The Truth About AI Data

[00:00:00] Dr. Eva-Marie Muller-Stuler: You need to have the right technology. But the first question where they go wrong is: Do we actually have a strategy? What are the use cases we want to focus on, and how will we build a clear roadmap on what to focus on first?

[00:00:12] Intro: Technology has transformed how we live, work, and connect. And behind that change are visionary people with bold ideas. Welcome to the BridgeCast, where host Scott Kinka explores hybrid work, AI, and our evolving future with tech’s brightest minds. Presented by Bridgepoint Technologies, the nation’s leading tech advisory firm. Experience a better way—the Bridgepoint way.

[00:00:47] Scott Kinka: Welcome to the BridgeCast. I’m joined today by Dr. Eva-Marie Muller-Stuler, who is the founder and Chief AI Officer at the Hummingbird Group. Eva has spent more than 25 years helping global organizations turn emerging technologies into measurable business outcomes. She previously led major AI initiatives across some of the world’s most influential organizations, from Ernst & Young to IBM, where she developed their first AI Center of Excellence. She’s also advised the United Nations, UNESCO, and the European Parliament on responsible AI governance and risk mitigation.

She’s been named a “Top 10 Most Influential Woman in Technology,” “World’s Best Data Scientist,” and “Top Brilliant Women in AI Ethics.” I mean, she’s a literal rock star, and we’re happy to have her on the show. In a moment where AI promises enormous opportunity but also real risk, she is helping set the standard for responsible innovation. I couldn’t be happier to have her on the show. Dr. Eva, welcome to the BridgeCast.

[00:01:49] Dr. Eva-Marie Muller-Stuler: Hello, and thank you so much, Scott, for inviting me and for that really nice introduction. I’m really excited about spending this time with you.

[00:01:57] Scott Kinka: Certainly. I usually ask at this point: How did I do on the introduction?

[00:02:01] Dr. Eva-Marie Muller-Stuler: Oh, very well. Yeah, I was impressed by what I heard.

[00:02:07] Scott Kinka: Well, you’re very impressive, and I appreciate that. But here’s what I would love to do: Tell us something about you that we can’t learn on your LinkedIn profile.

[00:02:16] Dr. Eva-Marie Muller-Stuler: One thing that you don’t really see enough of on LinkedIn is how much I really care about people. I think one thing we see with AI is always the competition of “AI destroying the world.” But for me, I do care about environmental topics, ethical topics, and equality. I think that is something that doesn’t just go through my LinkedIn; it actually goes through my life. I love animals, but I would also like to fight for fairness and justice.

[00:02:58] Scott Kinka: We may get into some of those topics around AI, because it’s a sort of charged environment given some of the topics you just talked about. You are currently in London, correct?

[00:03:08] Dr. Eva-Marie Muller-Stuler: I am, yeah.

[00:03:09] Scott Kinka: London is not home?

[00:03:10] Dr. Eva-Marie Muller-Stuler: London is not home. I got stranded here about 18 days ago. We were suddenly surprised by the conflict in the Middle East, and my flight back was canceled. Now I’m sitting here waiting to be able to go back home.

[00:03:26] Scott Kinka: Gotcha. So, Dubai is home for you?

[00:03:29] Dr. Eva-Marie Muller-Stuler: Yeah, I love the sunshine. I love the warm weather, and it’s definitely a good place to live.

[00:03:36] Scott Kinka: It’s the rainy season in London right now, too, so you’re smack-dab in the middle of it. You told me in the pre-show you’ve left a parrot back in Dubai. Somebody’s taking care of the parrot?

[00:03:47] Dr. Eva-Marie Muller-Stuler: Yes, I left my little boy. I miss him daily, and he’s very sweet. I can’t wait to have him back.

[00:03:56] Scott Kinka: Well, I appreciate all of that, and thank you for being on the show. We wish you luck getting back home. We scheduled this before the conflict started, and here we are on the phone—you’re still taking the call from London, so I appreciate that. Thank you.

Before we jump all the way in—because there are so many topics we could explore—we’re going to get into technology and governance in particular. But tell us a story from your journey that shaped how you think about innovation in general. Maybe some seminal moment in your career that serves as the basis for how you think about tech today.

[00:04:34] Dr. Eva-Marie Muller-Stuler: I have always been very technical. Even as a child, I loved building electronics with my uncles and my father. I was playing in a rock band as a teenager—I played the bass—and when everybody had these amazing pedalboards, I had to build them all myself because I couldn’t afford them from my pocket money. In a way, I learned a lot about how to actually build electronics.

I then studied mathematics, mainly because I loved it. People said, “What? There are no jobs for mathematicians! Why don’t you take something sensible like business or computer science on top?” So, I studied the combination of all three. I’m still a mathematician at heart; it’s how I attack problems, how I think, and the logic behind it. It’s still my passion.

I started my career in financial modeling. For a very long time, the way I saw AI was as corporate performance improvement. I still like to make a split when companies are thinking about investing in AI. One path is building a research team if you really want to, but the other is building capabilities purely focusing on corporate performance: Do we make money? Do we manage our risk? Do we decrease our costs? If you aren’t purely focusing on research, you should not mix them. It’s a different kind of work and a different kind of people you need to hire for both.

I think where my thinking really shifted was at KPMG in 2013, when we started to build a team focused on building “always-on” machines for decision support. It was the beautiful Wild West where companies had no idea what data was. We would call them and they were like, “Okay, you can come over and take all my data. How do I give it to you? I’ll give you my laptop and go for lunch, and you just take what you want. By the way, the printer is broken—can you fix it as well?”

In the end, we asked one of the world’s biggest telecom companies if we could have all their data—who called whom, from where, and who paid which bill. They just asked, “How?” and we said, “Oh, we’ll send the intern.” So we sent the intern to buy flash drives and pick up the data. We didn’t sign an MOU; we didn’t pay any money. For days, we were sitting there thinking, “Whoa, this can’t be okay.” We had so much knowledge about individual people. We didn’t even need to build AI models anymore; we could just blackmail people by filtering the data.

That’s how I personally got involved with governments. GDPR is a really good first step. It was pre-GDPR back then, so it was completely legal, but I realized there were so many issues in building data-driven AI models. Even if we remove religion, for example, there are so many proxies for it. Does somebody never buy pork in the supermarket? You have a proxy for religion, more or less. Do they eat shellfish? What day are they shopping? I realized that governments need to be aware of this, and we need clear rules and regulations. That’s why I started working with the UK government, the EU, the UN, and other NGOs to talk about AI safety, rules, and governance.

[00:08:46] Scott Kinka: Super interesting. We’re going to explore all those topics. But before we double-click on that story, the thing that struck me was your example: “Here’s all the data, and can you fix the printer?” That’s a great example of what we see when talking to customers today. Historically, IT lived in a bucket where they stood up the CRM, and as long as you could access it, they were out; the business owned it. Now, business leaders look at AI as a technology and think IT can lead it, but IT doesn’t necessarily have perspective on the problems trying to be solved inside the business or what’s actually in that data. Who owns the risk of doing AI wrong?

[00:09:43] Dr. Eva-Marie Muller-Stuler: I always say AI needs to come out of a business need. I have seen so many data and AI teams that came out of IT—quite often involving people they didn’t want in the IT team anymore, or perhaps people who were just more innovative. I was with a client in South Africa one day, and they wanted to build a Hadoop cluster to do social media analysis to show the business what they were doing wrong. I said, “Your job is not to be against the business.” Even if you’re right, the business will push back and your AI will never work.

But also, I was sitting there thinking, “Why a Hadoop cluster?” Do you even have a social media handle? Their customers were over 80 years old! I asked, “Do you know what inventory you have?” They had zero data quality in-house. They thought social media analysis would give them new insights, but that doesn’t work—especially if your demographic isn’t even on social media. There is a big misunderstanding of the required skill sets. Traditional IT departments often aren’t the right people, neither in terms of technical skills nor strategic view.

[00:11:04] Scott Kinka: I got it. Let’s double-click on that. You’ve spent your career building AI systems at scale at KPMG and IBM. As we just said, most enterprise leaders think about AI as technology, but you’ve argued it’s something else entirely. What are they misdiagnosing? What is it at the end of the day?

[00:11:30] Dr. Eva-Marie Muller-Stuler: You need to have the right technology, but the first place they go wrong is not having a strategy. What use cases do we want to focus on? How will we build a clear roadmap? The best way is doing this together with the business in a design thinking workshop. Ask the business: “What are the things we can fix?”

The business can go crazy and suggest sensors all around the world or things that don’t make sense technically. Then technology says, “That’s a good idea, let’s park it for later; it’s not feasible because we don’t have the data. But this is a good idea, we have the data, and it’s easy to build.” That’s how you find the low-hanging fruit. You have to work hand-in-hand with the business, not against them.

[00:12:30] Scott Kinka: Agentic AI is all the buzz right now. Autonomous workers that are just going to do it on their own—are you buying that story?

[00:12:46] Dr. Eva-Marie Muller-Stuler: The big risk I see—and it’s funny because I’m writing a book about how to correctly build AI—is that most books on the market say you need “good data,” but they don’t tell you what that actually means or go into the technical depth that professionals need. When we look at agentic AI, you cannot run it if you don’t have the right information.

A lot of companies are trying to cut corners. It’s like buying a car and leaving off the wheels; it will never drive. I see that often with AI and agentic AI. It can work if done correctly, but would I buy it as the first low-hanging fruit? In most cases, no. There are other problems where you can focus your money more wisely to have a bigger impact. Everyone has Copilot or a spam filter now, and those definitely increase productivity. But as for taking humans entirely out of the loop? I would be very careful about whether that’s the right starting point for your company.

[00:14:40] Scott Kinka: Let me give you a situation because we run into this often. People in customer experience want bots to start answering the phone, email, or chat. The first question is always: “Where are the processes documented?” They usually say, “Well, we taught over the shoulder, but we have these call recordings and transcriptions.”

Provided you are validating that what was done was correct, is that a way to backfill missing documentation to inform a model?

[00:15:42] Dr. Eva-Marie Muller-Stuler: There’s a big problem with leakage in chatbots or customer centers. Often, we only get escalated things that go wrong for a certain reason. We don’t see the things that go “semi-wrong” or right, but leave the customer unhappy. It’s actually very difficult to train models to do it right. The world changes, so if the documents aren’t updated and the model isn’t constantly retrained and monitored, you end up giving outdated or wrong advice. The frustration for people can be extremely high.

In call centers, 80% of questions are about the same 20% of topics. That is very easy to automate. But there should be a strong security monitoring system. We see it with ChatGPT—sometimes you get frustrating or wrong answers. If you want to keep the customer, you need a way to interfere. If they ask the same question three times and aren’t happy, a human should call them back. Then, take that interaction and ask: “How can we train the model so we don’t get caught in that loop next time?”

[00:17:31] Scott Kinka: Got it. And I guess that is your answer to autonomous workers: Keep the human in the loop and check the work. We’ve talked about where it goes sideways—what does “good” look like? What does a well-designed AI deployment look like at scale?

[00:17:50] Dr. Eva-Marie Muller-Stuler: I like it when the end-to-end function works. They set up a good Center of Excellence with top-notch, experienced people. Nowadays, we often see experienced people leaving companies or being made redundant, but their expertise is vital because they have failed so often. They know what to look out for.

A good team has a close relationship with the business, focuses on the right use cases, and has strong sponsorship. The best AI solution doesn’t work if the company doesn’t accept it. You need support from the CEO to say, “We know it’s expensive—we don’t get anything good for $10,000—but we are willing to invest in it and in our people.”

Also, understanding that data for AI is not the same as data for analytics. If a police officer is lazy and enters “0” for a criminal’s age, an analytics dashboard will show that it doesn’t make sense and filter it out. But an AI system will learn from it and say, “We have a hotspot of newborn criminals in maternity wards,” and suddenly police cars are being dispatched to hospitals. Finding the reason for that is complicated “archaeology.” If I see a company that understands how to monitor data specifically for AI, I’m very impressed.

Finally, the outcomes need to be communicated so the business understands them. I once had a friend whose GPS said “turn right at the next one,” but he didn’t know if it meant this turn or the one further down. It was a good AI running in the background, but it didn’t do him any good because the communication was unclear. Corporate AI often fails at that last bit of change management.

[00:21:18] Scott Kinka: You mentioned your focus on equity. We have five generations in the workplace right now. I’ve heard it said that AI isn’t necessarily problematic for older generations because they are the “power informers”—the tribal knowledge lives with them. Is that an equity opportunity for businesses?

[00:22:08] Dr. Eva-Marie Muller-Stuler: Definitely. I worked with a client on the classification of stone samples from drilling. Their biggest problem was that the only person who knew how to classify them was an older gentleman who had retired ten years ago. They couldn’t let him go because he was the only one who could look at a sample and say, “This is A, B, C, or D.” We trained the AI model based on his knowledge. But if he leaves and no one else has his skill set, there is no way to monitor, review, or retrain the model if things change. Companies run a high risk of making these people redundant, only to have to buy them back later for a lot of money.

[00:23:33] Scott Kinka: Let’s bring it back to practical tips. If I’m a CIO and I want to start fixing or starting an AI deployment this week, what are the three building blocks?

[00:23:53] Dr. Eva-Marie Muller-Stuler: First, understand what you want. Do you want to buy something or build it in-house? And why? Can you buy something “good enough”? In the Middle East, for example, fraud detection machines are often bought from the US. They don’t work well because they aren’t trained on Middle Eastern data. It’s the same with healthcare; we buy US solutions, but there was likely no dataset of an Emirati woman over 65 in there, so we don’t know how it will perform for our demographics.

Classify by risk. Start with something easy that doesn’t jeopardize the whole company, is feasible, and has an impact. Then, make a plan for where to get talent. A big risk is calling consulting companies. I used to be one of them! They offer a proof of concept in six weeks—the easy part—and then they leave. The client is left not knowing how to scale or manage what they have. They spend a lot of money when buying books for their teams might have been a better investment.

[00:25:40] Scott Kinka: I want to turn to ethics and government. You’ve worked extensively on governance frameworks. Where are we headed from a regulatory perspective?

[00:26:05] Dr. Eva-Marie Muller-Stuler: I am very worried. I recently realized I could give the same talk I gave in 2017, word for word, just by replacing a Google screenshot with one from ChatGPT. All the systems are so biased. Back then, if you searched Google for a “nurse,” it was always women serving men. I tried the same on ChatGPT—American nurse, Filipino nurse, European nurse—always women. When I asked for a “lawyer,” it was always men. We are training stereotypes into people’s brains.

Unbiased AI isn’t really possible because side demographics are never fully represented. It would be too expensive to train and monitor everything, so you essentially have to decide who you are okay with discriminating against. Our laws are also too vague. We have GDPR, but large language models are not GDPR compliant, and it’s not enforced enough. If you ask these models to build a case for why they are compliant with the EU AI Act or CCPA, they can’t do it. We have good laws that are currently meaningless and unenforceable. We are at a high risk of building a world that is far more unequal than before, especially regarding the Global South.

[00:29:26] Scott Kinka: That’s a scary thought. If laws aren’t followed or are inconsistent globally, it invalidates their usefulness. What are you advising governments to do to counteract this?

[00:30:37] Dr. Eva-Marie Muller-Stuler: I haven’t given up hope. People argue that regulations slow down innovation, but I disagree. If you have clear frameworks, your investment is safer. No one would fly on unregulated planes. We need clear, specific frameworks, often signed off by industry bodies. It has to stop being the “Wild West.”

