Executive Summary
Theo Lau, founder of Unconventional Ventures, says the gap between AI hype and results sits in organizational readiness. Many institutions announce they are doing AI without a strategy or a defined outcome. Lau says institutions buying capabilities from several vendors may be placing them on top of broken processes. The result is automated mistakes and the wrong work done faster. She sees clear results in fraud and AML detection, small business underwriting, internal knowledge bases, and call resolution, and she says all of them are easy to measure. Personalized financial advice is still closer to a demo. Lau says an advisor would need a customer’s full financial picture across institutions, and she is wary of the ones she has seen.
Lau says bias and trust belong in the same conversation. Human judgment was never free of bias, so AI trained on untested decisions carries that bias forward. Synthetic voice and video make it harder to tell what is real, and she describes checking who is behind any unexpected email before she trusts it. Lau says institutions need to explain their decisions so AI is not a black box, and she cites the Apple Card bias question and a lawsuit against a student lender over its AI underwriting model as governance issues. She says responsibility goes beyond asking whether a decision can be defended. Drawing on work with her school board, she recommends two steps. Keep an inventory of the models and tools that touch decisions about customers or members, including those embedded in vendor products. Then decide early what should not be automated and which decisions people keep.
Lau says accountability cannot be delegated. An institution cannot blame the model or the AI agent, because financial services exist to serve people and businesses in a community. She worries about knowledge transfer. If automation removes entry-level work while experienced staff retire, institutions lose the way new employees learn and the knowledge that goes with it. On whether community institutions can compete, Lau says large banks treat AI as a scale game. Smaller institutions can move from decision to production faster, work with smaller and cleaner data sets, and stay closer to the people they serve. She says that trust is an advantage as long as a human stays in the loop. Her challenge to banking leaders is to leave the country and learn how other people work.
Listen, watch, or subscribe on your preferred platform:
YouTube | Spotify | Apple Podcasts
We are also available on all other major podcast platforms. Simply search for “Next Best Action.”
Key Takeaways
- Readiness Is the Gap Between AI Hype and Results: Lau says many institutions announce they are doing AI without a strategy or a defined outcome. Buying tools from several vendors does not fix processes that were already broken, and automating them produces mistakes faster.
- Fraud, Underwriting, and Knowledge Bases Work Today, and Personalized Advice Does Not Yet: Lau says fraud and AML detection, small business underwriting, internal knowledge bases, and call resolution produce results that are easy to measure. AI advice would need a customer’s full financial picture across institutions, and she is wary of the demos she has seen.
- Bias Existed Before AI, and Trust Depends on Explainable Decisions: Lau says human judgment was never free of bias, so AI trained on untested decisions carries it forward. Institutions need to explain how decisions are made, and she points to the Apple Card bias question and a lawsuit against a student lender over its AI underwriting model.
- Responsibility Starts With an Inventory and a Line for What Stays Human: Lau says responsibility goes beyond asking whether a decision can be defended. She recommends an inventory of the models and tools that touch decisions about customers or members, including those embedded in vendor products, and an early decision about what people continue to own.
- Accountability Cannot Be Delegated, and Automation Puts Knowledge Transfer at Risk: Lau says an institution cannot blame the model or the AI agent. She also worries that automating entry-level work while experienced staff retire removes the way new employees learn, and the institutional knowledge goes with it. She adds that heavy use of AI tools can lead teams to overlook security.
- Smaller Institutions Can Move Faster and Keep Closer Trust: Lau says large banks treat AI as a scale game. Smaller institutions can move from decision to production faster, work with smaller and cleaner data sets, and stay closer to the people they serve. She says that trust is an advantage as long as a human stays in the loop.
To learn more about Theo Lau’s work at Unconventional Ventures, visit https://www.unconventionalventures.com/.
