be you leading the field? But IBM at that time was very focused on selling software. And literally, we had big meetings where our architects and sellers were saying, “No, we don’t want the client to think about the use case. We just want them to buy the product.” And I was like, but if they don’t think about the use case, then they will never use the product, and they will never buy again. And the answer was like, who cares? It’s a different, someone else’s problem is next year, but I want to close my deal ASAP. And really getting them around and saying, “You know, if we work together, if we look at everything we need to look at what AI brings from the strategy, from the technology, from the use case, from the change management, and all these factors around, we can actually help them transforming their organization and bringing in all of our best people, not just selling a product.” And same for the person who’s selling a strategy or a change management in people, in technology management, they can bring in our products. And if we work together on this front, we’ll all be positioning and making money. But this change management, this, I’d actually call it fear management, was a big task where I basically, in the end, had to go two roads. One of them was really working together with Global and saying, and I spent a lot of time in the U.S with our chief AI officer and so on and saying, “This is how we need to change. This is how we need to build new teams. We need to have squad units. You need to ask these questions before AI projects.” And also the other thing that I did, I found the young and greedy. And the young and greedy basically meant I set up EMEA Data Science and AI Community, basically. And I said, “We just have weekly calls. We get experts in to talk about image recognition, about all the different technical things, but also about the industry of saying, what does AI in telco mean? What does AI in healthcare mean?” And with that, I got a lot of the younger generation that is Middle East and Africa, we were per square meter. We were the biggest region in the world. But we had one person who was really interested in Nigeria, another one in Kenya, two in South Africa, but they didn’t have leaders and people around them who could help them grow. And by setting up these calls, I brought them all together. And so, in a way, I’d say I started the revolution in two ways, from the top and from the bottom. And I think in the end, now nobody, yeah, argues about what we did back then was actually really essential for changing how IBM does Data Science and AI.

[00:12:55] Yeah, I 

[00:12:56] Jason Hemingway: love it. And sort of tying this together with the next question I was thinking about was you said one of the biggest mistakes that businesses can make is treating AI as a technology project rather than, kind of, a business strategy. And when you look at that in practice, perhaps at IBM or elsewhere, what is it that companies tend to miss about that kind of idea that it’s more of a business strategy?

[00:13:22] Dr. Eva-Marie Muller-Stuler: I think, yeah, it’s one thing that I’ve seen with most of my clients. The CEO comes back from a conference or something, they sit in the board meeting, and they’re like, oh, we really need to do something with AI. Let’s do something. And then they spin off what they normally do. They go like, okay, let’s create a roadmap. Let’s collect some use cases. Let’s talk to the vendors, build pilots, and then celebrate the demos. And that’s where they normally end. But this is not really progress. This is motion. And I actually often say, but now a proof of concept that is not designed with like the whole plan of scaling at the beginning is not just a waste of time and money. And the problem that companies are seeing is they see AI as another technology layer, but that’s not all it is. AI is something that changes how decisions are made. So, it changes the evidence they use, they change who gets influenced, who is accountable, and if you make a wrong decision, you can repeat it many, many times very quickly. So, instead, the companies really need to focus on the core questions. Don’t ask what model, ask which decision are we trying. Don’t ask how accurate is it, it’s, like, what are the consequences? And really start with the more strategic and the operational plan. So, yeah, if you cannot name the decision or the owner, the acceptable failure route, very important to stop authority. You’re not really scaling AI. You’re just scaling an unmanaged list.

[00:15:07] Jason Hemingway: Yeah, I mean, and we can talk about that in a bit as well when we talk about the accountability again and risk. But it’s almost like you need governance built in in the way you think about it from the very beginning, isn’t it? Decision governance as much as technology governance. So, one of the things you’ve also talked about, particularly at your time in KPMG, where data was so granular that it carried risk before you did any modeling on it. And I think that’s an interesting thing to sort of explore. Can you just explain that kind of what happened there, so that people can sort of understand what we mean by, you know, it carried risk before you started any modeling. It’s quite an interesting thing, I think, to talk about.