[00:31:30] Scott Kinka: How should an IT leader think about this? Should they be cautious, or just assume the regulation will affect the AI companies and they should just use the tools responsibly?

[00:32:03] Dr. Eva-Marie Muller-Stuler: They don’t think about it enough. They should be stricter than the governments. If you don’t know what your model is doing, you’re flying blind. I was at an event recently with compliance officers from big tech companies. They were asked: “If you realized your company had thousands of AI models running that were never audited but are critical to the organization, what would you do?”

Most said they would “grandfather them in” because they didn’t want to interfere with the business. To me, that’s like an airline realizing their planes are running on unverified technology and saying, “Let’s hope it goes fine.” If a fraud engine or a credit application model hasn’t been audited, you might not discover the failure until five years later when everything collapses. We shouldn’t be afraid to say, “Stop everything. We will monitor these and start them back up one by one once we know what they are doing.”

[00:34:27] Scott Kinka: I could go on forever, but we’re running out of time. Let’s do a rapid-fire round. What is one development in AI that executives aren’t anticipating in the next 12 months?

[00:34:51] Dr. Eva-Marie Muller-Stuler: I think we will see a major failure. Executives will realize their bottleneck isn’t the number of proofs of concept they have, but that they are losing their competitive advantage by not having the right people in-house.

[00:35:23] Scott Kinka: Do you hold any unpopular or contrarian opinions about AI in the workplace?

[00:35:40] Dr. Eva-Marie Muller-Stuler: We need to focus on upskilling the people we have. I don’t see enough investment in that. I also miss seeing people with deep experience in AI teams. I set up the Hummingbird Group because I saw experienced people who didn’t want to deal with timesheets or corporate fights anymore. By letting them go, companies lose the mentors who show the younger generation what to do. The idea that computer scientists will be the first to lose their jobs worries me; if you don’t know what is right or wrong, your prompting will go in the wrong direction.

[00:36:55] Scott Kinka: What are you reading or listening to right now?

[00:37:07] Dr. Eva-Marie Muller-Stuler: I’m reading an awful lot of books on AI product development and responsible AI. It’s amazing how different they are, even with the same titles. I hope people will start reading end-to-end books again rather than just short articles. On a non-tech note, I recently reread Hemingway’s The Old Man and the Sea on a flight. It’s a great book.

[00:38:16] Scott Kinka: Last one: You’re in a dystopian future. Your phone only has one working app. Which one is it?

[00:38:36] Dr. Eva-Marie Muller-Stuler: A messaging app. I need to communicate with people, say “I’m here,” and find my friends.

[00:38:54] Scott Kinka: Perfect. Eva, this was a lot of fun. For listeners who want to learn more about you, the Hummingbird Group, or your governance initiatives, what’s the best way to find you?

[00:39:07] Dr. Eva-Marie Muller-Stuler: The best ways are the website, thehummingbirdgroup.ai, my personal site, doctoreva.ai, or LinkedIn.

[00:39:23] Scott Kinka: Awesome. This was an absolute pleasure. Thanks for taking the time from your temporary abode in London.

[00:39:29] Dr. Eva-Marie Muller-Stuler: Thank you so much, Scott.

[00:39:30] Scott Kinka: We wish you luck getting back home.

[00:39:33] Dr. Eva-Marie Muller-Stuler: Thank you. Let’s hope for peace soon.

[00:39:36] Scott Kinka: Amen to that. And thank you all for staying with us. That’s a wrap on this episode of the BridgeCast. If today’s conversation gave you something to think about, hit the subscribe button, leave a review, and share it with your team. We appreciate you spending your most valuable asset—your time—with us.

Special thanks to Bridgepoint Technologies for their sponsorship. If you’re thinking about digital transformation or an IT project, check out bridgepointtechnologies.com or reach out to me at skinka@bpt3.net. Until next time, do what you can to be the signal in all the noise.