Full Episode Transcript
Alex Jiménez: Welcome to Next Best Action. My guest today is Theo Lau, founder of Unconventional Ventures, an advisory practice focused on financial inclusion, longevity, and the intersection of financial services, technology, and humanity. Theo is a three-time author, most recently of Banking on Artificial Intelligence: Navigating the Realities of AI in Financial Services. She was named one of American Banker’s most influential women in fintech. She hosts her own podcast, One Vision, and writes a monthly column for FinTech Futures. Her work has appeared in the BBC, Forbes, MIT Technology Review, and Harvard Business Review. And if you make the rounds of the banking and fintech conferences, you will see Theo. Welcome, Theo.
Theo Lau: Thank you for having me, Alex.
Alex: Well, I’m glad you’re back on this side of the world after several months in Asia, and a side trip to Turkey. Glad to see you.
Theo: It’s a world tour. And I came back just for you. You are literally the first person to talk to.
Alex: Yeah, I know that. So today I want to talk to you about how to separate AI hype from reality in financial services, and why responsible AI and human oversight can’t be an afterthought. Are you up for it?
Theo: Yep. Love this topic. Who doesn’t?
Alex: A lot of people don’t. But we’re going to change their minds.
Theo: Okay.
Alex: So when you talk to bank and credit union leaders about AI, what’s the gap you see most often between the hype and what the technology can do today?
Theo: That is a big question, Alex. I would look at the gap from not a capability perspective, but more so readiness. There’s a difference. Everyone says, we want to do AI, we’re doing AI, but that’s not a strategy. It’s about how you can organizationally get it ready, and then implement it and look at outcomes. We’re not talking enough about outcomes. Everyone, the board, says, oh, we’re doing AI. And then next thing you know, it’s like, what exactly are you doing when you’re talking about AI? So looking at outcome, looking at capability, looking at exactly what you’re doing, that is important.
And it’s funny. Institutions that are buying AI are getting all these different AI capabilities from different vendors. But what they forget is that if your processes are broken underneath it, it doesn’t matter what you put on top of it. You are automating things that are broken. You are automating mistakes. You’re automating the wrong things. Perhaps you’re just doing the wrong thing faster. That is not a positive thing. And I think that’s the hype that we see today.
Alex: The other piece I see is the other end of the spectrum, where the only outcome people are thinking of is that AI is for fraud and underwriting.
Theo: It’s easy to do, though, isn’t it? It’s easy to measure, too. If you need to put a dollar sign on something you’re working on, we reduced fraud by X dollars and we’re able to underwrite an extra amount of loans. That’s an easy way to measure.
Alex: Yeah. And there are many outcomes that AI can help with. But you’re right, you’ve got to focus on the outcome, you’ve got to focus on the strategy, before you decide you’re going to put any kind of technology against it. So where have you seen AI create real efficiency gains inside financial institutions, versus where it’s still just a demo?
Theo: Can I say fraud now? We did talk about fraud. And honestly, fraud and AML detection, that is an easy target. That’s what we see a lot of companies being able to do, and being able to get good results from. Because AI fraud is much more prevalent now, you need something that can go at machine speed. So that is an easy one. We hear a lot about underwriting, helping small businesses get loans. That’s another good one, from document processing to intake to linking everything together, being able to get money in the hands of small business owners faster. That’s a good win-win on both sides. We hear a lot about knowledge bases. I think that’s also an easy one to implement, bringing all of your knowledge base information together so your internal staff can look up information faster, helping your CRM agents pull information faster, getting information to customers faster. Call resolution. Those are things we’ve seen work, and they’re easy to measure.
I want to talk about productivity. We hear a lot on the news, oh, we’re able to create code and push product out 40 percent faster. We can do things so much easier with less people. We keep seeing a lot of the headlines, especially in the last year, year and a half. You see Block. You see all these companies coming out and saying, we need less people. Part of me wants to say, yes, I can see how that works, because I’ve been using some of the tools to help me code faster. It’s funny, because I haven’t coded properly for a long time. So I see the benefits of those tools, bringing something together, at least from a prototype perspective. But the caution I have on that one, and I know we’ll touch on it later, is that when you start using a lot of these very useful tools to automate a lot of things, we’re also overlooking things that we should be paying attention to. For example, security. That’s a big one. Or helping the younger staff coming in learn how things work. When you automate the whole layer away, how are they going to learn how things work? And what’s going to happen in, I don’t know, five, ten years’ time?