[00:15:51] Dr. Eva-Marie Muller-Stuler: In a way, maybe the beautiful Wild Wild West, it was we went to clients and said, “Can we have access to all your data?” And they were like, “Yeah, here’s my laptop. Call me when you’re done.” And they were like, “Oh, do you need like an hour, or do you need a day on my laptop?” So, there was no understanding of what data actually is, the value in it. We had clients saying when we turned up, I remember one client meeting, where the client where we were getting the data from, it was a beverage distributing center. He spent half an hour talking about how upset he is that his printer doesn’t work. And I was like, I was terribly jet lagged. And I really tried to understand what he was on about until my colleague was like, no, no, he just doesn’t understand why we’re here for, and he just goes technology – printer is your field. And funny enough, while we were waiting for the data to download, I actually did end up fixing the printer just because I had the time to do so. But we called one of UK’s biggest telco companies. We were working with them, and we said, “Actually, can we have all your data?” And they’re like, “Oh, but that’s a lot. How?” And we’re like, “Oh, we sent over the intern with flash drives.” And they’re like, “Okay, perfect.” And so we were sitting there, and we built this mass connected ecosystem where we connected healthcare information, Transport for London information, street maps information, property prices to understand if they’re going up and down in certain regions, then with footfall. And especially like everything we knew about who called whom, at which time, from which place, we knew everything. And we were sitting there and joking and saying, “You know what? We can just filter and find out what people are doing when they are at places where they shouldn’t be.” Are they staying in hotels close to their home? Are they cheating on their wives and whatever? And make so much money off of other information because it was all personal, identifiable. And even healthcare information, where you could actually download it from the government’s website at that time. They were on postcode level. That is like 11 households. 

[00:18:18] Jason Hemingway: If 

[00:18:19] Dr. Eva-Marie Muller-Stuler: you then know that’s a postcode, that’s the pharmacy, they go shopping. And that’s where I see this cell phone. You know, who was it? So, yeah. And I think that’s also what really then got me into working with the government and saying, “You know, we’re good people, we have no interest in hurting people, but if we were bad people, we would have a massive amount of money and power.” And that is not something that should be accessible. Later, we had GDPR, so things like that actually, then, became against the law. But overall, I think we are still far behind on the laws and legislation we need to have to govern what we’re doing.

[00:19:12] Jason Hemingway: Yeah, I think that’s interesting. I mean, like you say, Wild West at one point, but still an issue, albeit, you know, perhaps not quite as alarming as the case you pointed out. But let’s take it down to a more sort of global content level, when you think about AI-generated content and content moves across languages and markets, do you see a similar need? The case you pointed out was a highly sensitive risk. But when you take it to content, do you see a similar need to think about how it’s governed as it scales?

[00:19:47] Dr. Eva-Marie Muller-Stuler: Yes, absolutely. So, one big problem we have with technology, or it’s actually a beautiful thing, but it’s also a problem. It has no boundaries. And it also has a lot of the GenAI that we see online is built either on Chinese demographics or American demographics. So, all the content we are producing is heavily dominated by white, rich, democratic demographics. So, then we enroll these models across the world, and people are using them to ask legal advice and not even including the jurisdiction. We use them for healthcare advice, even though healthcare is highly dependent on our behavior, and our eating patterns and so on. So, the advice you would get is always the advice that a white male American should get. And it might not work for an Emirati woman. I say Emirati because I live in Dubai, women over 65 because they’re not represented in that data set at all, and so we see these issues, but we also see issues, for example, with translating, that things that are absolutely fine to say in America are not okay to say in Japan, or look at Germany. We just had a beautiful scandal with our German train system that our apprehension for one hour is the German word for hour is ‘Stunde’, and we often shorten it with STD. And so the German model actually used the German word for STD instead of hour, which was very, very funny and ended up in all the newspapers. And I mean, even though on the one hand, it’s a funny newspaper anecdote. On the other hand, it was about medical advice, and the doctor said something, “You should come back in one hour.” And you think you got a medical diagnosis or something else. And these are, like, just small cases. So, there isn’t really a clear governance on how these things get audited. I think the one thing I’m most worried about is that our younger generation, the 16 to 20-year-olds, I think over 80% use ChatGPT for psychological advice. Or I say ChatGPT, but all kinds of large language solutions. And these are advice that has never been audited. We don’t know what they were trained on. We don’t know the data that goes into it, and a lot of the data in the Internet is highly biased, is highly misogynistic, is also, I often say, the more time you have, the less educated you are. The more time you have commenting on things on Reddit and websites like that, and my favorite thing about these online advices is, “I have no idea about this question, but let me tell you, I think the answer should be this and this.” And I’m like, “If you have no idea about what’s talked about, why do you answer?”

