Pitchbooks, Comps, and the Slow Death of the Analyst Grind

OpenAI hired more than 100 ex-bankers to teach a model how to do M&A, LBO, and IPO work. Meanwhile, the tool that was supposed to shrink the junior class at banks — Rogo — has reportedly been creating more work for associates instead.

Both of those things are true at the same time, and that gap is where the real story lives. I’ve reviewed enough AI tools on this site to know the pattern: the launch narrative says replacement, the deployment reality says redistribution. Let’s look at what OpenAI actually shipped and what it probably means for the people whose jobs are on the marketing slide.

What ChatGPT for Financial Services actually does

The 2026 launch is a specialized version of ChatGPT aimed squarely at the labor-intensive end of investment banking. Per the reporting, it can research companies, analyze financial data, and generate the formatted presentations that investment banks run on. The named targets are specific:

• Company research
• Running comps
• Testing valuation scenarios
• Building pitchbooks

If you’ve never worked a banking analyst program, that list looks like a grab bag of tasks. If you have, you’ll recognize it as roughly the entire first two years of the job. The stated ambition floating around the launch is automating about 60% of junior banker tasks.

What makes this different from the generic “use ChatGPT for finance” advice people have been passing around for years is the training investment. Hiring 100+ ex-bankers to generate and grade domain work is expensive, slow, and unglamorous. It’s also exactly the right move if you want a model that produces output a managing director won’t throw back across the table. Banking deliverables have conventions — formatting, footnoting, how a comp set gets defended — that no amount of general reasoning ability will teach you. You learn them from someone who got yelled at for getting them wrong.

Why I’m not calling this a job apocalypse yet

The Rogo detail is the most useful data point in this whole story, and it’s the one getting the least attention. A tool built to shrink the junior class ended up adding to associate workload. That’s not a freak outcome. It’s the default outcome for AI tools that produce plausible work product in a field where being wrong is expensive.

Here’s the mechanism, and I’ve watched it play out with code assistants, research agents, and document tools alike: generation gets cheap, verification does not. When a model can produce a pitchbook in minutes, you don’t get fewer pitchbooks. You get more of them, each requiring a human to check whether the comp set makes sense, whether the valuation assumptions hold, and whether the numbers tie. The person doing that checking is the associate who was supposed to be automated away.

There’s a second-order problem that the reporting flags directly and that I think is the genuinely hard one. Wall Street has used these tasks to train people. Building comps by hand is how you develop intuition for what a business is worth. Assembling a pitchbook is how you learn what a deal thesis looks like. Remove the grunt work and you’ve also removed the apprenticeship. Banks now have to answer a question they’ve never had to answer: how do you produce a competent VP without putting someone through the analyst grind first?

What this means if you’re the one applying

One of the ex-banker recruiters involved in this conversation mentioned getting five student emails a day asking for investment banking advice. Those students are reading headlines that say their target career is being automated. I’d offer a more measured read.

The skill that gets scarce isn’t producing the deliverable. It’s judging whether the deliverable is defensible. That’s a different muscle, and it’s harder to build without the reps that used to come free with the job. If you’re entering finance now, the practical move is to get good at the review layer fast: knowing what a bad comp set looks like, catching an assumption that doesn’t survive scrutiny, understanding why a client would reject a thesis. That’s the work that gets more valuable, not less, when generation costs approach zero.

My honest assessment of the tool itself: the domain training makes it credible, and the task list is well chosen. But the 60% automation figure is an ambition, not a measured result, and the industry’s own recent experience with a similar tool points the other direction. Headcount at banks is a function of deal volume and margin pressure, not tooling capability. If this thing works well, expect the junior class to shrink slowly and the definition of junior work to change fast.

Worth watching, worth testing, not worth panicking over.

🕒 Published: September 11, 2026