Alex: Right. I’ve seen advisory firms using large language models to automate things like client newsletters and daily summaries of the market, which I think are great uses of the solution. The problem is if nobody reads it and checks what the large language model says about the market. You could be sending out something that makes no sense. So there’s got to be guardrails. There’s got to be some sort of intelligence around whether what we’re sending out is appropriate or not. And I worry about that, because the tools are great, but you have to check.
Theo: Wasn’t it recently that one of the big accounting firms, consulting firms, got into trouble because they were sending things out that they didn’t even bother checking? There might or might not have been some made-up stuff inside. And it wasn’t just that one firm. There was another firm, their counterpart, that did something similar for another government a couple of months ago. So you’re absolutely right. These tools are useful to help you gather information and do research, but you need to double-check and make sure that it is real.
Alex: There are many examples in the law field where attorneys have used large language models to do research that comes up with cases that don’t exist and are entirely made up. And it’s caught by the judge, which is beyond embarrassing.
Theo: Oh my goodness. You’re charging people for your time and your services. You’d better make sure that it works.
Alex: Yeah, exactly.
Theo: To the other part of your question, what is still a demo. What I would like to see more of, and I think that’s going to take a little bit longer, it’s not low-hanging fruit, is how we always thought, oh, now you have all these AI advisors that can help you make perfect personalized financial recommendations. So even people like me who can’t afford a proper advisor can get good advising. I don’t know if you’ve seen one that you would actually trust and that works. I’m still a little iffy about that. I’ve seen demos, I’ve seen things that look great. But to make a lot of these tools work, you need to give them proper context. Which means all of your financial information, not just a bank account, but everything else, all of the relationships that we have with different institutions, from tax to banks and what have you. Unless you have that, I don’t know if it can give you really good, concrete, useful recommendations. So I don’t know. It’s like a chicken and egg.
Alex: Yeah. At Finalytics, we’re going down that path. But the promise of open banking would allow for that to happen. If you have a bank or credit union that collates all the information from all your other accounts, and it’s available through your digital banking, in theory that’s where you could start seeing some of that all put together. But the client, the member, still has to approve, still has to connect all the accounts. All of that has to happen before something like this would work to give you the full, all-encompassing advisory that you’re asking for. Or you have a third-party tool, like, I don’t know, like a Rocket or something, that takes all these feeds from other places and puts it all together. But again, it’s not going to be something automated at this point. You and I have talked about and heard about the idea of agentic purchasing, agentic advisory, all of these things where all of this just goes to the background. And as a consumer, you don’t have to even think about it. Your agent’s thinking about it. It’s coming, but the infrastructure is not there yet.
Theo: There are so many things that need to happen before that. Yes, I do like the promised land, and I think we are closer to where it needs to be compared to ten years ago. But we’re not quite there.
Alex: The other problem is bias and disinformation. So what does that risk look like inside a bank or a credit union in practice today?
Theo: Can we acknowledge that humans have biases too?
Alex: Well, yes, of course. And when we do machine learning on data that is already biased, that’s where it comes from.
Theo: Right. It’s funny. I’m not saying that we should allow AI to be biased and give you wrong information, but we expect it to be perfect. If we think about the absence of that, we’ve been running it through human judgment and human discretion, and that was never perfect. And now you put the training and everything on top of decisions that were never tested and that were probably a little bit biased to begin with, and then we expect it to work. And I don’t know. It’s not magic. So that’s the first pushback I would have.
Disinformation is definitely a huge challenge. It’s a big fraud factor. I was just reading the news yesterday. It was saying that in Hong Kong, for example, investment fraud is becoming such a huge problem. And I think institutions everywhere need to figure out exactly what we can do. It’s synthetic voice, synthetic video, and we have seen how good some of these tools are. It’s getting really, really hard to figure out what’s real and what’s not real. And I don’t know about you, Alex, but you know me. When I see an email, when I see people reaching out, like, oh, we would like to chat with you on XYZ, the first thing I do is search on that person and see if the email address is from a company that seems legit. Is this the right person? Does it have a proper social profile? My first reaction is I don’t trust. And I don’t know, from a society, from an institution perspective, that’s a good thing. This is a whole industry and a whole society that’s based on trust. Human relationships are based on trust. When we are getting to the point where, when we see something, our first reaction is, I don’t trust it. You have to prove to me that I can trust you. I think eventually this is not going to be a good direction.
Alex: Yeah, and definitely not for banking. Banking is all based on trust.
Theo: Right. So I know originally we were joking about fraud being a use case that we often talk about when it comes to disinformation and stuff, but I think the challenge is real. It’s how do we reestablish the trust and how do we retain that trust? How do we explain the decisions that we make, so it’s not a black box? I know years ago we talked about Apple Card, and the people that they approve, and whether there is a bias in there or not. And I think last year there was also another case with a student lender that was being sued because of their AI underwriting model. And that’s a governance issue. But in the absence of regulations, how should I say this, we still need to make sure that we connect the dots and make sure that people are not harmed. So going back to looking at the decisions that are made, and the outcome, and the people that are impacted, not just because we’re told we need to do it.
Alex: Explainability is a big one. As an industry, we have a problem if our underwriting system is automated and we can’t explain why this person got approved and that other person who looks very much like them did not. Our regulations haven’t caught up with that yet. And when they do, we’re going to have an issue as an industry in the U.S. We already see some of that happening in other parts of the world. In Europe, they’re a little bit ahead of us in the governance and compliance around AI. Not where they should be, but they’re ahead of us. So when we catch up, it’s going to be a lot of pain for organizations and for the vendors that are providing these services. If there’s no explainability, “I don’t know, that’s what the system says” is not going to cut it.
Theo: No. Or, “I didn’t do it. The agent did it.” That’s also not going to cut it.
Alex: Yeah, exactly. There are all kinds of stories. There’s the story from a few years ago from the Netherlands, where the welfare agency, whatever it is called in the Netherlands, was turning down approval for many different citizens who happened to be from outside of the Netherlands, or who had moved, or whose parents had moved, to the Netherlands. It was a big deal, and the government took a strong stance against trying to explain what was going on. It was not a good situation. And it’s a good example of understanding the impact of the AI and machine learning that you’re putting in, and being able to explain it.
[Editor’s note: The agency was the Dutch Tax and Customs Administration, and the program was childcare benefits. Its fraud-risk algorithm used nationality as a risk factor, and thousands of families were wrongly accused of fraud and ordered to repay benefits.]
Theo: Absolutely. I think as a society we need to do more of that. Every action that we take as an institution or as a company has a consequence. And who are we impacting, and how are they being impacted by our actions? That needs to be something we think more about, regardless of what the regulations tell us we need to do or not.
Alex: So what does responsibility mean for a banker in the AI world? Is it just that the compliance team needs to come and review what I’m doing? Or is it… I’m kind of putting words in your mouth here.
Theo: It has to be way more. When you think about compliance, from a perspective, it’s more like, can we defend this? But being responsible is more than just can we defend it. It’s a little bit about going back to what you were saying before. The people that are impacted by the decisions that we make, are we okay with it? That’s what I would see as part of what responsibility means.
In practice, it’s funny you ask this question, because I went through a little bit of this with our school board recently. How do we look at implementing AI in our school system? What does that mean? How do we do it responsibly? The first thing we said we need to do is make sure that we have a good inventory of the tools that are being put in place that will touch how decisions are being made that would impact the end user. And I think this is applicable in banking as well. Do we know what models we’re using in the FI? Do we know what tools we’re using in the FI? Do we have a good inventory of that, that will impact how a decision is being made on the customer or member? That includes, for example, if you have bought in a credit card and you embed it in your vendor’s product, et cetera. We need to know what we have. I think that’s the first thing.
And then we also need to decide what we should not be automating. Just because you can doesn’t mean you should. We say that a lot. But in reality, right before the holidays, I had a really good discussion with someone who’s looking at HSA and FSA. And he said, at the end of the day, you need to know what part humans will continue to own. That is your responsibility. And then the rest you can probably automate, but you need to make sure that you know that line very, very clearly and early on. So I would say that is the second thing.
And it’s not just about fast. We talk a lot about efficiency. We talk about speed and cost and all of that. But we also need to, going back to what we were saying earlier on, think about trust and outcomes. It might be hard to measure sometimes, but that’s important. So that’s what it means for having responsible AI. That squishy thing is also important.
Alex: Models are getting better and better. It seems like every few weeks there’s a new, updated model that’s revolutionizing everything, but then it gets replaced by another one three weeks later. Why do you think keeping a human in the loop matters for that?
Theo: Because you can’t delegate accountability. You can’t blame the model. You can’t blame the AI agent. It’s still the human at the end of the day. This is a human business. If you think about it, why does a bank exist? Why does financial services exist? It exists to serve humans in a society, to serve the businesses in your neighborhood, in your community. So it’s a human business, and that’s why you need a human in the loop. It almost sounds silly to ask why we need humans, but we do think about that all the time.
It’s also about what we said earlier, knowledge transfer. We keep automating everything. I have two teenagers, and I think about what they’re going to do in a few years, after they graduate college. Will there even be any jobs for entry level, or for interns rather, because those are all the things that are going away, that are being automated away? And if that’s the case, and then you look at people that are older and they’re all retiring, you are losing that institutional knowledge, and you’re losing the handover to the next generation. What’s going to happen? I don’t need a crystal ball. I can already see this is not a good thing. So having a better model, getting released faster, doesn’t substitute for human judgment. I think that still needs to be there.
Alex: A lot of the bankers that we talk to are mostly community bankers or credit union executives. They talk about the difficulty of competing on AI. Is there really a gap? And is it really a scale game?
Theo: I love this question. I really do, because I genuinely hope, and think, that there is a place for smaller institutions to be in. I think it’s a tough one. If you look at the likes of JPM and Citi and all of those, when you talk to them about competing on AI, they look at it very much, as you say, as a scale game. Build models, hire a bunch of people, use a lot of resources. They have the data, they have all of that, and they just build and bulk up their AI capability inside. So in that sense, unfortunately, it is very much a scale game.
But if we look at the flip side, the small institution, chances are they can be able to deploy something and go from making a decision to putting it in production a little bit faster. I wouldn’t say very fast, but faster. So that’s an advantage. When you are smaller, you have a data set that’s smaller, cleaner, hopefully, and easier to bring together. So that’s also an advantage. And when you are in an environment where you are smaller, you are more in touch with the people that you serve. You have that trust. And trust is important, even and especially in the AI age, when you start automating everything, when everything becomes, I’m looking for the word, removed from human. But you still have that trust in there, and you still keep the human in the loop somehow. I think that’s also your advantage. So you’re not optimizing for everything, because you know that you don’t have to. You are going for a different competitive advantage, if you will.