[00:23:20] Jason Hemingway: Yeah. 

[00:23:21] Dr. Eva-Marie Muller-Stuler: Yeah.

[00:23:22] And I think that training models on that, and these models then coming back and saying, we advise you to break up with your girlfriend or walk away from your friends or whatever where they don’t know the whole scenario, the cultural background, or anything going on. It’s extremely risky what we’re doing with our younger generation there.

[00:23:43] Jason Hemingway: Yeah, I think you’re right. And I think everyone can cite a funny example, like the example that you said in Germany, but it becomes very, very not funny very quickly when you start thinking about the harm that it can do. So, which is why you need governance, governments’ guardrails, all of those kinds of things that are coming, but probably not quick enough.

[00:24:05] Dr. Eva-Marie Muller-Stuler: Absolutely. And I think one thing is the governments being aware of having to do that. But also, every company needs to do it themselves. If you don’t govern your AI, it doesn’t work. It’s decoration.

[00:24:25] Jason Hemingway: Yeah.

[00:24:25] Dr. Eva-Marie Muller-Stuler: And you need a control system in there. And if you don’t do that, you’re running a massive risk. You’re flying a plane blind.

[00:24:31] Jason Hemingway: I am 100% correct. I think there is a liability on the company to actually provide the information correctly when they’re doing it, you know. Anyway, let’s lighten the mood a little. And we call it the mid-show moment, and it’s kind of an inbox confession, and we changed it up. Usually, I ask people what they would automate in their lives, but I’m not going to do that this week because lots of people are saying calendar and this. It gets very similar, and I have people seem very time poor. It’s what I kind of gathered. But a different question is, what’s something that you’re genuinely bad at that you’ve just decided, okay, I’ve made peace with that now. I don’t mind being bad at that anymore.

[00:25:11] Dr. Eva-Marie Muller-Stuler: One thing I am bad with is, in a meeting, from political reason, I should just say yes. I can’t. And when I’m against it, I always have to say the truth. And I do realize like I might have lost clients about this honesty. But this rule, yeah, sitting there saying. This is not a good idea because of ABC. And just agreeing and saying, okay, no, nobody will find out if you just do it, and we cover it later. That is not my style. And I think in the past, I sometimes said like, okay, it would have been easier for me if I just had the ability of just wave it through, say, let’s find a way, but in the end, I think it also, pushing through weak ideas because they’re convenient is not my style. And I think in the end, it builds a lot of trust with me because the project that I might have lost because of that, they failed. And so in the long run, people come back and say, “Oh, Dr Eva, you pushed back on that project. It failed.” And I was like, “Yeah, because I was telling you at the beginning.” But, you know, if you turn it around, we can still make it successful with ABC, or I had the feeling at the beginning that your stakeholders were on the wrong page. I had clients asking me, like, “If you don’t have any data, what can we do? And can we not just generate data?” I was like, “You know, synthetic data only works when you use less than 30% in certain cases. But if you have nothing to stand on, if you just make numbers up, the algorithm tells you what you made up. It doesn’t tell you the truth. It goes one way and then back.” And I think I’m actually also quite glad that these things didn’t go forward because the delivery would have been hell. The trust wasn’t there at the beginning, and yeah, for me, it’s actually a way of respecting my client that I’m not gentle with false confidence.

[00:27:34] Jason Hemingway: I love it. I think it’s that teenage rebel serving you very well to be truthful and honest and keep things. And also, as you say, if you have that belief that it’s not something you’re going to do and you come back and you see that if you’d agreed to do it that way and it’s failed, then it kind of compounds your thinking and sort of, you know, that you’re right.

