OpenAI on Thursday introduced ChatGPT for Financial Services, an enterprise product for company research, financial analysis and client presentations inside investment banks. Nick Turley, the company's vice president of product, said it was built on ChatGPT Work with Morgan Stanley and Evercore as design partners, and that it runs on GPT-6 Astra, OpenAI's newest model. What the product advertises is not a capability but a job description: comparable company analysis, leveraged buyout models, buyer screens, earnings breakdowns, pitchbooks. Those are the tasks Wall Street has handed to its youngest employees for decades.
The difference from the generic enterprise version is data, not reasoning. ChatGPT for Financial Services gets direct access to financial statements, quarterly earnings call transcripts, company metrics and private-company data from LSEG, Daloopa, PitchBook and Crunchbase. It also connects automatically to the database subscriptions a bank already pays for. Every figure carries a link back to its source document, charts can be reconciled against the underlying data, and administrators get controls over access to confidential deal material.
That last cluster is the part banks are actually buying. A model that produces a number no one can trace is useless on a deal team; a model that produces the same number with a citation and a permissions boundary around it is a piece of infrastructure. OpenAI spent its integration budget on provenance and access control rather than on making the output sound smarter.
In the demonstration, Turley asked the system to assess a potential merger or acquisition target, pull the relevant metrics, and produce a formatted PowerPoint deck in the bank's own template. The hard part, he said, was not making the slides look good but making their contents make sense: the model had to choose an appropriate comparable set, load prices into a table, check a chart against the source data, and explain a decline and subsequent recovery in the share price.
Turley said demand is high and that OpenAI plans equivalent versions for other industries. He did not name any bank that has already signed.
The launch fits the company's wider direction. Sarah Friar, OpenAI's chief financial officer, told investors in August that the enterprise business now generates more revenue than the consumer one. The company is expected to be preparing an initial public offering. And OpenAI is not first here: Anthropic shipped Claude for Financial Services last year.
Banks are among the most expensive customers software vendors have. They pay for data and for reliability, and they employ tens of thousands of people whose daily work maps unusually cleanly onto what large language models do. Investment banking was the obvious first vertical for products like this, and it is also the clearest test case.
Asked by CNBC whether the product would reduce junior analyst hiring, Turley called the system a productivity tool rather than a replacement. Analysts work 100 hours a week, he said, and the effect will resemble the arrival of Microsoft Excel: the same people doing analysis faster and better.
The Excel comparison does not survive the demonstration he had just given. Excel sped up the analyst's reasoning; it did not pick the comparable set, decide which chart the argument needed, or explain why a stock fell and then recovered. A tool that does those things takes over precisely the portion of the work through which the analyst was learning in the first place. This reads less like a productivity claim than like the answer a vendor gives when the honest answer is that the customer decides.
Inside the industry, the same risk is being named out loud. On the August 24 episode of Goldman Sachs Exchanges, Chris Churchman — a partner who runs the bank's Marquee platform and co-chairs the AI working group in Global Banking and Markets — warned of "cognitive atrophy" if bankers start handing their reasoning to models. A large share of financial knowledge is tacit and acquired in practice, he said, comparing the likely effect to the way GPS weakened people's sense of direction. Churchman acknowledged that Goldman Sachs has not yet decided how much entry-level work should be automated.
On the question of learning by doing, there is already an early measurable signal. Research from Stanford's Digital Economy Lab, updated in August, found that among workers aged 22 to 25 in the occupations most exposed to generative AI, employment is now 19% below where it would have been on a trajectory comparable to less exposed occupations. More experienced workers in the same occupations show no comparable gap. The researchers attribute the difference to firms hiring fewer young people rather than to layoffs of people already employed, and the decline is sharpest where AI automates tasks rather than augmenting them.
The Federal Reserve Bank of Dallas reached a similar conclusion in an analysis published in January: AI-linked employment losses appear only among young workers, and mostly because fewer people who previously had no job are entering those occupations.
Investment banking is one instance of a broader pattern, but this launch makes the pattern legible. If a model can assemble a defensible pitchbook in minutes, the economics of the two-year analyst program change. The program was never primarily a way to get pitchbooks made — it was how banks paid to train people, with the pitchbook as the tuition. Seb Kirk, chief executive of the AI platform GaiaLens, put the mechanism plainly to Moneywise: nobody learns to make decisions by reformatting a deck at two in the morning, but a person who works through the numbers themselves notices when one of them is wrong.
What happens to the role depends on a choice each bank now has to make explicitly rather than by default. The job can be rebuilt around checking the model — auditing assumptions, testing comparable sets, sitting in client meetings that used to be reserved for senior staff — or hiring can simply be cut. OpenAI has built the tool that does the number work. The banks that choose the second option will not find out what was inside the part they removed until the people who would have learned it are the ones running the deals.