# Managing Change Across Modern Financial Operations - with Ajay Swamy of JPMorganChase and Founder of FundLens.ai

- **Show:** The AI in Business Podcast
- **Published:** 13 Aug 2026
- **Source:** https://podcast.emerj.com/managing-change-across-modern-financial-operations-with-ajay-swamy-of-jpmorganchase-and-founder-of-fundlensai
- **Transcript:** https://podcasts.opennash.com/shows/ai-in-business/managing-change-across-modern-financial-operations-with-ajay-swamy-of-jpmorganch-e186dd/
- **Provider:** OpenNash Podcasts (machine-generated)

## In brief

Ajay Swamy argues that the real obstacle in banking operations is not speed or scale but trust. Banks already process millions of transactions an hour. What slows them down is fragmented data and judgment calls hidden inside routine steps. His answer is to build AI that is explainable, traceable and auditable from the start, so every decision can be defended.

The main tension is between the excitement around new AI models and agents and the less glamorous work of governance. Swamy says the models themselves are becoming interchangeable. What separates institutions is whether they can supervise AI, catch problems quickly and rebuild workflows around controls. His closing message is that trust must be engineered deliberately, not assumed.

## Key insights

1. **Fragmented data, not volume, slows bank operations** Swamy says banks process millions of transactions an hour flawlessly. The friction comes from data spread across trading systems, document stores, risk engines, SharePoint and personal spreadsheets, each with different entitlements.
2. **Pointing AI at all your data backfires** Swamy warns that you cannot simply 'rag your way through' a bank's data, because it is a dozen different shapes with different criticality. Dumping it into one system produces nonsense and exposes the bank to regulatory, reputational and financial risk.
3. **Untraceable decisions are indefensible decisions** Swamy's rule is that a number or decision you cannot trace is one you cannot defend. A loan rejection, for example, must show which data points went in and whether it matches past decisions, or the bank has a regulatory problem.
4. **Evaluate AI on supervisability, not demo performance** Swamy says frontier models are being commoditized month by month, so the technology is the easy part. The real test is whether the bank can supervise it and audit it end to end, from training data to outcome.
5. **Risk controls should scale with materiality** Swamy says a model used in trading carries far more risk than one extracting invoice data or processing checks. Oversight should match that, and aim to reduce and quickly act on errors rather than demand zero errors.

## Questions this episode answers

### Why is it hard to run banking operations consistently if banks already handle huge volumes?

Ajay Swamy says scale is not the problem, since banks process millions of transactions an hour. The difficulty is that data is scattered across many systems and even routine steps need human judgment.

### Why can't a bank just point an AI tool at all its data?

Swamy says the data is not one consistent pool. It sits in trading systems, document stores, risk engines and spreadsheets with different access rules, so combining it can produce nonsense and create regulatory, reputational and financial risk.

### What does 'explainable' AI mean, and why do regulators care?

It means being able to show which data went into a decision, how it was handled and why it came out as it did. Swamy uses loan rejections as an example: if a bank can't explain one, it has a regulatory problem.

### How should a bank judge whether new AI technology is safe to adopt?

Swamy says to look past demos and ask whether the bank can supervise and audit the system from the data it takes in to the outcome. If the software can't show that, his answer is no.

### What does Swamy mean by 'engineer trust'?

He means trust does not come built in, so a bank has to design it into its processes through grounded data, controls and oversight. He says the goal is making operations defensible at speed, not just faster.