[00:27:53] Dr. Eva-Marie Muller-Stuler: It was really something to come at peace with, exactly as you asked, because I had moments of colleagues going, “Why did you say that?” I had meetings where my colleague would say, “No. No, we can guarantee you that the model will be 99% accurate.” I’m like, “How? We haven’t even seen the data.”

[00:28:11] Jason Hemingway: Yeah, I think it’s a good trait. I don’t think that’s a bad thing. So, let’s talk about, you wrote a book on Responsible AI Product Development, and just for readers who haven’t read it, what’s the single most important thing about AI governance that most organizations still get wrong?

[00:28:32] Dr. Eva-Marie Muller-Stuler: I think the most important thing is really, it’s not a policy document that you have somewhere in the draw, it’s a clear control system, and it goes from right at the beginning of what should we do, what is our AI strategy, all the way through to data, responsibilities, visualization and only if you have all of these blocks in place then you can have successful AI. But if, for example, the system uses data without permission, the pipeline should block it. If a feature changes the meaning, you need to realize that, and you should block it because your model will always silently accept all of that. If your subgroup that you’re looking at is changing, your model will just keep on going and stop working. So, you really have to be clear. There are so many issues. And I actually, just for fun, started writing down all the reasons why I saw client projects fail. And I think after 300, I just decided, no, I can’t.

[00:29:53] It’s 

[00:29:54] Jason Hemingway: too depressing.

[00:29:55] Dr. Eva-Marie Muller-Stuler: Yeah, I started structuring and so on. But it can fail at every single place. And at every single place, you need to have people there who are empowered to pull the line, who say no, it doesn’t work. And, oh, I saw the data is changing. Oh, I thought the use case is not working. Oh, I thought the technology is wrong. And then say, let’s stop here and say how we can fix it.

[00:30:25] Jason Hemingway: It’s interesting because you’re talking about where people get involved, and you know, the kind of phrase du jour, as that human in the loop side of things. But as things get more autonomous and AI makes decisions that affect, you know, employees, customers, even communities or, you know, citizens, let’s say, in the governmental sphere, wherever they are in the world. When you look at that kind of, okay, you’ve got a human who can stop things, but let’s take it above and sort of say, well, what’s the accountability, and where should that be? You know, should it be the person building it, the person who’s operating or overseeing it, as you’ve just said, that can kind of say no, or is the accountability, the leadership team, the CEOs, the board even? Where does that accountability sort of land for you? Or is it all of those?

[00:31:15] Dr. Eva-Marie Muller-Stuler: I have to say it lies on all the levels, and I think it really depends on, like, very similar to the anti-money laundering or bribery, something that the CEO doesn’t take serious, it will fall through the whole organization. And that’s why it’s important to start at the top, but that’s where you set the values. But of course, if one of the junior levels goes like, “Oh yeah, I saw the data is going the wrong way, but I don’t want to rock the boat.” Then you can say it’s his or her responsibility. But also, it’s a question of what is your culture and why did you hire that person? So, I think the idea of having rebels in your company is becoming more and more important. And people who feel safe, that if they speak out, they’re not in danger. They’re still safe, and that is actually their position. And so, I say accountability on every single level. But it starts from the head.

[00:32:29] Jason Hemingway: So, let’s move it a little bit onto some of the things that you’re talking to both governments and Fortune 500 companies about. I thought it was interesting. What do you think business leaders can learn from government when it comes to deploying AI responsibly? What have you seen in that sort of sphere?

