October 21, 2025
Investment banking has always defended the analyst grind as an apprenticeship. The long nights, the endless model revisions, the formatting rules, the senior banker comments that become comments on comments: all of it was supposed to teach judgment through repetition.
Project Mercury challenges that bargain directly. OpenAI has more than 100 former investment bankers helping train artificial intelligence to build financial models. The effort includes people who previously worked at firms such as JPMorgan, Morgan Stanley and Goldman Sachs, and participants are being paid $150 per hour to write prompts and construct models for transaction types including restructurings and initial public offerings.
That is not a vague experiment in workplace productivity. It is a targeted attempt to automate the exact work that has historically filled the analyst seat.
The Grunt Work Was the System
Junior bankers often spend more than 80 hours a week at their desks during live deals. Much of that time goes into Excel models for mergers, leveraged buyouts and other transactions, plus the steady churn of PowerPoint revisions requested by senior bankers.
Banking culture has joked about this for years. The “pls fix” meme exists because the work can feel absurdly iterative: move this box, change that number, update that footnote, then redo it again after the next round of comments.
But the repetition was never meaningless. It trained analysts to understand how financial statements connect, how valuation assumptions flow through a model and how small changes can alter the story a company tells buyers, sellers, lenders or public investors.
The uncomfortable question is whether banks still need humans to learn that way if a model can be trained directly on the output.
OpenAI Is Not Training on Theory
The important detail is that Project Mercury is using people who know the conventions of the job. Participants are asked to create models in Excel and follow industry norms, including formatting details such as margin sizes and italicized percentages.
That matters. Banking work is not only about calculating numbers. It is also about presenting information in a form that senior bankers, clients and committees recognize as professional. A model that gets the math right but ignores the expected structure still fails in practice.
The process described for the project is also revealing. Contractors write prompts in simple terms, execute the model, receive reviewer feedback and fix issues before the work is used in OpenAI’s systems. They are expected to submit one model per week.
In other words, the human banker is not merely doing the work. The banker is translating the work into a repeatable training process.
The Banker's Edge Moves Upstream
This does not mean every junior banker disappears overnight. It does mean the value of the seat changes.
If AI can increasingly handle first drafts of models, routine transaction templates and mechanical presentation changes, then the remaining human edge shifts toward asking better questions, spotting bad assumptions and understanding what the output means. The analyst who only knows how to follow instructions becomes easier to replace. The analyst who can challenge the instruction becomes more valuable.
That distinction is critical. A model is not judgment. A spreadsheet can be internally consistent and still be strategically wrong. A deck can be clean and still tell the wrong story. Automation may reduce the hours spent producing materials, but it does not remove the need to decide what matters.
Still, the profession should be honest about what is happening. The junior role has long relied on a large volume of manual production. If that production is compressed, banks will have to rethink how analysts are trained. You cannot remove the reps and assume the judgment appears anyway.
What This Means for Future Analysts
For students and candidates, the takeaway is not to ignore technical skills because AI is coming. It is the opposite. The people training these systems are former bankers because the systems need expert examples.
That means financial statement knowledge, modeling logic and deal mechanics still matter. The difference is that memorized steps are less defensible as an edge. Candidates should be able to explain why a model is structured a certain way, what a line item implies and how a change in assumptions affects the transaction.
The safest skill set is not simply “I can build the model.” It is “I understand the model well enough to direct it, audit it and use it to make a decision.”
Project Mercury points to a future where the analyst class may be smaller, faster and more dependent on judgment earlier. That future may remove some drudgery. It may also remove the old path by which many bankers learned the business.
The machine can be trained on the work. The harder question is whether the next generation of bankers can still be trained without doing as much of it.