Gabe Stengel on Rogo and AI in Finance

Twenty years ago, getting a mortgage meant a trip to a bank branch and a long conversation with a loan officer. It felt like the kind of decision that needed a human in the room. Today, forty to fifty percent of mortgages are delivered entirely online, through platforms like Rocket Mortgage, with no loan officer required (28:32). Gabe Stengel, founder and CEO of the AI finance platform Rogo, thinks the rest of Wall Street is about to have its own quiet mortgage moment. Deals that take five months to price could take five minutes (28:58). The people doing that pricing today, analysts, associates, junior bankers, are the ones whose work Rogo is built to absorb.

Stengel has tried to build this company twice before, without success. In high school he built a tool to track merger exchange ratios. In college he published a paper on AI for financial econometrics. Both times, in his words, nothing worked at all (03:23). Rogo itself launched right as GPT-3 arrived, before ChatGPT existed, and for a long stretch the product was more demo than tool. What changed the trajectory was not a better idea. It was the models. OpenAI's o1 Pro, Stengel says, was the first model reliable enough to act as a real search tool, capable of calculating a financial metric across twelve quarters without making things up (03:43). Anthropic's Opus 4 and 5, arriving toward the end of 2025, pushed further: they became capable of doing basically anything a junior investment banker was doing, given the right instructions and context (04:04). Stengel calls this a first mover's disadvantage for applied AI companies. Build for where the technology is going before it arrives, and the early product looks terrible, so people write you off (04:28).

Plumbing Over Intelligence

The obvious worry for a company like Rogo is that the big AI labs will eventually build the same thing for free. Stengel's answer is that intelligence is only part of the product. The harder, less glamorous part is what he calls plumbing: compliance workflows, audit trails, data rooms, and the systems that plug into a bank's CRM and portfolio-monitoring tools. When a managing director emails a deck to be marked up, Rogo can return it in twenty minutes instead of the two days a human turnaround would take, while alerting the junior analyst on the deal and preserving a record of every edit made (05:40). None of that is intelligence in the raw sense. It is integration, worked out one workflow at a time.

That distinction is also Stengel's answer to the question everyone asks him: how does Rogo compete with OpenAI and Anthropic? His view is that the labs are chasing markets worth hundreds of billions of dollars, and the narrow, messy problems of financial services, MNPI compliance, deal-room security, regulatory audit trails, sit beneath their notice, even though each of those niches can be worth five to ten billion dollars on its own (18:03). Rogo's bet is that finance is really dozens of such niches stacked together, each with its own data, its own definition of good, and its own regulator, and that owning the plumbing across all of them builds a business the labs will never bother with.

This is also why Stengel says accuracy is not the main event. Auditability is.

"I think it's actually more important to be auditable than it is to be accurate, and obviously those two things are conflated." Gabe Stengel, [34:39]

His logic: an answer that is occasionally wrong but shows its sourcing is still useful, because a banker can check it. An answer that is usually right but opaque is useless, because nobody trusts a black box in a regulated industry. As AI tools move from co-pilot, where they retrieve information, to autopilot, where they act on a client's behalf, that traceability stops being a nice feature and becomes something regulators require (35:06).

The Cost of Conviction

Rogo's business model is unusual for an AI company. Most of the buzzy AI startups, Cursor, Factory, and Cloud Code among them, ride token consumption: usage rises automatically, sometimes a hundredfold on a single contract, without adding a single salesperson (25:44). Rogo cannot do that. Its buyers, big banks used to paying for Bloomberg terminals and FactSet subscriptions, expect per-seat pricing, and every deal requires an account executive, a solutions architect, and a sales engineer walking the customer through integration by hand (25:21). Stengel's long-term goal is to skip past both seat pricing and token pricing and charge for outcomes instead: a fee per good investment idea, per LP report, per banking simulation (26:46). It is a simpler pitch than trying to explain a token bill nobody can translate into value.

The tradeoff is speed. Enterprise sales are limited by how fast people can be trained, and Stengel says the real constraint on Rogo's growth is not the model, it is how quickly a new hire becomes useful. To solve that, Rogo records every internal conversation and feeds it into a tool the team nicknamed Shrek, so a banker in New York can pull up context from a deal a colleague closed in Asia without ever having met them (36:42). It's a small detail, but it captures Stengel's broader argument: the interesting AI problems in finance are rarely about raw model capability anymore. They are about memory and coordination, a problem he calls compaction, and one that current models still handle badly once more than a couple of people are involved (23:27).

None of this came easily. When Rogo raised its Series A, an early investor introduced Stengel to forty venture firms. All forty passed, some after weeks of meetings and dinners that felt, in Stengel's words, like being broken up with by forty girlfriends (45:52). Part of the resistance was structural: Silicon Valley investors, he argues, lack intuition for how large the financial data market is, because prior giants in the space, Bloomberg, FactSet, ION Group, were never venture-backed to begin with (46:32). Investor Keith Rabois eventually backed the company, reportedly telling Stengel the idea was not even contrarian, it was just the legal-AI company Harvey applied to finance. Stengel's response to the doubt was to raise the stakes rather than lower them.

"It's fine if it gets thirty percent more likelihood that I fail if the odds that I become a hundred billion dollar company also increase by twenty percent." Gabe Stengel, [44:43]

That appetite for risk explains his most sweeping claim: that in ten years, the world's best investment firms and banks will hold ninety percent of their enterprise value not in people but in software, data, and systems (53:46). If that is right, the job of a firm's leadership is not to buy a chatbot. It is to work out which parts of a top banker's judgment are genuinely irreplaceable, built on relationships and hard-won context, and which parts are just information a system could hold instead.

Stengel does not pretend the human side disappears entirely. He doubts a small business owner handing off a company built over twenty years will ever be comfortable doing it by clicking a button instead of shaking a hand (56:46). But he thinks that exception is narrower than most people in finance believe, and narrower every year. The interesting question is not whether AI reaches capital markets. Structurally, it already has. The question is which of finance's thousand small human-run rituals turn out to be irreplaceable, and which ones were simply habits nobody had gotten around to automating yet.