## Facts

- **Ajay Swamy** [0:02:32]: "I've seen banks process millions of transactions every single hour flawlessly."
- **Ajay Swamy** [0:02:32]: "There's one workflow that touches a core system or a trading system."
- **Ajay Swamy** [0:04:27]: "There's a new policy called the S R twenty six dash two for Gen A I, where people have just said, hey, I don't know, you tell me, I e the banks, you tell me how you want to govern Gen A I."
- **Ajay Swamy** [0:06:31]: "you also need some sort of a golden corpus of eval documents where you can evaluate and say, hey, this is actually giving me to a state of confidence, accuracy, confidence that I'm comfortable with."
- **Ajay Swamy** [0:06:59]: "you also need to plug in some level of a human in the loop so they can double check, sort of like the four eyes principle."
- **Ajay Swamy** [0:06:59]: "a bank rejects somebody for a loan. Now that decision has to be explainable, and that's that's a regulatory matter of fact principle."
- **Ajay Swamy** [0:08:56]: "the models that they're using underneath, the frontier models, et cetera, they're all getting commoditized on a month by month basis."
- **Ajay Swamy** [0:09:27]: "any model, whether it's machine learning or GEN EI, that sits in a business critical process, it has to be auditable end to end."
- **Ajay Swamy** [0:10:01]: "So if your software can't show that, then it's a no."
- **Ajay Swamy** [0:10:37]: "For my startup FundLens.ai, what we do is that we learn how a fund operates based on its philosophy and its strategy and its risk posture."
- **Ajay Swamy** [0:11:11]: "So it can surface when a new input cuts against what we already know, or what we already are comfortable with"
- **Ajay Swamy** [0:11:49]: "a model that's being used in trading is going to be highly, highly more riskier than a model that's being used to like. Maybe extract invoice information"
- **Ajay Swamy** [0:13:16]: "There have been a couple of examples across my career where things have gone south, where amounts have been missed."
- **Ajay Swamy** [0:14:46]: "There's going to be this notion of tokenomics, where the amount of tokens you're using and how valuable it is, right?"
- **Ajay Swamy** [0:14:46]: "the institutions that will win over the next couple of years are the ones that have rebuilt their workflows and their governance controls"
- **Ajay Swamy** [0:15:23]: "multi model orchestration with new models, agentic, fully agentic, and autonomous agents, M C P servers, skills that you use in Claude Code or Codex"
- **Ajay Swamy** [0:16:20]: "There's product managers on my team who don't write from scratch anymore."
- **Ajay Swamy** [0:16:54]: "Because that'll do seventy to 80% of the heavy work."
- **Ajay Swamy** [0:16:20]: "you will be able to deliver faster with more quality, as long as you sort of ground the A I with the specifics manifest files and the skills"
- **Ajay Swamy** [0:18:54]: "I think you need to make them defensible at speed. So the bottleneck was never really the speed."
- **Ajay Swamy** [0:19:21]: "Don't try and automate trust, engineer it deliberately into your processes, make sure that's grounded"

## Transcript

**[0:00:12] Host:** Welcome everyone to the Emerge AI in Business podcast. Today's guest is Ajay Swamy, Senior Executive Product Director for Gen AI products and governance at J P Morgan Chase. Ajay discusses why operational processes in banking look simple step by step, but require constant judgment once you account for fragmented data across trading, document and risk systems. He explains what separates AI that's safe to deploy in regulated environments from AI that just performs well in a demo. Today's episode is sponsored by Reindeer.

**[0:00:44] Host:** A quick note for our audience that the views and opinions expressed by Ajay on today's program are his own and do not reflect those of J P Morgan Chase or its leadership. In this episode, we cover why fragmented data and manual judgment calls are slowing down financial operations and what it takes to make AI driven decisions. Explainable. Deeper on this topic and learn how financial institutions are digitizing paper-based records to unlock usable data for AI. Download our free PDF report, AI in Financial Services Executive Cheat Sheet at emerj.com/fcs1.

**[0:01:17] Host:** That's E M E R J dot com slash F C S one to download your copy. Now, the conversation with Ajay. A J Welcome to the Emerge A I in Business podcast studio.

**[0:01:54] Ajay Swamy:** Thank you very much. Happy to be here.

**[0:01:55] Host:** Yeah, I'm excited to get into this conversation. Uh with you. I think we're gonna hit some interesting things today. At this point, I think everyone agrees that uh individual pieces of a workflow, uh a document check, a policy rule, uh sign off. All of these things are very simple on their own, but if we string ten of those together across different systems and different teams. And somehow, that entire simple system turns into a whole new thing.

**[0:02:20] Host:** That turns into weeks of back and forth and just a lot of maintenance around it. And you're the right person to ask. What makes operational workflows in B F S I hard to run consistently at scale?

**[0:02:32] Ajay Swamy:** Yeah, look, I think that's a great question, but I'm gonna challenge the premise, right? I think the hard part isn't necessarily the scale, because quite frankly, I've seen banks process millions of transactions every single hour flawlessly. So the hard part is that every one of these workflows or these steps requires judgment, even when it looks routine, right? So what do I mean by that? So first is fragmentation, data fragmentation. There's one workflow that touches a core system or a trading system.

**[0:03:04] Ajay Swamy:** There is one workflow that touches a document store. There might be another workflow that touches a risk engine, and there might be another workflow that touches your beloved SharePoint sites. And then you have other cottage industries across big financial institutions where people have their own Excel spreadsheets. So different data, different data stores, different entitlements. So when people just say, hey, point A I at it, I think that's a big mistake. You cannot just rag your way through this because the data isn't one corpus, it's not consistent.

**[0:03:42] Ajay Swamy:** It's a dozen different shapes, a dozen different varieties of criticality, material, non-material information. So you can't just dump it all into one system, because ultimately you'll get nonsense. But more importantly, you're going to expose yourself to a lot of risks, both regulatory, reputational and financial. I want to sort of like paint that picture there as to like why it is so difficult for operational workflows to really work well and consistently.

**[0:04:11] Host:** Is there a part that adds to the friction and the frustration that is still done manually? And that also creates a new type of fragmentation and frustration? Is there a part that I think the outside? Outside of the industry, people would be surprised to hear that, oh, this is still being done manually.

**[0:04:27] Ajay Swamy:** I think what happens is that let's take an example. There's different document variabilities, the input is unstructured, nothing arrives in the same format twice, and that requires some level of manual input. Now, the third piece of it is also sort of policy and regulatory interpretation. These things change. You know, this A I rules and policies being put into place today. There's a new policy called the S R twenty six dash two for Gen A I, where people have just said, hey, I don't know, you tell me, I e the banks, you tell me how you want to govern Gen A I.

**[0:05:05] Ajay Swamy:** Now, when you have such generic hand wavy regulations, you need to have some level of manual control, and I think that's a necessary twin in that. However, if you root yourself with some ground rules, I e explainability, traceability, lineage, and you know where the data is coming, you know how the data is being modified, you know how the data is being input into the A I, and you know exactly how that A I makes a decision.

**[0:05:35] Ajay Swamy:** Then you really have yourself some sort of a governed, traceable, explainable process that will remove some of the manual elements out of it. Otherwise, you're going to have manual human input. That's gonna slow down our workflow, it's gonna slow down our process, and by the way, humans make mistakes all the time. And it's very hard to stitch together what exactly happened across that process.

**[0:05:59] Host:** Uh, the picture that you're painting makes me wonder about. Obviously, if we have the data is fragmented already, and if we have humans involved there as well, like you said, uh, the same document will come in different formats. And if we then throw A I in the mix, the system might not always get what it expects, depending on which part of the process we're in. So when a document doesn't match what the system expects, what happens next?

**[0:06:24] Host:** Does it stop the entire process, or does someone quietly patch it by hand? What is the the background of that look like?

**[0:06:31] Ajay Swamy:** I think, sort of technically speaking, it's that, you know, it's the same document, maybe twenty different formats, twenty different counterparties. You need a combination of, like, vision models, text extraction, table parsing capabilities, you need all of those working together. And by the way, you also need some sort of a golden corpus of eval documents where you can evaluate and say, hey, this is actually giving me to a state of confidence, accuracy, confidence that I'm comfortable with.

**[0:06:59] Ajay Swamy:** And you also need to plug in some level of a human in the loop so they can double check, sort of like the four eyes principle. Hey, this is actually accurate. Like, I'll give you a perfect example, right? This is why I sort of harp very much around, like, explainability, traceability and lineages. Because, like, let's take an example that everyone understands, like a bank rejects somebody for a loan. Now that decision has to be explainable, and that's that's a regulatory matter of fact principle.

**[0:07:27] Ajay Swamy:** Which data points went in, why, and whether it was consistent with the last thousands of decisions that the bank has made. If you can't answer that, you don't really have a model or a parsing or a logic problem, you have a regulatory problem. So I would approach it from that lens.

**[0:07:44] Host:** And that's the perfect lead into my next question. Because we're thinking about evaluating the systems and the processes and the A I and everything that we're incorporating. And and what's really been nagging at me is that everyone we speak to, they're all saying that, oh, no, you know, um, we are being thoughtful about new technology, but thoughtful can mean two very different things, right? It can either be a real evaluation process or just caution dressed up as diligence.

**[0:08:12] Host:** And I think once, uh once, a technology in question can act on its own instead of just answering questions, that's when things are really getting tricky. So, and and you've been on both sides of having to make these decisions. How do you evaluate new technologies for use in regulated operational environments?

**[0:08:31] Ajay Swamy:** I think a line that I use pretty frequently is a number or a decision that you cannot trace is a number or a decision that you cannot defend. So, no black boxes. So let me tell you, let me give you my honest version, right? It's it's a little different than what gets said in public. In a regulated environment like a bank, or a buy side, or a P shop, or insurance, you're not really evaluating the technology, the technology is the easy part.

**[0:08:56] Ajay Swamy:** Because you see all these great demos, you see all these great feature sets, and by the way, the models that they're using underneath, the frontier models, et cetera, they're all getting commoditized on a month by month basis. Like, you look at Claude's Fable and the next version of G P T. Ultimately, it's gonna get to a level of intelligence where they're sort of similar. But what you're evaluating is, hey, can you actually supervise that technology? And can you actually understand the inner workings of it?

**[0:09:27] Ajay Swamy:** And that's the floor. Because any model, whether it's machine learning or GEN EI, that sits in a business critical process, it has to be auditable end to end. From the data that it takes in, the data that it manipulates, to how it was trained, to how the outcome occurred, and to the moment there's divergence in those outcomes, or those, or those results. You have to figure out, you have to be able to tell, and you have to be able to explain, not only to your risk committee, but also to your customers and to the regulators.

**[0:10:01] Ajay Swamy:** So if your software can't show that, then it's a no. So I would urge to look at it from that perspective, because quite frankly, there's two axes to this. The first axis is like, hey, does it do the job that I wanted to do from a capabilities perspective? But the second aspect of it is, is it safe? And can I use it in an explainable fashion? So that's how I sort of approach it.

**[0:10:22] Host:** And I guess that's something that we'll evaluate, maybe for some organizations on a monthly basis, some quarterly, depending on their compliance requirements. But is there something that we should be overseeing on a day to day basis? Uh, once this is fully in production,

**[0:10:37] Ajay Swamy:** Yeah, absolutely. Look, I think I think I think you need to have some level of oversight. So, for example, when I talk about model divergence, or we talk about changes in data, or we talk about decisions that analysts or traders make based on that software, you need to have some level of oversight. That is also learning at the same time. So I'll give you a perfect example. For my startup FundLens.ai, what we do is that we learn how a fund operates based on its philosophy and its strategy and its risk posture.

**[0:11:11] Ajay Swamy:** So it can surface when a new input cuts against what we already know, or what we already are comfortable with from a level of risk and from a level of awareness. From a level where we think this is sort of the antithesis to what it should be doing. So the point isn't to like police the analyst or the trader. Point is to recognize when and where there's a divergence and be able to say, hey, wait a second, let's pause, or let's raise this risk immediately, or let's call that human in the loop that says, let's figure something out before something worse happens, right?

