OpenAI did not build a banker; it built a very expensive intern with no memory of last quarter’s deal and no ability to be blamed for anything.
That distinction matters more than the headlines suggest. ChatGPT for Financial Services, launched in 2026, targets a specific slice of investment banking labor: company research, financial data analysis, and the generation of formatted presentations. OpenAI reportedly hired 100+ ex-bankers to train the system on M&A, LBO, and IPO work, with a stated target of automating roughly 60% of junior banker tasks. Those are the four things a first-year analyst does between midnight and 4 a.m. — pulling comps, testing valuation scenarios, researching a target, and assembling the pitchbook.
So yes, the task coverage is real. But as someone who spends her time looking at how agent systems actually behave under load, I want to separate two claims that keep getting collapsed into one: this system can perform junior banker tasks and this system can replace junior bankers. The first is an architecture question. The second is an organizational one, and the architecture does not answer it.
What the task decomposition actually tells us
Look at what got picked. Research, data analysis, and slide generation share a useful property: each one is a bounded transformation with a verifiable output. You can check whether a comp set includes the right peers. You can check whether a DCF’s terminal value assumption is stated. You can look at a slide and see if the numbers on it match the numbers in the model.
This is exactly the shape of work that current agent architectures handle well. Short-horizon tasks, dense supervision signal, an output artifact that a human reviews before it leaves the building. Hiring 100+ practitioners to generate training data makes sense precisely because these tasks have conventions — a bank’s comp methodology, the standard LBO structure, the pitchbook’s expected narrative arc. Conventions are learnable. That’s the whole trick.
What was not picked is equally informative. Nobody claims the system runs a live deal process, manages a client through a bad week of price discovery, or decides which of three structures to walk into a boardroom with. Those are long-horizon tasks with sparse feedback, ambiguous success criteria, and accountability that has to attach to a person. Current agent designs are weak there, and the product scoping reflects an accurate read of that weakness.
The training pipeline problem
Here is where the story gets more interesting than a jobs panic. Wall Street has used those grinding analyst tasks as its training pipeline. You learn what a business is worth by building the model wrong three times. You learn which comps are defensible by having an MD tear your set apart. The tedium is not incidental to the education; it is the mechanism of the education.
Remove the mechanism and you get a curious structural hole. You still need people who can judge whether the machine’s comp set is defensible, but you have removed the process that produced that judgment. This is the same problem that shows up in every domain where agents absorb the bottom rung: the review capacity that makes automation safe was itself a product of doing the work manually.
Reports on adjacent tooling already hint at how messy this gets in practice — one account of the AI platform Rogo describes it creating more work for associates rather than shrinking the junior class. That’s the pattern I’d expect. Generation gets cheap, verification does not, and verification load lands on whoever sits one layer above the generator. If your model output is 90% right and the 10% is unmarked, someone reviews 100% of it. The economics of that are not obviously favorable.
What I’d want to see measured
The debate about entry-level finance jobs is real and sparked for good reason, but it’s being conducted with the wrong metrics. Task coverage tells you almost nothing about deployment value. The numbers that matter:
- Verification cost per artifact — how long does a human need to trust a generated comp set versus build one
- Error localization — does the system flag where it’s uncertain, or does it produce uniformly confident output
- Failure mode distribution — are mistakes obvious and loud, or plausible and quiet
- Institutional context retention — does it know this bank’s methodology, or a generic one
Quiet plausible errors in a valuation model are the expensive kind. A junior analyst who is unsure says so, badly, at 3 a.m. A language model asked for a terminal growth rate produces a number.
The honest read
OpenAI has productized the parts of investment banking that were already closest to being a solved information-retrieval-and-formatting problem, and it did the sensible thing by paying practitioners to encode the conventions. That’s good engineering and a clear-eyed product decision.
What it isn’t is a replacement event. It’s a restructuring of where human attention gets spent, and the honest open question is whether banks redesign their apprenticeship model deliberately or just discover in five years that nobody in the associate class can tell when the machine is wrong.
🕒 Published: