AI in Finance: Who Owns the Model's Output?

The Data Railroads Just Became Retrieval Layers

Morgan Stanley and Evercore didn't pilot OpenAI's new financial services platform. They co-designed it. And in doing so, they may have just signed the death warrant for an entire layer of the industry's cost structure—including parts of their own.

OpenAI announced this week that it's selling a version of ChatGPT purpose-built for financial services. It plugs directly into LSEG, PitchBook, Crunchbase, and Quartr. It connects to your existing FactSet and S&P Capital IQ subscriptions. It builds financial models. It generates pitchbooks from your templates. And it does in an afternoon what used to require three 80-hour analyst weeks.

The companies providing the data—LSEG, FactSet, S&P—have charged this industry a toll for thirty years. They built the railroads. Now they're retrieval layers inside someone else's model. The value didn't disappear. It just moved up the stack.

I've watched this movie before. Let me tell you how it ends.

Nobody Gets Fired the Day the Railroad Arrives

When electronic trading hit the NYSE floor in the late 1990s, the initial response was partnership. Specialists helped design the algorithms. They understood better than anyone how order flow actually worked, where the inefficiencies lived, what buyers would pay for speed and certainty.

They built the system that made them optional.

Five years later, the floor that once employed 5,500 people was running on a few hundred. The knowledge didn't become worthless—it became embedded in the system. The specialists who survived weren't the ones who fought the technology. They were the ones who figured out what clients would pay for once the technology made their old job free.

Morgan Stanley and Evercore are playing the same game, one layer up. They're betting they can capture the value that moves to judgment, relationships, and strategic advice even as they compress the 80-hour research grind into a background task. They might be right. But the math gets uncomfortable when you realize what they're compressing is the training ground for the people who eventually provide that judgment.

What Your Bankers Are Already Doing

I was on a call last month with a firm that spends seven figures annually on FactSet, S&P, and PitchBook. The question from their CFO wasn't "should we explore AI tools?" It was more specific: "How do I get ahead of what my team is already doing with ChatGPT on their personal laptops?"

Because they are. I've seen the usage logs. Analysts are pasting in financial data, asking for comp analysis, generating first-draft commentary. The tool is already inside the building, policy or no policy. The question isn't whether to allow it—it's whether you have a review process for what it produces.

And here's where most firms are pretending the problem is simpler than it is. They're treating this like adopting Excel—a productivity tool that makes existing work faster. But Excel didn't change what accountants signed. It changed what was worth paying an accountant to do. The arithmetic became free. The judgment about what the numbers meant became the entire job.

Same move, one layer up. The model does the research. Someone in your firm still owns the recommendation a client acts on.

The Control Isn't in the Audit Log

OpenAI's pitch to financial services emphasizes "exportable logs for audit." Every query tracked. Every source cited. Full transparency into what the model accessed and how it built its answer.

That's not a control. That's a receipt.

The control is your review process. Your template for what "done" looks like before the client sees it. Your judgment about when the model's output is ready to carry your firm's letterhead and your professional liability.

I'm working with firms right now trying to build that review process in real time, while their teams are already using tools that didn't exist six months ago. The ones getting it right aren't starting with the technology. They're starting with the question: What are we promising the client when we put our name on this deliverable?

The model can generate a comp analysis. Can it determine which comps are relevant in this specific M&A context, given this client's strategic goals and this moment in the market cycle? Can it assess whether the data it retrieved is current enough, complete enough, representative enough for the decision the client is about to make?

Maybe. Maybe not. The uncomfortable part is that "maybe" now lives inside your workflow, and most firms don't have a process to resolve it.

Who Collects the Rent?

Here's the pattern I've seen play out across four technology disruption cycles: The firms that survive aren't the ones that fight the new tool or the ones that adopt it fastest. They're the ones who figure out what clients will pay for once the old service becomes a commodity.

Legal research used to be billable hours. Westlaw and LexisNexis turned it into a flat-fee subscription. The lawyers who survived didn't charge for research—they charged for judgment about what the research meant for this specific case.

Tax preparation used to be arithmetic. TurboTax made it a $60 software purchase. The CPAs who survived didn't charge for data entry—they charged for advice about structuring the client's affairs to minimize the liability in the first place.

Financial research is becoming a retrieval problem. What's left?

The firms that co-designed this tool are betting the answer is strategic advice, relationship management, and judgment about what the model can't see—management quality, cultural fit in M&A, the political dynamics that kill deals even when the numbers work. They might be right.

But I keep coming back to the uncomfortable math: If you compress the 80-hour analyst experience into an afternoon task, where do the people who eventually provide that judgment get trained? The specialists who helped build electronic trading understood order flow because they'd spent years on the floor. When the floor disappeared, the pipeline disappeared with it.

I don't have a clean answer. But I know what happens when an industry pretends the question isn't real.

What to Do Monday Morning

If your firm pays for FactSet, Bloomberg, S&P, or PitchBook, you're renting access to the railroads. The railroads just became retrieval layers. That doesn't mean they're worthless—it means the value is moving.

Here's what I'm telling the firms I advise:

Ask your team what they're already doing with AI tools. Not whether they're using them—they are. What they're using them for, what review process they're applying, and what they're telling clients about how the work was produced.

Define what "done" looks like before the client sees it. Not "the model generated it and the log shows the sources." What level of review, what judgment calls, what verification steps happen between the model's output and your firm's letterhead.

Figure out what you're actually selling. If it's research, you're competing with a tool that costs $60/month and ships next quarter. If it's judgment, what's your process for developing the people who provide it when the training ground just compressed from 80 hours to an afternoon?

The tool is already built. The design partners made sure of that. The question isn't whether it works—it's whether you've figured out what your firm gets paid for once it does.

Excel didn't kill accountants. It changed what they got paid for. The firms that survived weren't the ones that did arithmetic faster. They were the ones who owned the judgment about what the numbers meant.

Own the judgment, and you still collect the rent. Rent it, and you're the town the railroad runs through.

But what do I know—I've only watched this movie four times.