Transcription:
Penny Crosman (00:03):
Welcome to the American Banker Podcast. I'm Penny Crosman. What separates companies that are truly successful at using AI from those that aren't? A lot of people are looking at this question from different angles. What use cases lead to the greatest ROI? How do you get people to use AI more? How do you get more accurate and useful results from the gen AI models you use? Our guest today, Bradley Lemer, has been a thought leader in banking and technology for decades. When I first met him, he was head of digital banking at Mechanics Bank in Walnut Creek, California. He went on to head innovation at Santander and then led FinTech partnerships and open innovation at Sumitomo Mitsui Banking Corporation. More recently, he's been in venture capital. He's currently an advisor at REI Capital and he's partnered with global venture firms like Insight Partners, Gause Ventures and Morrow Capital. He's also mentored and advised dozens of FinTech startups helping raise $44 million with three successful exits. Welcome, Bradley.
Bradley Leimer (01:03):
Well, thank you so much for having me on the podcast. It's good to see you.
Penny Crosman (01:06):
Sure. So a Stanford University study recently found that about 40% of professionals report having received AI generated slop in the last month. So this is content that looks good when you first look at it, but when you read more closely, you see that it lacks substance and that it may contain major errors. Do you see this as a genuine problem or do you think worries about AI slop are overblown?
Bradley Leimer (01:33):
Yeah, there's a lot to kind of unpack. When chat GT launched in November of 2022, it changed everything that the innovation team and SBC was working on. It seemed like every work stream that we did had AI embedded in every partnership, every project, and so it was important to really start off, I think, well, working on those type of projects, and when you think about what happens when you put something into a prompt and it's not what you expect, whether it's work slap as a hallucination or something else, then that starts to really take away from the credibility of what the technology can do. When you look about the projects that have been done across banks and the efficiencies that AI can deliver, the incredibly high expectations that AI could do, everything really starts to build that expectation. So let's define what works lot can mean.
(02:27):
You could call it inaccurate, fabricated, or low quality output generated by the systems that you're using. Again, whether it's sort of an everyday AI, like a copilot, or whether it's something like a specific LLM that was built by a bank inside or built by a partner, you're looking at something that people have a high expectation for that it's going to create the right result. It's a significant risk though We have a highly regulated environment and we accuracy in the financial sector of course, and as banks rely on AI for tasks like customer service and fraud detection and credit decisions or risk assessments of any kind, you need AI to be reliable. So it's not good. And yes, I've seen it, but honestly it's not very often because there's ways to mitigate it, which we'll talk about later. But let's think about hallucinations. So large language models confidently present false information and it could do it more and more because people seem to forget how it works or don't know how they work. Most models could only look at two or 300 pages of data at one time.
(03:42):
So the most advanced models from Anthropic and Deepseek can actually push that to about 2,400 pages of data at one time. So if you think about data in an AI model, accessing that information, think of it as looking at all of the data that's inside a library and all of the Dewey decimal system that it has to offer. It's going to kind of look at a very broad section of information. And having worked at libraries for more than seven years in Berkeley and Stanford and the places, the more you understand how to specifically target individual Dewey decimal numbers within that set of data. The more you look at specific books and specific pages within Met data, the better it is. And so we'll talk about how you do that and sort of reduce hallucinations. But workshop comes in other flavors too, algorithmic bias because if the data that is being trained on these models is already a little biased, it's going to simply amplify that. You have to include human in the loop to make sure that you don't have those types of bias. You could see things like data leaks where if your environment's not secure, some of that slop looks like exposing that data potentially to the public or to places that are outside of the bank. So this can lead to fraud, this can lead to poor quality output, but work slop horrible. But again, it can be reduced. It can be significantly mitigated.
Penny Crosman (05:14):
So what are those ways of mitigating it? And I'm interested in trying to apply the Dewey Decimal system to this idea.
Bradley Leimer (05:23):
So first off, you have to have a really robust governances framework at SNBC and other places that really have a clear program. You establish clear policies, procedures, and controls for both AI use. So what applications are you going to use? What type of environment are you going to enable AI tools to come into? What type of data controls do you have? How clear is the data usage for your teams? So that's kind of the first you have to have that around. Then you think about things like retrieval, augmented generation rag. So this is the whole idea of taking the library of information and narrowing it down. The technique grounds AI answers into specific verified internal knowledge bases. So it pinpoints data and trains the model on that data so that you do get more specific and more clear answers that are generated in a way that is such as a human would because you're basically using them as a expert to find that data.
(06:24):
So think of it, you're sending out librarians into the library to find things that are specifically trained for them to look at, and that significantly reduces hallucinations. Another thing that a lot of people building models are doing now is that they're creating specialized or small language models. So for things that are very important, building out presentation deck that relate to client deals or to do tasks like credit decisioning, banks can train smaller, more specialized models to create carefully vetted domain specific data-based responses from an AI workflow that is specifically designed for that. So what's interesting is that it actually makes the model itself smaller so that it has to process smaller amounts of data, but it's more specific and more targeted. A couple other things really quick, it's enforcing the fact that you have humans here, so humans in the loop you'd hear about with ai, but that's critical because you need to really leverage a final view on things to have loan officers or risk analysts or compliance folks look at the output of what happens with AI generated information.
(07:39):
So everything that you do is designed to lock down the data, lock down insecure environment, prioritize things like transparency. AI systems should be explainable. People should understand what they're doing. They should have a documented auditor for how decisions were made, and that's what the regulators are looking for. That's what your compliance teams are there for. That's why they're brought into the big tent of developing them. And then the last thing to sort of mitigate all of these issues with work slot train employees, raise awareness, educate employees and teach them how to do prompting, teach them about the limits of AI and how to spot things like AI generated fraud or anything that has to do with something that might be considered work slap.
Penny Crosman (08:28):
Yeah. We recently wrote about Citi, which has mandated prompt training for 175,000 employees, which is most employees. Do you think we're going to, and this was training them in AI prompting, do you think we'll see more banks and other companies do the same?
Bradley Leimer (08:49):
Yeah, I mean, again, I would just say that your reporting on this topic has been great. I think it's really important that banks understand what other institutions are doing, not just from a competitive standpoint, but from a let's make banking better and let's ai let AI be part of that solution. I think cities move a significant, they publicly said that this is what they're doing. All 175,000 employees are required to complete prompting like a pro training within 60 days. That might seem like a big thing to people outside the industry, but we have report or training that happens all the time inside financial institutions. And when you think about the kind of compliance training and regulatory training that we do, something like this doesn't seem like that big of a deal, but look at the results. I mean, according to your reporting, six and a half million prompts have already been submitted this year.
(09:44):
They're doing these workspaces and assist agents to do an agentic approach to some of the workflows that they have. Do I think more companies and more banks are going to require training? Absolutely. Think of it as Excel or PowerPoint when they came out decades ago or other types of applications that are pretty much table stakes now. That's what AI is going to be. The challenge I think, is that their training seems pretty basic. I think that when they can do it in 10 minutes or they could do it in 30 minutes, it's not enough. It needs to be more comprehensive. You need to create a center of excellence. So this is educational tools, resources, comprehensive training programs, and you need to create a community, a community of influence to offer assistance really to say, this is a prompt that I'm working on, or this is something that I'm trying to work on with my AI tool. How can I do this better? And you need to have a lot of resources. AI doesn't reduce resources. If anything, it actually adds them. So the creative proper environment to have space to learn this stuff that's important.
Penny Crosman (10:54):
What are some of the things people should be trained on besides prompt training? Like you mentioned, having employees review output. Does that take specific training you think? Especially if there's high volumes of output coming out of a gen AI model?
Bradley Leimer (11:11):
Yeah, I think a good AI program has to be robust and not just part of individual people's job. It needs to be comprehensive and ongoing city's approach is upskilling as opposed to strong arming. I think everybody who is involved in the beginning set of the data being available, the beginning of the tools being available, they need to be involved in creating this training. So that just like when we do regulatory training, that's mandatory. We have the experts that are involved in creating the training itself, the folks that look at the results at the end when a loan decision is made or when a deal gets presented to a client or whatever it might be across capital markets or consumer workflows. Everybody that's been involved in that as a human decision maker needs to be involved in not just developing the model but validating the model. And that takes work. And I think that's part of the thing that banks are learning is that, sure, it seems like AI is going to make it magic that we could do more with less, but it takes a lot of work to get that done well, and I think it's something that is bigger than the educational breadbox of doing one class one time.
Penny Crosman (12:29):
That makes sense. So we also wrote this week about the very large headcount reductions and Accenture and a TCS where for instance, Accenture has had a headcount reduction of 11,000 people in the past three months. And the CEO said at the most recent earnings call, we are exiting on a compressed timeline. People were re-skilling based on our experience is not a viable path for the skills we need. We're investing in upskilling our reinventors, which is the primary strategy. Those we cannot re-skill will be exited, which to me sounded a little chilling. What do you think this idea of exiting people who are considered AI laggards?
Bradley Leimer (13:12):
Yeah, I saw this and I've seen a lot of inroads by companies saying, Hey, we're going to completely outsource this to ai, or we're going to have reductions because we're so much more efficient now. Accenture maybe just has too many employees in general, and that's sort of a leadership failure. That's a different thing. How can you cut 11,000 people and say some people are beyond re-skilling because I'd like to see the people that refused to take an AI class. This erodes culture, if anything, and I see it again as a leadership failure in planning for the evolution of the workforce. Retraining is generally faster and cheaper than mass amounts of layoffs and turnover. It's crazy to think that that's the answer. It's not. Most employees can learn basic AI skills if they're truly invested in it and the company's invested in giving them the time and space. The danger isn't just the layoffs, it's the message that adaptability is optional until it's suddenly not. It's a convenient excuse that we're seeing more and it's not a good one. So banks don't do this. Invest more in your employees, invest more in your bank's future. And AI could be a part of that.
Penny Crosman (14:25):
And I've had conversations with people where they brought up ageism. Is this an excuse for getting rid of older people and then using the excuse that they can't be re-skilled? Not even sure how you would really conclusively identify people who can or cannot be re-skilled unless, as you said, there are people who refuse to take any training, then I guess you'd have some kind of basis.
Bradley Leimer (14:50):
Well, this idea that people that are older can't do tech. I've been in this space for 30 years and I could swing around any sort of tech that those 20 year olds can. I think it's really about people and culture. Most people can be reskilled, especially on basic AI tools, and that's very targeted to what they do. What's harder to instill is that mindset, curiosity, adaptability, willingness to experiment, and the company giving them time to do that. The real question isn't who can't learn AI or who refuses to evolve has nothing to do with age leaders need to ask if it's an individual gap or a management one. Here's an idea. Have your executive leadership team demonstrate their usage of new AI tools, their commitment to be trained in how to use AI and other emerging tech. Because I've been seeing it for decades where I don't understand why my folks that are in executive leadership roles, because I've been in one for decades, they don't look at this technology the same way. So let's talk about all those things before we could identify people who can't be re-skilled. So just with the rise of FinTech and the rise with other emerging tech over the last 15, 20 years or more, you really have to do more than just wait it out and retire. These folks need to be willing to adapt and learn as well. To really run a bank well, you have to be all in, and that means everybody at the top too.
Penny Crosman (16:23):
So it sounds like you've been doing some AI engagements in financial services or helping firms with their AI deployments. What are some of the things that you have seen in terms of best practices or best use cases where you're seeing really solid results and strong ROIs?
Bradley Leimer (16:44):
So I'd always suggest, especially to banks that haven't really dipped their toe much into it, is to start narrow where the risk of AI work, slop and things like that, the things that we've been talking about are really low, and yet product productivity gains are more measurable. So pilots around the call center and customer assistance, banker assist tools to help them have the right information as they have conversations with clients, look at document heavy tasks that are related to onboarding or underwriting deal flow or contract review and those type of things, and give your developers tools to offer things, whether it's copilot and GitHub with code suggestions or debugging or patch automation. The best sort of playbook that we look at with AI and that in practice the last three or four years I've been working on is you pilot, you measure the results of that pilot and then you iterate and you train models and you get more people involved.
(17:44):
Banks like Citi and Bank of America and JPM of course are layering governance feedback loops so that AI doesn't run unchecked. You have to learn from what they're doing. And in my last couple of years, providing all employees sort of everyday, AI was projected to save the bank nearly a quarter of a billion dollars in efficiencies globally. That's not a small number, and that's just a general AI tool with copilot. If you look at all of the saving potential with workflows with specific LLMs that are created for those roles, that's a much bigger number than that. So if banks can just work a little bit harder to develop the right culture, to avoid the type of AI related layoffs that we've been discussing, give teams the support, the time that they need to leverage these tools properly, AI is surely going to prove out to be the most impactful technology that has ever been implemented inside a bank.
Penny Crosman (18:43):
So the next wave of course is agentic AI and having agents that are working with generative AI models, executing tasks automatically, interacting with other agents for things like executing a payment if a certain condition is met. Do you think that, do you see that as the next wave of productivity and are there ways of making sure that the danger of an agent not running amok is mitigated as well?
Bradley Leimer (19:19):
Yeah, I do. And panel, like what we talked about earlier, you have to get more teams involved that are part of that process today. So agentic, AI really stitches together a lot of the things that we were talking about, this idea of using rag or small language models to specialize what that workflow looks like. One, you talk about building and stitching together multiple calls of data, multiple iterations of generative function and data results into something that is agentic. What you're talking about is building out, in some cases, very complicated workflows that today humans have the touchpoint in between. You can't automate all the banking, you can't automate what humans can do to the certain extent that you can guarantee that the results are going to be the same. But just like humans, machines and humans can both make mistakes and that's why you have checks and balances. I don't think that should change. I think that there's still going to be people in the back office that verify, validate, look at what the results are of just about everything, and that's what the regulators are going to require us to do as well. That's why people have trust in this system. And so I don't think AI is going to change that unless companies really do a poor job of deploying these things.
Penny Crosman (20:39):
Alright, well, Bradley Leimer, thanks so much for joining us today and to all of you, thank you for listening to the American Banker Podcast. I produced this episode with audio production by Anna Mints. Special thanks this week to Bradley Leimer at Darrery Capital. Rate us. Review us and subscribe to our content at www.americanbanker.com/subscribe. For American Banker, I'm Penny Crossman and thanks for listening.