[00:32:48] Dr. Eva-Marie Muller-Stuler: I think governments have a very interesting position at the moment. They’re being, first of all, the body who is doing the loss, but then the body who’s financing a lot of the innovation and research, then they are massively attacked with fake news, and then they’re a user of AI. So, they’re actually wearing all four different hats. And I think what business leaders can learn is, first of all, I think there is no way around good governance. No matter how small you are or how big you are, if you’re a country, you should actually be ahead of the laws, because we don’t have the laws yet. If we have the laws, we have no way of enforcing the laws, which is a big problem. We all know that the large language models that we’re using out of the U.S and worldwide are probably not GDPR compliant. We don’t know where they have that data from, but we, probably, they have enough data where we kind of think, like, it cannot be that GDPR compliant. But there’s nothing you do about it. We’re just saying, “Oh, we have the laws, and we have the reality. Let’s step back.” And I think for companies, it’s important to not just see the legal risk, but also see reputational risk and financial risk as one of the big problems they can face by having bad AI. And so I think the company should have their own rules, of saying this is what we’re willing to do, and this is what we’re not willing to do. They need to have their own way of monitoring it and saying, “This is an accuracy, or this is whatever we are not okay with.” And I think there’s also the risk of being attacked by fake AI, it’s not just for politicians, it’s also for the company overall, of having fake video calls, of having failed sign-offs, of the whole cyber threat. So, I think that there are many layers where we all need to work together, and we’re actually on the same page. And I really don’t like the idea of strong laws and rules hinder innovation, because I think strong rules and laws enable innovation in the right way, with the right focus. So that you don’t end up in the end building something that you then have to shut down.

[00:35:42] Jason Hemingway: That’s interesting. And I think once you take that to a sort of international level and regulation moves at different speeds in different markets, obviously, you know, and then particularly AI regulation. So, we’ve talked about how businesses can look at it from a holistic level. When you’re talking about multinationals preparing for each market, what does that mean in your view, how they structure, say, I don’t know, automated content, as we talked about before, or just AI operations globally?

[00:36:11] Dr. Eva-Marie Muller-Stuler: And I think one thing that they really should not do is the sign for the weakest regulatory environment. So, there’s always this talk about the risk of, oh, if we have very strong rules in the UAE, people will just go to Saudi, or if you have very strong rules in Saudi, people will just go to the UK. That could be, but it’s actually a strategic mistake. Because in the end, you want your AI to work everywhere, and not to be shut down. So, I think we need to have a global control standard. And I think most of the world is actually aligned when you listen to what we’re saying. We just have to find a way to align how we put it into place and how do we enforce it. No country wants AI to do harm. No country wants AI to be extremely biased and so on. So, we do agree pretty much on the basics. We just need to say, how do we build an AI governance model that is not fragmented? And we don’t want to have different tools everywhere. We don’t want to have different approvals and so on. It might look like it’s flexible if you do that. But in the end, it will be impossible to manage for every international organization.

[00:37:45] Jason Hemingway: Yeah. So, let’s dive into that a little bit. So, if you’re thinking about companies under pressure, lots of pressure to move fast on AI. And they often face this kind of build-buy decision. How should leaders decide whether to invest in kind of a platform to help them do that or multiple combinations of different solutions? And what sort of makes that difference between a tech foundation worth building on, or rather than this other idea, which you said, which is they’re all becoming liabilities because they’re all doing different things all over the place?

[00:38:19] Dr. Eva-Marie Muller-Stuler: I actually think when you start off with the question, build versus buy, you’re already on the wrong foot, because you should start with the question, what do we need to control? And when you don’t need to control it, you can definitely buy it. But if you need to control it, if it, for example, defines your competitive advantage, defines your user experience from a regulatory exposure, all of that, then you need to have control over it. If you outsource that, you make yourself dependent on somebody who’s able to control you. In the end, if you rely too much on them, they can raise the prices whichever way they want because you gave something away that you need to control. You definitely need to bring it back in-house. If it’s just a commodity, then buying it is absolutely sensible. There’s no need to build your own large language models to make emails faster or something. Something is offered so many times, if one of them raises a price, does it just fits to the other model? But you have to be very careful about the vendor lock-in trap. So, if it’s something that gives them the opportunity to control them. It’s a dependency you don’t want. So, really ask yourself questions like, can we govern it? Can we integrate it? Can we audit it? If it’s something you like, when you get financial numbers, and you have no idea where they come from, you make yourself dependent on something that is a black box that you don’t know. Can you in-house scale it economically, or even when you buy it? And then how do you exit? And that is, I think, one thing that I see as a big risk at the moment of a lot of people building companies, building GenAI solutions on one of the vendors. We all know that most of the GenAI out there is making a loss still, so there will be a moment in time when they need to increase the prices heavily. And at that time, you have invested a lot of time and money in building something on a cheap assumption. Will they suddenly 10x the prices. The question is, do I carry on, or do I switch?

