Latest / Boardroom Conversations with Fexingo: CEOs, Strategy, and Corporate Leadership Explained / How JPMorgan Built Its Risk Management Moat
Transcript
- Lucas: If these conversations have moved your work forward in some small way, consider supporting the show directly. A couple of dollars a month is genuinely what keeps these going — buy me a coffee dot com slash fexingo — that's F. E. X. I. N. G. O. Luna: Thanks for that. It really does help keep the lights on for us. So, let’s get into it. Lucas: Let’s talk about JPMorgan Chase. Specifically, their internal AI engine called COiN. Most people think of AI in banking as fraud detection or chatbots, but JPMorgan took a radically different path. They built a system specifically to read commercial loan agreements. Luna: Wait, just reading contracts? That sounds boring compared to predicting market crashes. Lucas: It is boring, which is exactly why it works. Before COiN, lawyers and analysts spent forty million hours a year manually reviewing these documents. They were looking for specific clauses related to collateral and financial covenants. It was tedious, expensive, and prone to human error. Luna: So they used AI to do the reading instead. How much faster is it now? Lucas: The system can process those same agreements in seconds. We’re talking about a reduction from millions of hours to just a few thousand. But here’s the interesting part: they didn’t outsource this to a vendor. They built it in-house using their own data. Luna: Why build it in-house? Couldn’t they just buy a legal tech solution off the shelf? Lucas: That’s the key question. Off-the-shelf software is designed for general use. These loan agreements are incredibly complex, with unique language specific to JPMorgan’s risk appetite. An external tool wouldn’t understand the nuance of their specific compliance requirements. By building it internally, they controlled the accuracy and the speed. Luna: So it’s not just about speed, it’s about precision. If the AI misses a clause, the bank takes on more risk. Lucas: Exactly. And that’s where the real moat comes in. It’s not just that they saved time; it’s that they reduced operational risk. When you automate something this critical, you eliminate the fatigue factor. Human eyes get tired after hour three. The algorithm doesn’t. Luna: But did they face resistance from their own lawyers? Lawyers love their billable hours. Lucas: They did initially. There was a cultural shift required. The lawyers had to trust that the AI wasn’t replacing them, but augmenting them. The narrative changed from 'the robot is taking my job' to 'the robot is doing the drudgery so I can focus on strategy.' It took time to get there, but once it happened, adoption was rapid. Luna: That makes sense. People don’t hate technology; they hate being replaced without explanation. If you frame it as a tool, they’ll take it. Lucas: Right. And look at the numbers. The system saves roughly one hundred million dollars a year in operational costs. That’s not pocket change. For a bank of that size, it’s significant, but more importantly, it frees up talent to do higher-value work. Luna: One hundred million dollars. That’s a lot of lawyer salaries. Lucas: It is. But here’s where it gets even more interesting. They started applying this same logic to other areas. Like supply chain financing. They realized that if they could automate document review, they could also automate data extraction for broader financial modeling. Luna: So it became a platform, not just a single tool. That’s how Salesforce scaled, right? Lucas: Similar playbook. Once you have the infrastructure for natural language processing in a regulated environment, you can apply it to credit risk, to anti-money laundering checks, to customer service transcripts. It becomes a core competency rather than an IT project. Luna: I’m curious about the training data though. Did they have to clean up years of messy documents first? Lucas: Oh, absolutely. Garbage in, garbage out. They spent months cleaning and labeling historical documents to train the model. It wasn’t plug and play. It was a massive data engineering effort before any machine learning actually happened. Most companies skip that step and wonder why their AI fails. Luna: So the secret isn’t the AI itself, it’s the data hygiene. Lucas: Precisely. You need clean, structured, labeled data to teach the system what to look for. JPMorgan had decades of legal documents sitting in databases. They just hadn’t made them accessible to machines until recently. Luna: That’s a huge barrier to entry. Small banks probably can’t afford that kind of data cleanup. Lucas: They can’t. Which means this widens the gap between the big players and everyone else. JPMorgan isn’t just getting bigger because of capital; they’re getting bigger because their operating costs per transaction are lower due to tech efficiency. Luna: It’s a network effect of efficiency. The more data they process, the smarter the system gets, which lowers costs further, allowing them to underprice competitors. Lucas: Exactly. It’s a flywheel. And they’re not stopping at loans. They’re looking at insurance claims, trade finance, even regulatory reporting. Every area where humans are bottlenecked by paperwork is a target for COiN. Luna: Do you think other industries should copy this? Like healthcare or construction? Lucas: Absolutely. Healthcare has similar document-heavy processes. Patient records, insurance claims, consent forms. Construction has bids, contracts, change orders. Any industry drowning in paper is ripe for this kind of intervention. Luna: But again, only if they’re willing to invest in the data layer upfront. Lucas: Yes. You can’t bolt AI onto bad data. You have to fix the foundation first. It’s unsexy work, but it pays off. Luna: Unsexy work is usually the hardest work. I respect that about JPMorgan. They didn’t chase hype; they chased efficiency. Lucas: They didn’t chase hype. They chased pain points. And they solved them systematically. That’s how you build a durable competitive advantage in 2026. Luna: So, is this the future of all large corporations? In-house AI teams solving specific operational bottlenecks? Lucas: I think it’s becoming the standard. The era of buying generic software is ending. The era of building custom tools for specific problems is beginning. And JPMorgan is leading that charge. Luna: Interesting. It makes you wonder what other hidden efficiencies are sitting in our own company’s archives, waiting to be digitized. Lucas: Probably dozens of them. The question is, are we willing to do the dirty work to find them? Luna: Well, that depends on whether we have a JPMorgan-sized budget. Lucas: Maybe not the budget, but the mindset. Start small. Pick one painful process. Clean the data. Build a prototype. See if it works. Luna: Simple advice. But rarely followed. Thanks for breaking it down, Lucas. Lucas: Anytime. Let’s see where this goes next.