OpenAI shipped ChatGPT for Financial Services on September 10. The load-bearing sentence in the announcement is not about the model. It names five data vendors wired into the same workspace where the analysis happens: LSEG News, PitchBook, Daloopa, Crunchbase and Quartr. Earnings transcripts, financial statements and company fundamentals now sit inside the window instead of behind four browser tabs.
Read that as a workflow decision, not a benchmark claim.
What got plugged in
Firms that already pay for their own terminals can connect FactSet, S&P Global, Preqin and Datasite on top of the built-in five. OpenAI said it re-indexed the licensed datasets on its own infrastructure so retrieval lands on the right row and the citation traces back to a source.
Morgan Stanley and Evercore signed on as design partners. One bulge-bracket, one independent advisory shop. Between them they cover most of how a research or coverage analyst actually spends a day, and OpenAI built to that.
The scope OpenAI published has three legs: research across multiple sources, financial model construction, and client material produced from the firm’s own templates, pitchbooks included.
Underneath is GPT-6 Astra, the model OpenAI released on September 3. We covered that launch when it happened: Astra’s headline skill is operating a computer end to end, carrying a multistep job through to a finished document without stopping to ask what comes next. That piece argued the capability would only become legible once someone pointed it at a specific industry. Seven days later, the industry arrived.
The governance layer shipped with it. The product inherits ChatGPT Enterprise encryption and role-based access, and compliance teams can export workspace logs into their existing audit pipeline. That feature list is not aimed at analysts. It is aimed at the risk, infosec and technology committees who decide whether a tool touches client data at all, and who have killed more bank AI pilots than any model limitation.
Why this one clears a bar the others did not
For two years, AI on Wall Street sat in the same spot: a competent assistant that could not reach the things the job runs on. It would tighten your prose. It did not know what your firm’s pitchbook template looked like, and it could not query the data subscription your desk pays six figures a year for.
Both gaps closed on September 10. The subscriptions connect. The templates load. That moves the tool from a window beside the work into the work itself.
The cohort point is Harvey. On September 9 the legal-AI company disclosed 80% penetration of the Am Law 100 alongside a $15.5B valuation. Law and banking are the same business wearing different suits: billable pyramids where the bottom row converts raw material into something a partner can sell. Harvey got to 80% on the legal side. OpenAI is entering banking with two of its customers as co-designers, roughly eighteen months behind, and with a distribution base no legal vendor has.
The exposed cohort
A first-year investment banking analyst spends the year on four things: pulling information, updating models, running comps, and assembling all three into client material.
That is the product scope, restated. CNBC’s headline said it plainly: the launch targets the work of Wall Street junior bankers.
The transmission mechanism will not be a layoff announcement. No bank sends a September memo eliminating the analyst title. It does something quieter and slower: it shrinks next year’s analyst class. In July we reported that entry-level hiring in law and consulting fell 35% from 2023, with PwC planning 39% fewer audit hires by 2028. None of that decline was announced at a podium.
Thin the base and the spire holds for a while. The bill comes due in three to five years, when the VP bench is drawn from a cohort that was never hired.
The number Morgan Stanley already published
Morgan Stanley is a design partner here. Its own research desk put a figure on this on May 28: AI could let European banks cut 10% to 20% of headcount over five years, more than 200,000 roles by 2030, against an estimated 30% productivity gain. The analysts, led by Giulia Miotto, named back-office operations, risk and compliance as the concentration points.
Hold those two facts next to each other. The institution that published the 20% estimate in May is co-designing the tooling in September. The research desk priced someone else’s headcount; the business is customizing its own workflow. Same trade, two vocabularies.
One more marker. On May 21, Jamie Dimon told Bloomberg that JPMorgan would hire “more AI people and probably less bankers in certain categories”. In May that was a sentence. In September it is a product, built with two of his competitors.
What to watch
Three markers, in the order they should appear.
Deployment disclosure from Morgan Stanley or Evercore. Design partners run narrow first and publish only once internal numbers hold up. Whichever one first names a seat count or an hours-saved figure becomes the citable case for this cycle.
The 2027 analyst class. Banks set that number each spring around campus recruiting. It is the most honest indicator on this side of the trade, and considerably more honest than any AI strategy slide.
The second vertical. OpenAI said it plans to expand across the wider financial services sector. Asset management, private capital and insurance underwriting are structurally identical to banking on the dimension that matters here: heavy document assembly, heavy templated output, heavy junior labor. Once the first vertical clears compliance, the rest do not require new engineering.
The tool is not sitting next to the analyst. It is sitting in the chair.
