Investment banking’s junior years have always worked like a medieval guild apprenticeship. You don’t grind through 90-hour weeks building comps because comps are hard. You grind because that’s how the guild decides who’s worthy. The work is the initiation. The output is almost beside the point.
So when OpenAI launches ChatGPT for Financial Services in 2026 — a product built to research companies, crunch financial data, and produce the formatted pitchbook decks that analysts assemble at 3 a.m. — the interesting question isn’t whether the AI can do the work. It’s what happens to an industry whose training pipeline was never really about the work.
What OpenAI actually shipped
The scope is specific and it’s aimed squarely at the bottom of the org chart. Company research. Running comps. Testing valuation scenarios. Building pitchbooks. These are the four pillars of the first-year analyst experience, and OpenAI named all four.
The detail that tells you how seriously this was built: OpenAI reportedly hired more than 100 ex-bankers to train the model on M&A, LBO, and IPO work. That’s not a generic model with a finance system prompt bolted on. That’s paying people who lived inside the deal process to encode what the process actually looks like — the formatting conventions, the unwritten rules about which multiples matter, the specific ways a managing director will send a deck back.
The target circulating in the discussion is automating roughly 60% of junior banker tasks. Treat that number with the skepticism any vendor-adjacent figure deserves. But even a fraction of it changes the math on how many warm bodies a coverage group needs.
Why I’m not fully sold
I review AI tools for a living, and the pattern I see over and over is that automation aimed at junior work has a habit of generating senior work. There’s already commentary in the space pointing out that Rogo, a finance-focused AI product, was supposed to shrink the junior class and instead created more work for associates. Someone has to check the output. Someone has to catch the hallucinated comp, the wrong fiscal year, the transaction multiple that quietly includes a company that isn’t remotely comparable.
In banking, a mistake in a pitchbook isn’t a bug ticket. It’s a credibility event in front of a client considering a multi-billion-dollar transaction. That asymmetry — cheap to generate, expensive to be wrong — is exactly the environment where verification costs eat the savings. The analyst who used to build the deck now audits the deck. Whether that’s faster depends entirely on how good the model is on the specific, ugly, non-standard situations that make up most real deals.
There’s a second friction point that’s almost comedic. Bankers have gotten sharp at spotting AI output, and it’s become a liability in interviews. The industry that’s buying the tool is simultaneously developing an allergy to anything that smells like it. You can’t automate the deliverable and keep the credential intact at the same time.
The part nobody at OpenAI has to answer for
The debate this launched is about entry-level finance jobs, and it deserves to be. Analyst programs are how the industry manufactures its own future — you learn valuation by building models badly and getting yelled at, not by reading about valuation. Compress that stage and you get a generation of associates who can review work they never learned to produce.
Banks may not care. Analyst classes are a cost center that exists partly by tradition. If a tool trained by 100 former bankers gets a first draft to 80% for a fraction of a salary, the incentive to shrink the class is immediate and the cost of a thinner talent bench arrives years later, on somebody else’s watch. That’s the kind of tradeoff organizations make badly and consistently.
My read
This is a serious product built with serious domain investment, and anyone claiming banking is immune is not paying attention. But the framing of replacement is doing a lot of work that the evidence doesn’t yet support. What I’d expect in practice is fewer analysts doing more oversight, deals that move faster on the mechanical parts and no faster on the judgment parts, and a two-year stretch where nobody can tell whether headcount cuts were the AI or just the market.
If you’re a student sending CVs into the void hoping for a banking seat, the honest advice is uncomfortable: the tasks that got you hired are the ones being productized. The skills that survive are the ones that were never on the job description — reading a room, owning a judgment call, knowing which number the client will actually fight about. Those were always what the apprenticeship was secretly teaching. Now they’re the only part left worth teaching.
🕒 Published: