Remember when spreadsheet software first landed on accountants’ desks and everyone braced for the end of the profession? The panic was real. Ledger books were being replaced by cells and formulas, and the people who filled those books by hand were told their skills were about to become decorative. What actually happened was messier and more interesting: some jobs vanished, many changed shape, and a whole category of financial modeling work appeared that nobody had imagined.
I keep coming back to that memory because in 2026 OpenAI launched ChatGPT for Financial Services, and the target this time is remarkably specific. Not “knowledge work” in the abstract. Not “productivity.” The junior investment banker. The 23-year-old with the second monitor and the 2 a.m. smooth order.
What the tool actually does
Stripped of the announcement language, this version of ChatGPT is built to handle three things: researching companies, analyzing financial data, and generating the presentations investors read. If you’ve never worked in banking, those bullet points may sound like a modest feature list. Inside a bank, they describe most of the first two years of an analyst’s career.
The specific tasks getting automated include:
- Company research — pulling together everything known about a business before a deal conversation
- Running comps — comparing a company against similar businesses to estimate what it’s worth
- Testing valuation scenarios — asking “what if growth slows” and recalculating the answer
- Building pitchbooks — the formatted slide decks banks use to win and explain deals
OpenAI didn’t guess at how these tasks work. According to a LinkedIn post from Paul Cuatrecasas, the company hired more than 100 ex-bankers to train the system on M&A, leveraged buyout, and IPO work, with a stated target of automating around 60% of junior banker tasks.
Why hiring the bankers matters more than the model
This is the part I want non-technical readers to sit with, because it’s the most useful thing to understand about how AI agents get good at real jobs.
A general-purpose AI can describe what a leveraged buyout is. It read about them. But it doesn’t know that your bank formats the sources-and-uses table a particular way, or which adjustments get made to earnings before a comp set is considered credible, or what a managing director will send back at midnight with three words of feedback. That knowledge doesn’t live in textbooks. It lives in the muscle memory of people who did the work.
Hiring 100+ ex-bankers is a way of extracting that muscle memory and encoding it. It’s expensive, unglamorous, and it’s the actual method. When you see an AI agent that seems unusually competent in a narrow professional domain, the story behind it is almost always this: someone paid practitioners to teach it.
For anyone trying to figure out where AI agents will land next, that’s your signal. Ask whether the tacit knowledge of a job can be captured by a few hundred people sitting down and explaining it. If yes, the tooling is coming.
The training pipeline problem
Here’s what makes the Wall Street case genuinely thorny rather than just another automation story. Those labor-intensive tasks weren’t only output. They were the training program.
Banks have used comps, models, and pitchbooks to teach new hires how deals actually work. You learn valuation by building valuations badly, getting corrected, and building them again. The grunt work was the apprenticeship. Remove it and you’ve solved a cost problem while creating a development problem: how does someone become a senior banker without ever having been a junior one?
The debate this launch sparked about entry-level finance jobs is mostly framed as a headcount question. I think the sharper question is about pipeline. A bank can run leaner next year. It’s less obvious how it produces experienced dealmakers a decade out.
And the early evidence suggests the transition isn’t clean. One commentary on the space noted that Rogo, another AI tool aimed at shrinking the junior class, ended up creating more work for associates instead. That pattern shows up constantly with agent adoption: the tool produces more output, someone still has to check it, and the checking lands on the person one rung up.
What I’d tell a nervous 22-year-old
Automating 60% of a task list is not the same as automating 60% of a job. The remaining 40% tends to be the part nobody wrote down: judgment about which analysis matters, reading a room, knowing when the model is confidently wrong.
That said, I’d stop treating “I can build a clean comp set” as a career moat. It was a real skill. It is becoming a checkbox. The people who do well here will be the ones who get good at directing these systems and catching their mistakes, which is a different skill that nobody has ten years of experience in yet.
Spreadsheets didn’t end accounting. They did end a specific way of being an accountant. This feels like that, arriving faster, with a much clearer target painted on it.
🕒 Published: