- Key insight: Prashant Mehrotra has set up a disciplined operating model for AI deployments at U.S. Bank.
- What's at stake: Investors have recently rebelled against excessive AI spending at tech companies like Meta.
- Expert quote: "We have never been into token maxing. When you have some technology or some capability available to you, you first want to experiment with it, understand it. There is that excitement that happens. But in a corporate setting or even in a personal setting, I want to understand what I am getting out of this. What's the value?" —Prashant Mehrotra, chief AI officer at U.S. Bank.
As investors and analysts evolve from AI enthusiasm to return-on-AI scrutiny, artificial intelligence leaders within companies have to make decisions more carefully, knowing they're being watched closely.
Prashant Mehrotra has been chief AI officer at U.S. Bank for a little over a year; before that, he was head of AI at Allstate and director of data engineering at Capital One. He recently took our questions about how he is navigating shifts in AI thinking, how he maintains a balance between innovating and keeping a realistic reign on costs, coping with AI work slop and whether we're in an AI bubble.
What is a day in the life of a chief AI officer like?
PRASHANT MEHROTRA: The way I describe my job is I have the four "E"s: evangelizing, educating, enabling and executing. The reason I put that in that order is it's really important for folks to understand what AI can do for them. People want to know how they can use the AI that they use so well in their personal lives in a professional setting. Then comes the part about education: How do I use it? How do I build it? How do I integrate it? What's the right way of using it responsibly? Because once something is available in a corporate setting, you can use it on a lot of data, but is that the right policy? We have thousands upon thousands of builders, software engineers, analysts, developers who are building some of these things. So how do we enable them? How do we provide them the right platform? How do we provide them with the right tools? How do we upskill them?
Some of the largest banks have focused on using frontier models from OpenAI, Anthropic and the like and then recently we saw PNC and a couple of others talk about how they're more focused on proprietary models. Are you focused more on frontier models or proprietary models that you build in-house?
PRASHANT MEHROTRA: We've been using some proprietary models and open source models. We balance that out with models from OpenAI, Anthropic and Google. At the end of the day, we are going to have to be in a balanced environment where we are not stuck with one. We also want to protect some of our highly proprietary, confidential data that makes those models so much better. Mythos and Fable are really really good at looking for vulnerabilities. So it's finding the right balance.
There was a shift a couple of months ago from the idea of token maxing and rewarding people based on how much they use AI to reigning in expenses and being more concerned about what this is all going to cost. Some of that is because the major model providers went from allowing enterprise licenses to charging per use. How do you look at that? Have you ever been into token maxing? Do you encourage people to use AI as much as possible, or are you more cost conscious?
PRASHANT MEHROTRA: Let me start with the simplest answer. We have never been into token maxing. When you have some technology or some capability available to you, you first want to experiment with it, understand it. There is that excitement that happens. But in a corporate setting or even in a personal setting, I want to understand what I am getting out of this. What's the value? We want our folks to experiment, explore that next frontier. But we want to make sure it is in the service of all our stakeholders, from customers to shareholders to employees and communities.
The shift away from token maxing stems from three different factors. Number one is, of course, the model providers raise the price to match the value to what they are spending on providing that service. Number two is the experimentation phase is coming to an end if it's not already ended, and stakeholders are asking, why are we doing it? And number three, and this is really important, is I think the people who were experimenting are realizing the value of it, so they are starting to build that into the process. So, notwithstanding all the noise about these use cases not providing value, people are actually seeing value, so they are getting more serious about it. The technology has matured to a stage where it is actually delivering measurable impact, not just to the enterprises but also to the customers.
Can you give a few examples of how you measure that impact? Where are you seeing that measurable effect?
PRASHANT MEHROTRA: Absolutely, and that is something we as a bank are very proud of. I'll give you two examples: one customer focused and one employee focused.
In terms of customer focus, we provide payment services and we have merchants who use our Elavon technology. When our sales force would go out and talk to the merchants, it would take them weeks, based on what they were paying today, to give them a comparable quote. Through the use of AI, they can understand their needs, take their existing spend, digest that information, connect it with what we are able to offer today, and provide them a quote and proposal in near real time. We use generative AI, we use language models, we use vision models, we use predictive analytics to build a complete package that solves the customer need.
The second example is we are able to generate code faster. We are able to generate more code through AI, but more importantly, we're able to test more code, so we can generate higher quality code faster. These are just two of the many, many use cases that we have operationalized.
So you're looking a lot at time savings, and then do you try to translate that into specific numbers of hours saved, which translates to reduced salary expense?
PRASHANT MEHROTRA: In the first example, we look at: Are we able to convert more clients, onboard them faster? There is a real dollar impact to that.
When it comes to the second example of higher quality code and more code being generated through AI, we are able to see things like cycle time, production incidents, all of those developer productivity measures. There are certain tools we will provide employees to do their job better, where if they save some time, it makes them more productive, it makes them more effective, but we don't try to measure every single last piece of it.
Are you giving software developers GitHub Copilot or Cognition's Devin or something like that?
PRASHANT MEHROTRA: We started with GitHub Copilot and now we are providing them many more advanced tools, whether it's Cognition or Claude Code, and we're experimenting with even more, we are absolutely looking at the next newer generation of these capabilities because the capabilities themselves are improving further, and they are much more integrated in the tools they use. One of the guiding principles for us is making sure that we are collapsing the number of interfaces and not creating more and more screens or tools for them to use, so it needs to be integrated in the experience.
At American Banker's Digital Banking Conference last month, you mentioned the need for a disciplined operating model, and I wondered, what does that look like?
PRASHANT MEHROTRA: First, we are collecting ideas and use cases agnostically across the enterprise, but then we segregate them out into growth drivers, cost savings, and things like cost avoidance or improving productivity that cannot be measured, and we try to make sure we focus more on the first two buckets than the last one. Second, we started with a centralized organization, and we are federating this out more so that we can enable the rest of the enterprise. Third is, and this one's really important for us, the reusability of it. So, if a part of the organization develops a capability, we want to make sure it gets reused rather than reinvented in eight other parts of the organization.
How do you decide which AI use case ideas to move forward with? Do you put them through a filter?
PRASHANT MEHROTRA: We do have a filter. We have a governance body that looks at all the use cases through a common lens. But there are some things that we know will continue to make incremental progress, like improving the onboarding experience, providing a better experience to the customer, and those use cases get prioritized through a team that meets regularly and looks at that across the enterprise.
Read More:
Should a bank ever be liable when a customer gets scammed ?OpenAI models' Hugging Face breach is a red flag for bankers The top-performing banks with $2B to $10B of assets in 2025 Inside Upstart's journey to get a bank charter approval
We give them guidance like, find out where there is high frequency work. So if something is very complex, maybe it's a great use case, but it only happens once a month or once a day, it's probably not worth it. So is there the right level of volume? Do we have the baseline matrix around it? If you can't even measure if we are making an improvement, it becomes more difficult. So we are equipping the teams with the ideas with the right information, so they can prejudge a little bit of that themselves.
I also want to ask you about AI work slop because there are times where the AI model is very efficient, but then it takes a lot of time to fix the errors that it introduces, or to make the output truly useful. Do you have any thoughts on work slop and how to minimize it and deal with it when it does happen?
PRASHANT MEHROTRA: The way I think about it is when we put an engineered solution out there for our team members, we define not just the model we'll use, which is probably the easy part, but on what subset of data it can act upon, what's the intended outcome. We'll make sure that we are matching the right model, the right dataset, and the right expected outcome, in a use case. I'm not going to tell you that we'll never ever have AI work slop. We are looking at those things proactively. Another use case that I can talk about is we are looking at using AI in risk processes, and the purpose of that is to make sure that we are much more efficient around building and testing controls. But these are probabilistic technologies. So we need to make sure that we are testing them and building them and putting in enough guardrails so that we don't create a situation where fixing the output is going to take longer than doing it manually in the first place.
How do you take projects from pilot mode to enterprise-wide production?
PRASHANT MEHROTRA: That's where the disciplined approach comes in. There are two different phases. First is how do I know this is the right idea? Do we have the right skill set? Do we have the right capabilities and technologies in place? And you experiment, but you also timebox those experiments to make sure that you're not always running experiments. Secondly, the moment we select the use case, all the experiments in this one innovation, we do it with AI on what production would look like, and when we build those, we always have the team make sure that we can take them to production. There is a defined process to go from experimentation to production, and because of the rigor upfront on selecting the use cases and opportunities, we are able to see more and more of these things now scale up and go to production, but we are very disciplined on killing experiments if you know if they do not work as intended.
I've been seeing the term "AI bubble" quite a bit lately. Investors and analysts are asking companies what they will do if the AI investing bubble bursts. What do you think? Is there a bubble?
PRASHANT MEHROHTRA: I am not very good at stock picking. But I will tell you this: It's really important for us to separate our value from valuation, and I think sometimes we end up conflating those two things. There is a tremendous value in the capability itself. Are the valuations matching or do they run ahead? Euphoria, in my opinion, happens with every single breakthrough in technology. We have euphoria. We get ahead of ourselves. In the short term, we think this will change the world. There is a tremendous amount of value both to individuals and to corporations and society. Does the short-term valuation match that expectation, or what it takes to adopt it in a corporate setting; those two things need to get in line. Like any other technology, adopting it at scale and in a corporate setting takes some time and effort, and people underestimate that level of difficulty.