**[0:11:49] Ajay Swamy:** That being able to have that risk posture and being able to understand that from like a model risk management perspective, that's critical. Like, you have to know how a model is used, how material the risk is, and that materiality climbs across different business functions. So, for example, a model that's being used in trading is going to be highly, highly more riskier than a model that's being used to like. Maybe extract invoice information or read checks, for example, or process checks.

**[0:12:19] Host:** I can hear that. Obviously, these are not just things that you learn from reading a book. Um, I'm sure that there is some lived experience there. So let's. I'm gonna help our listeners to maybe learn from your life lessons. Ha, uh, has a system that you approved ever done something that surprised you after it went live, or something that you did not expect it to do? And that probably wasn't that great?

**[0:12:40] Ajay Swamy:** Yeah, look, that happens all the time. I don't want to name names, actually, but you know what? What I would say is there have been divergences of systems and software across my career, and that's just part of the trade. I mean, that's what software is. Because you, as a human, when you build and write that software, you're not necessarily accounting for every single edge case, you're not accounting for every single exception. So it's hard, and especially in today's world where data comes at you so furiously and you're taking in all of these different inputs.

**[0:13:16] Ajay Swamy:** It's very hard for you to say, hey, I know this is going to work within this certain level of accuracy all of the time. So, to answer your question, I have seen it. There have been a couple of examples across my career where things have gone south, where amounts have been missed. There's been rules that have been triggered where they shouldn't have been. But ultimately, what mattered at the end was being able to have controls around it that allows you to sort of mitigate that.

**[0:13:44] Ajay Swamy:** So that's how I would, I would think about it, right? Don't think about it from the perspective of, hey, I wanna have sort of zero tolerance for errors. Look at it from the perspective of how can I control and reduce that risk and be able to act on it quickly. Because ultimately, there's no such thing as sort of a perfect product.

**[0:14:01] Host:** That sounds very realistic and and very honest. I think something that everyone's busy talking about, and I think we we also basically did it a bit earlier in this conversation. Is is speaking about the next generation of AI agents and what they'll be capable of and all of that, but not really talking about. Is the institution ready for it? Ready to receive it? In terms of the skills and the controls and the muscle memory for managing something that acts on its own instead of just using answers that's given to it.

**[0:14:30] Host:** And you're kind of in the middle of of building that readiness at a place with probably, I don't know, as much at stake as as possible. So what do you think a B F S I institutions really need to have in place to adapt to what's coming next?

**[0:14:46] Ajay Swamy:** I think the next wave, it's not going to be sort of the best models. I think, like I said before, the models are going to get commoditized to a certain degree. There's going to be this notion of tokenomics, where the amount of tokens you're using and how valuable it is, right? Like, let's call it like token R O I, that's going to become key. We're going to start looking at some of those things. The other part of it is the institutions that will win over the next couple of years are the ones that have rebuilt their workflows and their governance controls to absorb all of these new capabilities.

**[0:15:23] Ajay Swamy:** And when I say capabilities, I mean things like multi model orchestration with new models, agentic, fully agentic, and autonomous agents, M C P servers, skills that you use in Claude Code or Codex, or whatever. Being able to understand how that impacts either your data or your people or your reputation, or the financial risk. That piece of that is going to be just as important as being able to embrace and absorb and adopt A I across your institution.

**[0:15:57] Ajay Swamy:** Obviously, it doesn't sound as glamorous when you talk about governance and controls, but ultimately, that's what allows you to sort of manage some of that risk.

**[0:16:05] Host:** A lot of things to think about there. And before we wrap it up, we were just thinking about the future. But I'm gonna ask you about the past as well. Um, what is the skill that your team didn't need three years ago that is absolutely non negotiable now?

**[0:16:20] Ajay Swamy:** I, there's multiple ways to answer that. So one of the things I'll say is definitely coding assistance, copilots. These are absolutely critical things because, quite frankly, you will be able to deliver faster with more quality, as long as you sort of ground the A I with the specifics manifest files and the skills and all that stuff. The second piece of it is the things that you do manually. I'll give you a very simple example. There's product managers on my team who don't write from scratch anymore.

**[0:16:54] Ajay Swamy:** They don't write requirements or P R Ds or P R F Qs from scratch. They actually will use an A I to write that. Because that'll do seventy to 80% of the heavy work. First, based on the context that you give it, and then you sort of iterate and then you sort of customize it, and then you sort of think through your own little nuances. Those are the two examples that come to my head, like at the forefront.

**[0:17:16] Ajay Swamy:** But then also just think about it from the fact of, hey, if I were to give you a chatbot, a Gen E I chatbot with access to your institutional data, your specific research repositories. Imagine a world where you can go in and say, Hey, help me do X, help me do Y, help me create a deck, help me create this word document. Every single one of these takes the pain away from starting from scratch. So the work is actually shifting away from, hey, I wanna do the grunt work of putting together an outline and some of the early messaging.

**[0:17:48] Ajay Swamy:** It takes that away, and your work actually becomes let me be the expert judgment analyst that can look at the output that the A I has created and really sort of tweak it and really make it stronger. So I think that's where sort of some of the knowledge work is moving.

**[0:18:05] Host:** Yeah, I like it, and it's it's kind of like, you know, um, being able to, just if if we're painting the picture, just being able to take, uh uh, the foundation off the shelf and then sculpt it into whatever you needed to be. I like that. Ajay, this has been a great conversation. I think our takeaways today would be that the bottleneck in B F S I operations usually isn't a broken system. It's like we said, the handoffs creating fragmentation, different people, different policies, different platforms, and they were all designed to not talk to each other.

**[0:18:35] Host:** Which is, which is creating a lot of frustration. And we're going to have to evaluate our tools better. We're going to have to look at, um, the platforms, how they can talk to each other, but at the same time, prioritize explainability and traceability. Like you mentioned, anything that you want to leave the audience with today before we wrap it up.

**[0:18:54] Ajay Swamy:** Yeah, look, if there's one thing to take away, it's the following, right? Like, stop trying to make financial operations faster. I think they're already fast, given the circumstances, given the complexity o of the world that they operate in. I think you need to make them defensible at speed. So the bottleneck was never really the speed. These folks are doing the best they can given the circumstances. And it was trust, right? If you don't automate trust, you don't want to automate trust.

**[0:19:21] Ajay Swamy:** You have to engineer it. Now, it just doesn't come naturally out of the box, like we talked about. So that's the whole game now, really. Don't try and automate trust, engineer it deliberately into your processes, make sure that's grounded, and then you'll have a workflow or some sort of an automation capability that you can actually trust. And make sure that it's working well and you can sleep at night.

**[0:19:45] Host:** Great conversation. Thank you so much for your time today. Very excited to to go and recap on all of this and hopefully have you back to take this conversation even further in a few months and see where we're at then.

**[0:19:58] Ajay Swamy:** Sounds great. Thank you for having me on.

**[0:20:14] Host:** Wrapping up today's episode, I think there are three key takeaways from our conversation with Ajay. First, the hardest part of scaling financial operations isn't transaction volume, it's the judgment buried inside steps that look routine, especially where the data is fragmented across systems. Second, explainability and traceability aren't optional extras. They're what determines whether an A I assisted decision can be defended to a risk committee, a customer, or a regulator. And finally, trust in A I assisted decisions doesn't happen on its own.

**[0:20:45] Host:** It has to be deliberately engineered into how these systems are built and overseen. In this episode, we covered how fragmented data and manual judgment calls are slowing down financial operations. To go deeper on this topic and learn how financial institutions are digitizing paper-based records to unlock usable data for A I, and using alternative data like public web and social signals to enhance risk assessment. Download our free P D F report, A I in Financial Services Executive Cheat Sheet at emerj.com/fcs1.

**[0:21:17] Host:** That's E M E R J dot com slash F C S one to download your copy. On behalf of the team at Emerge, we'll see you on the next episode.