In the AI Era, the Most Valuable Thing Isn't Intelligence — It's the Audit Trail
Rogo founder Gabe Stengel says models are already smarter than anyone he knows, and what's left is all plumbing — whoever runs the pipes into banks' compliance and systems of record wins.
The video won't play here. Listen to the audio instead:
The argument · tap a timestamp to hear it
Model companies are refineries; vertical companies are chemical plants
Gabe reasons from Dario's line: if a whole data center full of geniuses showed up tomorrow at the door of Goldman, Millennium, Citadel, those institutions would still take a long time to figure out how to change the way they work and how to wire it into their systems. So he thinks the biggest challenge for every great investor over the next five years isn't getting access to intelligence, it's wiring AI into the investment lifecycle. That's also why he isn't worried about Anthropic and OpenAI: finance is a collection of niches with wildly different data, definitions, and regulatory requirements, and for Anthropic to stop and pick up this coin would be pulling over on the road from $100 billion in revenue to $1 trillion.
— Gabe StengelApplied AI companies have a first-mover disadvantage
Gabe says companies that did applied AI early were actually at a disadvantage: you think you know where the world is going, you build the product for that end state, but the model isn't there yet, and users try it and just say ‘this is garbage.’ Rogo went through this phase itself. His judgment is that if your read on the end state is right and you keep building toward it, the moment the model catches up is the magic moment. He ties Rogo's product eras directly to model eras: 01 Pro was the first version reliable enough to at least serve as a search tool; by Opus 4.5, and from late 2025 into early 2026, the model can basically do anything a junior investment or banking professional does, provided you give it the right instructions and context.
— Gabe StengelThe MD annotating on an iPad is the real last mile
Gabe gives an example as concrete as it gets: the typical workflow of an investment banking MD is to email deck comments to an analyst, and Rogo lets that MD email the comments to an AI analyst, get the revised draft back in 20 minutes instead of two days, automatically notify the junior analysts on the deal, and show a full audit trail of every small change. He points out the key thing: this finance professional hasn't logged into a computer in ten years, but he has an iPad, and he knows how to annotate on an iPad. That kind of product detail for a specific end user can only come from someone with a spidey sense for the job.
— Gabe StengelThe harness matters more than the model itself
Gabe compares Claude Code and Claude Cowork against ChatGPT: the models are actually similar, but Claude's harness and presentation are far better, letting the model exercise more long-horizon capability, and so usage and expansion took off. He extends the analogy to the brain: humans have high agency not just because of raw memory, IQ, and knowledge, but because there are a bunch of microservices in the head — how to store knowledge, how to retrieve it, how to trigger it, and emotion is one way of triggering a microservice. A great investor's judgment comes from seeing something in the market that triggers the memory of some event, leading to a creative decision. Those microservices are what vertical companies have to build.
— Gabe StengelSkip seat-based pricing and go straight to outcome-based
Rogo today is classic enterprise software: priced per seat, because buyers are used to comparing it to Bloomberg, FactSet, Capital IQ, PitchBook, and every deal signed needs an AE plus a solutions architect to shake hands and explain the integration. Gabe says Anthropic can sell to them easily, with spend rising exponentially and no human in the loop, because it's a token-consumption model; but many industries can't ride the token-consumption wave, and Rogo is one of them, so it has to build its GTM machine at five times the speed. His pricing judgment is that every company goes through two pricing revolutions: first to usage-based, then to outcome-based. He'd rather skip tokens and usage entirely and ask the customer directly: how much do you charge per good investment idea? Per quarterly LP report? Per investment banking CIM?
— Gabe StengelAuditability matters more than accuracy
Gabe says accuracy is of course still super important, but auditability matters more than accuracy, and the two get conflated all the time. The logic: if the answer is wrong, and you don't know how to check it, and you can't see where it came from, then whether it's accurate or not you can't use it, because you don't trust it. Conversely, if it's right most of the time, and even when it's wrong you can clearly see the assumptions and the data sources, it's still actionable and still saves time. As tools move from copilot to autopilot — not just retrieving information, but having agency, actually making investment decisions or sending emails — you have to be able to look back at why it decided the way it did, because you need to debug it, just as when an individual investor makes a bad decision you check whether the data input was wrong or someone lied to him. In regulated capital markets, if you can't explain the basis for a decision, you simply don't get through.
— Gabe StengelThe company brain is called Shrek, and every conversation is recorded
Every conversation inside Rogo is recorded, and new hires are told on day one ‘you are always being recorded,’ and that content is filtered into the company brain, named Shrek because the engineers thought that was funny. There's a dashboard showing the ‘swamp’ of what everyone is working on; it's connected to the company's various systems and is very clear about company goals, values, what has to be delivered to customers, and the north star metric, so it shapes every answer and deliverable. It's both passive and active: you can go in and ask ‘how should I pitch the value proposition and savings of model routing to a large institution,’ and it pulls the information out; it also proactively reminds you, ‘Patrick, I see you're meeting this private credit firm on Thursday, here's what you should know, the use cases that will resonate, and the ROI metrics companies you've worked with before want to hear.’ Gabe says this is the core problem of enablement: suck information in, and give it out when people need it.
— Gabe StengelWhat it's like to be rejected by 40 investors in a row
Gabe says the hardest low point was the A round: at the time there were no star investors on the cap table, David Tisch introduced him to 40 investors, he met Sequoia, Kleiner, Benchmark and so on, and all 40 passed. And it wasn't the send-a-deck-and-get-no-interest kind; it was ‘interesting, let's meet Gabe,’ ‘we like Gabe, let's talk for an hour,’ ‘come to IC,’ ‘let's get a meal,’ ‘come back over the weekend’ — and then ‘we've decided not to invest.’ He says that phase had nothing to do with anything except you, like being dumped by 40 girlfriends, each time you fell in love and each time the other person said it wasn't for you. In the end Thrive didn't invest either; Keith Rabois came about a month after everyone had passed, and Gabe asked Keith back: if this isn't a contrarian bet, why did all your friends say it was a bad idea and not believe in me? He sums up three reasons investors didn't invest: they underestimated the finance TAM, the product really was bad at the time, and there weren't enough data points to believe he could chew glass and ship the next version.
— Gabe StengelFinance meets the innovator's dilemma for the first time
Gabe says that over the past 10 to 20 years, capital allocation and investing have been very profitable businesses, and the industry is extremely hard to enter, especially in private markets — raise one fund, then raise another, the business carries its own inertia, and it's hard to screw up. So for a long time there was no market shock that made every investment firm and every bank say ‘I have to completely rethink what I'm doing.’ Now there is one: hundreds of AI-native disruptors, AI-native investment firms, AI-native investment banks will emerge to attack the existing business models. He defines AI-native as being willing to constantly reinvent everything, addicted to AI to the point of not caring how completely unrealistic something looks, just moving in the direction of integrating this alien foundational technology into everything, with no part of the business sacred.
— Gabe StengelIn their own words · checked verbatim
And Dario has the great line about everyone's going to have a data center full of geniuses or a country full of geniuses in the data center. What would Goldman do? What would Millennium do? What would Citadel do if they had a country full of geniuses show up?
Gabe Stengel2:03
I mean, I think there was a first movers disadvantage for a lot of applied AI companies because you thought you knew where the world was going and you wanted to build a product for it, but the models weren't quite there. And so people would try it and go, this is terrible. This is garbage.
Gabe Stengel4:04
The models were actually fairly similar, but the harness and the way that it was presented from Cloud was far better. And it just allowed the models to exercise more of their long running capabilities.
Gabe Stengel16:57
I think that every business needs to go through two different pricing revolutions. You need to move to some sort of usage based and then you need to move to some sort of outcome based.
Gabe Stengel23:43
I think it's actually more important to be auditable than it is to be accurate. And obviously those two things are conflated.
Gabe Stengel31:19
And it's so personal because at that stage, it has nothing to do with anything but you. Right. It's like somewhere it's like being broken up with by 40 girlfriends who like every time you fell in love and then every time they said, not for you.
Gabe Stengel42:18
If you knew for sure that right now, 90% of your enterprise value is in your people, your best investors, your best bankers are the people that bring in deals, bring in revenue. And actually, that's what accrues enterprise value. And in 10 years, the world's best investment firms, best banks will have 90% of their enterprise value, not in people, but in software and data and systems. What would you start doing?
Gabe Stengel53:09
Figures
| Number of people at Rogo who have worked at an investment bank or investment firm | Over 100 | 21:26 |
| Number of early financial AI startups Rogo has acquired | 6 | 50:14 |
| Number of investors who rejected Gabe before the A round | 40 | 42:18 |
| Revenue scale Gabe thinks Rogo can reach by going deep in finance's various niches | $5 billion | 17:19 |
| Share of US mortgages done through online platforms (like Rocket Mortgage) | 40% to 50% | 27:41 |
| Gabe's self-assessment of how complete Rogo's product roadmap is | 1% | 27:41 |
| Window Gabe thinks everyone in finance has to make AI purchasing decisions | 18 months | 44:07 |
Glossary
- harness
- The layer of engineering that wraps the model and determines how it's invoked and presented, shaping the experience more than the model itself.
- MNPI
- Material Non-Public Information, the most compliance-sensitive category of information at investment banks and investment firms.
- IC memo
- The core document an investment firm submits to its investment committee to argue for a deal.
- DDQ
- Due Diligence Questionnaire, a long list of standard questions sent by the other side in a deal or fundraise.
- CIM
- Confidential Information Memorandum, the core pitch document an investment bank shows potential buyers when selling a company.
- compaction
- The ability to let an agent compress its memory into a limited token budget over a long conversation without losing key context.
How to listen
Vertical AI founders, investment banking and PE practitioners, and investors looking at enterprise software and the AI application layer. For anyone who wants to know how that layer of plumbing beyond the model gets built, priced, and sold into large institutions.
The product-era retrospective from 1:24 to 6:14 at the start is mostly setup; you can fast-forward to the mechanics after 16:57.