[00:41:03] Jason Hemingway: It’s where that vendor lock-in becomes important.

[00:41:06] Dr. Eva-Marie Muller-Stuler: You have opportunity by speed, speed to market, but never ever outsource any accountability or anything where you need to have control over.

[00:41:14] Jason Hemingway: So, let’s talk about, you know, skills and talent and teams. And you’ve mentored something like over a thousand women through Women in Data Science and built communities around AI talent. So, how do you think AI is reshaping skills that leaders need today?

[00:41:31] Dr. Eva-Marie Muller-Stuler: I think, actually, in a long way, because I think two things will become far more important. That is the confidence to decide and also flexibility. And we now have, I think, the fastest change ever in skills and what we need in the market. Whenever you don’t read the news for two months or a month, you feel like, “Oh my God, how much did I miss?” And I think when we look at leadership in the past, what was always rewarded was intuition, communication, and especially authority. Like, oh, listen to the senior person or to make decisions under uncertainty. It’s like, I don’t care if we do it this way. They still matter to some degree, but not as much as before. Those who, like, listen to the most senior person, they automatically rewrite something that is completely breaking, falling apart at the moment. We now need leaders that actually have evidence literacy. So, saying, “Oh, we see this and this and this. My gut feeling is completely different, but I see this and this and this. Is the model wrong, or am I wrong?” So, challenge yourself, yes, but also try to find the errors. And I think in the past, there’s understanding the difference between correlation and causality, understanding automation and control. All of this, how do I place myself in the system? How do I challenge my system? How do I challenge myself? These are all questions we have far more in the future. And I think that comes down to the hardest skill that I see is humility of saying, “Is this right? Is it dangerous? Is it right for everybody? Just because it works for me and it’s good for me doesn’t mean it’s good for everybody. And so one of the biggest things I see in managing risk is highly diverse teams. And that’s one of the things I still miss in a lot of Data Science and AI teams. I still see them as heavily male-dominated. The more diverse the team is, the more they’re actually able to find errors and saying, “Hold on, it doesn’t work for me. Or hold on, I completely disagree. Or my mother would not be okay with this.” And so it’s, I think, yeah, the most dangerous executive in the AI is the one who knows nothing but thinks he knows everything. And that is far more than in the past. And the one who basically just sees a demo and goes like, yeah, I understand it. I understand the risk. That’s a huge problem. And I’ve seen it so often. It’s the one who says, “Yeah, I’ve seen the demo”. But it’s just a demo. What are the problems in scaling? “Or I see the outcome”. But they are so against what I’ve learned in the past. Is it me wrong or is it the model wrong? And how can we find the reasons?

[00:45:10] Jason Hemingway: Yeah, it’s interesting, that kind of expertise side of things and staying open. And one of my questions, or one of the final questions I have, really is this. You described that it takes it to another level. And the opposite side of that is you described a client whose only expert was a retired geologist that they couldn’t replace, because the AI model depended completely on that individual’s knowledge. So, when you think about that, and automating content and AI across global markets, how important is it to retain some kind of local expertise and keep it in the loop as that all scales?

[00:45:47] Dr. Eva-Marie Muller-Stuler: Huge. And I think one of the things we are seeing is we are in the market, we are making it extremely hard for young talents to enter the market, saying, “Oh, we don’t need you to have AI.” But then we see massive mistakes of like consulting companies sending out completely AI-generated content to their client. Well, like, where was the senior? Where was the partner stopping that and reading it through and taking the time? I think the senior knowledge of really saying, “This is how the model, this is how the reality works.” I have seen these and these things going wrong, even in fraud detection, in forecasting and so on. Some of it are very rare events that we’re looking for. And the senior experience becomes more and more valuable, in my opinion.

[00:46:42] Jason Hemingway: Yeah. Yeah.

[00:46:43] Dr. Eva-Marie Muller-Stuler: But we only get the senior experience if you train the juniors. And that is something when we look at it, we always go like, oh, the young generation, they’re not loyal to their companies anymore. I’m like, yes, because the companies are not loyal to them.

[00:47:00] Interesting.

[00:47:00] I remember my junior years, I had constant trainings, and they were face-to-face in a room, not, oh, go online, where I had to like spend, like, yeah.

[00:47:11] Yeah. Self-guided learning. Yeah, yeah, yeah.

[00:47:13] We expect our juniors now, yeah, on top of their workload. And then we’re surprised that they don’t find the time because on their 60 hours a week, they don’t want to spend another 40 hours training, or 5 hours training.

[00:47:23] Interesting.

[00:47:23] Take them out of the room. Make them build a relationship to their peers. Take them away somewhere nice, like we did in the early 2000s, 90s, and that creates loyalty to the company. Develop merchandising. Make them feel like emotionally connected and taken serious as a person. And when we don’t do that, of course, they move to the next best gig. Of course, if somebody offers them another hundred dollars, they jump boats, because that’s the only way they have, the only point they have to see how valued they are.

[00:48:04] They’ve 

[00:48:04] Jason Hemingway: got no frame of reference either, have they? So, they don’t know how it could be, but, okay, so let’s get towards the sort of final question I have ready for you. So, this is for any executive listening today and trying to scale AI responsibly across all of those markets, languages and different customers. What’s the one thing that you’d say focus on first to them? 

[00:48:26] Dr. Eva-Marie Muller-Stuler: AI for decision quality. So, it’s not about chatbots, it’s not about content generation or dashboards, it’s how can you use AI to improve the quality, the consistency, the cost of your core business decisions, pricing, risk, customer service, forecasting, capital allocation, and if you understand your decision infrastructure, then the conversation becomes very serious very quickly. So, I think decision quality is where it can be your best support and your best sparing partners and where AI really can improve the decision that your business is doing for it. So, the classical Decision Science, Data Science, classical AI is, I think, far more important right now than the 

[00:49:29] Jason Hemingway: GenAI.

[00:49:29] So, just two more questions before we go, really quick-fire ones. We always have these at the end. Global growth in one word?

[00:49:35] Dr. Eva-Marie Muller-Stuler: I think when we can trust our AI, we can scale faster. And, yeah, global trust becomes very expensive without trust.

[00:49:48] Jason Hemingway: And then finally, what’s the leadership trait, I think I know the answer to this, given what you’ve said before, that the AI era will reward most?

[00:49:57] Dr. Eva-Marie Muller-Stuler: Yeah. Adaptive judgment.

[00:49:59] Jason Hemingway: Yeah, there you go. I thought you’d say that.

[00:50:00] Yeah. To judge and also adapt, or 

[00:50:05] Dr. Eva-Marie Muller-Stuler: how you adjust.

[00:50:06] Jason Hemingway: Okay, well, look, Dr Eva-Marie, thank you so much for a really fascinating and fantastic conversation.

[00:50:11] Dr. Eva-Marie Muller-Stuler: Thank you. I really enjoyed it.

[00:50:14] Jason Hemingway: Me too. See you again. Bye.

[00:50:16] Dr. Eva-Marie Muller-Stuler: Bye.

[00:50:17] Jason Hemingway: Dr Eva-Marie, thank you again for a brilliant conversation. You gave us an interesting, clear view of why organizations that scale AI successfully are the ones investing in governance and their people and the right foundations from the start and asking the right questions. And with that, that’s it for another episode of In Other Words, a podcast from Phrase. I’ve been your host, Jason Hemingway, and a big thank you to Dr Eva-Marie Muller-Stuler for joining us today. 

[00:50:45] Dr. Eva-Marie Muller-Stuler: If this episode made you think about how your organization governs and scales AI across different 

[00:50:50] Jason Hemingway: languages and markets, we’d love to continue that conversation. So, please head to phrase.com to find out more. And for more conversations on leadership, growth, and what it really takes to scale globally, subscribe to In Other Words on Spotify, Apple Podcasts, or as usual, your favorite podcast platform. Thanks for listening, and see you next time.