A free guide by New Money School
Everyone has spent two years talking about AI coming for software engineers. On September 10, 2026, OpenAI shipped a product built specifically to do the work of entry-level finance.
Here is the full breakdown: what this product actually is, the exact tasks it performs, which roles sit in front of it, and the part of this that almost nobody is talking about.
Straight from the source: read OpenAI's original announcement here. Everything below is built on what they published, so you can check any claim in this guide against it yourself.
What This Product Actually Is
It is called ChatGPT for Financial Services. Strip away the branding and it is three things bolted together.
One: a model tuned for financial documents. GPT-6 Astra, OpenAI's newest model, reads filings, tables, footnotes, and transcripts, then reasons over them and produces finished work.
Two: licensed financial data, built in. This is the part that matters most and gets the least attention. Datasets from Daloopa, PitchBook, LSEG News, and Crunchbase ship inside the product, covering earnings transcripts, financial statements, company fundamentals, and private company data. No contracts to negotiate, no connectors to configure. OpenAI hosts and indexes it all.
If your firm already pays for data, OpenAI is building shared sign-in with S&P Capital IQ, LSEG, MSCI, Dow Jones Factiva, and Moody's so the tool automatically reaches what you are already entitled to. Its connector ecosystem now runs past 50, including Datasite, Box, Preqin, and Intapp.
Three: firm-formatted output. Administrators upload the firm's Excel, Word, and PowerPoint templates. The AI then produces models, research notes, and pitchbooks in house style, not generic AI formatting.
That combination is the whole story. A chatbot with no licensed data is a confident intern. A model wired into Daloopa and PitchBook, outputting in your firm's deck template, is a different animal entirely.
The Exact Work It Does
OpenAI is not vague about this. The listed use cases are:
- Valuation analysis
- LBO modelling
- Buyer screening
- Earnings analysis
- Pitchbook preparation
Read that list again. That is not "AI assistance." That is a first-year analyst's job description, itemized.
The workflow it replaces is specific: pulling company and market data, reconciling adjustments in Excel, checking assumptions against source filings, then moving the whole thing into PowerPoint and formatting it. Work that ate entire nights. OpenAI says it now happens in minutes.
There is one more feature worth understanding, because it is what makes the output usable rather than risky. Every figure traces back to the exact table or passage it came from, with the supporting text highlighted. OpenAI's own example is a banker inspecting the reconciliation behind an adjusted EBITDA to see which costs were excluded. The AI assembles the evidence. A human still makes the call.
Which Roles Are Actually In Front Of This
The honest answer is not "everyone in finance." It is anyone whose day is mostly producing analysis artifacts from data. That includes:
- Investment banking analysts and associates
- Equity research associates
- Private equity analysts, especially on screening and LBO work
- Venture capital analysts doing sourcing and private company research
- Corporate development and M&A analysts
- FP&A and corporate financial analysts
- Credit and commercial banking analysts
- Valuation and transaction advisory associates, including Big 4 deal teams
- Hedge fund and asset management research associates
- Capital markets analysts who build pitch materials
- Wealth management associates preparing client decks
The roles least exposed are the ones built on relationships, live judgment, and accountability: coverage bankers who own the client, traders, portfolio managers, compliance officers, and advisers whose clients call them personally in a drawdown.
Notice the pattern. The dividing line is not seniority and it is not finance. It is whether your value comes from producing the artifact or from judging it.
About The Design Partners
The product was shaped by design partnerships with Morgan Stanley and Evercore. OpenAI says their teams identified the biggest pain points and steered where it started, which is investment banking and equity research.
Both firms framed it as augmentation. Morgan Stanley's statement is about frontier research helping "our people do the work that matters for our clients." Evercore's talks about deepening insight while keeping the judgment and standards clients expect. Neither firm announced anything about headcount, and neither said this is meant to replace staff.
Hold both facts at once. Two firms that employ thousands of analysts helped design a tool that performs core analyst tasks in minutes, and both describe it as making their people better rather than fewer. Which of those becomes true is a decision those firms will make later, quietly, through hiring plans rather than press releases. Watch analyst class sizes over the next two recruiting cycles. That number will tell you more than any statement.
The Part Nobody Is Talking About
Here is what actually deserves the anxiety, and it is not "a robot took my spreadsheet."
The grunt work was never just grunt work. Building the model by hand is how an analyst learns which assumptions break. Reconciling adjustments is how you develop a nose for an EBITDA that has been massaged. Formatting a hundred pitchbooks is how you learn what a persuasive one looks like. That drudgery was the apprenticeship. It was the mechanism that turned a 22-year-old into someone with judgment by 30.
Automate the apprenticeship and the entry rung does not just get harder to reach. The path from junior to senior loses the thing that produced seniors in the first place.
Nobody in the industry has an answer for this yet. Firms need experienced people in ten years, and the training system that produced them is exactly what this product makes economically pointless. That is a much bigger problem than any individual job posting, and it is why "just be the one directing the AI" is easier advice to give than to follow when you have never done the work yourself.
What This Cannot Do Yet
Keep the pressure calibrated with two facts.
OpenAI's own benchmark, OfficeQA Pro, tests finding and analyzing information across US Treasury Bulletins including complex tables and footnotes. GPT-6 Astra scored 69.9%, up from 60.2% on the previous model. That is a serious jump. It is also roughly three misses in ten on precisely the document work this is built for. In a business where a wrong number in a pitchbook is a career event, that is why citations exist and why a human still signs off.
It is also enterprise only, sold to eligible financial institutions through a sales conversation. OpenAI has not disclosed pricing, seat minimums, or eligibility criteria. Its claim of being roughly twice as cost efficient as alternatives has no published evidence behind it yet.
So this is not landing everywhere next quarter. It is landing at large firms first, unevenly, with humans checking output. That is your window, not your reprieve.
What To Do With This Week
You almost certainly cannot buy this product. You can absolutely copy what makes it work.
- Audit your own week honestly. Write down what you actually did for five days. Mark each item "produced an artifact" or "made a judgment call." That ratio is your real exposure number, and it beats any think piece.
- Feed AI real sources, never memory. The entire premise here is model plus licensed data. Upload the actual filing, the actual export, the actual deck. Asking a model to recall figures is the amateur move.
- Demand citations on everything. Add "cite the specific page, table, or passage behind every figure and quote the line" to any research prompt. You just rebuilt the feature banks are paying for.
- Build the templates once. Save your recurring formats as reusable templates and have AI populate them. That is the firm template feature, and it costs you nothing.
- Practice being the checker. If AI drafts the analysis, your value is catching what is wrong with it. Make a habit of asking what a skeptical managing director would attack first, then answering it before they can.
- Learn one layer up. The safest work is the work that assigns and validates, not the work that produces. Get closer to the client, the decision, and the accountability, on purpose.
The Honest Summary
OpenAI bundled a strong document-reading model with licensed financial data and firm formatting, and pointed it directly at the tasks that fill an analyst's first two years. It works well enough to matter and not well enough to trust unsupervised.
The assembly work is going. The judgment work is not. The uncomfortable part is that finance built its entire training pipeline on the assembly work, and nobody has said yet what replaces it.