Alex: Yeah, and I totally agree with that. Recently I was talking to someone who works with both small community banks and large regionals, and even one of the top four. What we were talking about was data. The data is not as clean as you think. I think a lot of people in smaller organizations think, oh, JP, they have a handle on the data, or B of A. And the reality is probably the reverse. They have more core systems. A lot of community banks and credit unions have a single core system. I remember working at B of A, and we had several core systems. How do you get all that data into a single place? There’s a data warehouse. Did you get all the data? Yes, maybe. Did you define it? Oh no, I didn’t. Now that’s a data lake. Do you have a data lake that has all that data? A lot of those things are not even a challenge at a place where all your data sits in the core system and you can get to it fairly easily in comparison.
The other piece is decision making, as you were saying. A small organization can make decisions so much faster. Having worked at B of A and at regional banks, sometimes it would take us over a year to make a decision that a community bank can make in a couple of months, if that. So there are some great benefits for small organizations with AI and any technology they want to deploy. But I would go back to your original comment of outcomes. Focus on the outcomes first.
Theo: It’s so interesting. I was talking to a digital banker last month, and she said, we need to focus on the outcome. And the funnier thing she told me is that some of these institutions, they’re way smaller, are actually doing a lot more than what you would see in the press, because they want to do it in stealth mode, if you will. They feel like they’re doing something special for their customers. It’s their own secret sauce, and they don’t want to come out and tell people. So when you’re smaller, you actually have the ability to do those things. And we need them.
Alex: Yeah, I agree with that. I want to switch a little bit and do a lightning round. I’ll ask you a couple of questions, and I just want quick answers, not necessarily related to AI or what we’re talking about. You ready?
Theo: Mm-hmm. Oh boy, you did not tell me about that, but okay.
Alex: No, no, no. So what’s your favorite city that is not Hong Kong?
Theo: Hmm. London.
Alex: London. Nice. What’s a challenge you’re currently excited about tackling, if you can tell us?
Theo: Yes. This will be the first time I’m going to say it out loud. I am very excited about writing my next book.
Alex: Awesome. I won’t ask you what it is about. I want to keep that surprise. What are you reading right now? I know that you read a lot.
Theo: I do read a lot. I don’t have enough time to read. That’s the problem. I just finished reading Paolo (Sironi)’s, a fiction novel, about “Quantum Sapiens”. It was an interesting read. I read it on the flight. My next one, I am reading Jensen Huang’s semi-bio, I think. Some guy decided to write about him, so it’s not an official bio per se, but he did get a lot of time with him and his circle. So it’s fascinating to read about someone… to read about why he wanted to start the company, his early days, the drive and the passion and his vision. To think about Nvidia where it is right now, it’s a fascinating story.
Alex: Yeah, I bet. If you were not in fintech, what would you be doing? I think I know what the answer is, but let’s see.
Theo: Oh. I don’t know, actually. I toyed around with it. I wanted to open a small coffee shop. A coffee shop and a bookstore together. So that would be one thing I would want to do if I’m not in fintech. And if I’m not doing a coffee shop, I would probably just travel. Can I get paid traveling?
Alex: You already do that, but you have to pay.
Theo: Yes, exactly.
Alex: I love the answer about the coffee shop, because that’s what my 11-year-old thought she wanted to do.
Theo: Alex, what are you trying to say?
Alex: I don’t know. All right. One final question. What is the one step forward you would challenge every banking leader listening to take after this episode?
Theo: Leave the country.
Alex: Ah!
Theo: And go out and learn how other people work. Seriously. I understand the U.S. is a huge country and there’s so much history, innovation, and diversity of the people that you serve, and all of that. And it’s amazing. It’s wonderful. But sometimes we forget that the U.S. is just one country in the entire world, and there are so many different countries, so many different communities, so many different ways of doing things. It takes leaving the country to appreciate what we have, but it also takes leaving the country to appreciate how other people work.
Alex: Yeah, absolutely. Love that. And on that, we’ll catch you another time. I know I will be seeing you soon in person, so we’ll save that for another time.
Theo: Yes, Alex, you cannot get rid of me. I’ll be around. Look forward to chatting next time.
Alex: All right. Thanks, Theo. Thanks for being on.
