Hebbia is the research tool bankers use to go through the documents in a deal. When it tried OpenAI's GPT-6 Astra, the model led the other models it tested at two jobs: choosing the documents that back up an answer, and building slides that follow the banker's brief. Of the presentation requirements that applied, it met 97.9%.
The deck arrives close to finished
Hebbia already put source footnotes on its slides in June, and it already had a separate agent that checks every figure in a deck against a source file. So traceability is not what changed. The draft is what changed. In Hebbia's experience, earlier models were good at one of two things: slides that looked polished but ignored the brief, or slides that followed the brief but looked like a machine made them. A banker cannot send either to a partner. A deck in the firm's template, in the structure the banker asked for, with a footnote on each claim, is a draft that can go out once someone has checked it.
Checking becomes the work
The promise is that an overnight deck turns into a few minutes of review. I think the review takes longer than that, and that is the point. Take a deck on a software company with 40 figures in it, each footnoted. Opening a footnote, reading the highlighted passage and comparing it with the slide takes about two minutes, so the checking comes to around 80 minutes. That is an evening instead of a night, and it is the part that matters, because a wrong figure now looks as good as a right one. A number taken from the wrong reporting period appears in a clean slide in the firm's own colours, and nothing about it looks wrong. The tidier the draft, the better a bad figure hides in it.
What the spreadsheet says about the next step
The spreadsheet did more than move work from clerks to accountants. A new scenario took seconds instead of an afternoon, so finance people asked far more questions of the same numbers, and scenario modelling became a line of work of its own.
Deals will go the same way. If a first draft of a deal takes minutes, a team can look at ten deals where it used to look at three, and the person who decides which ones deserve a partner's time becomes worth more. The agents Hebbia imagines, which watch the news in the background and flag a signal hours before competitors, are the same idea: more opportunities for a person to judge.
The people who lose out are the ones whose main task is assembly, and at a bank that is the first-year analyst who builds the slides. Their next job is checking, and it is learned differently, by tracing a number back to where it came from instead of by building the page. Banks that see this will teach it on purpose.
What it means outside banks
A finance director at a hospital group building a board pack, a recruiter writing a hiring report and a school department head writing up exam results all do the same job: turn a pile of documents into a page with numbers on it. GPT-6 Astra has been rolling out to paid ChatGPT plans since September, and OpenAI built it to follow an existing template. A first draft in your own format is within reach of any of them.
So the test for any tool that writes reports is simple. Pick one number in the draft and time how long it takes to reach the page it came from. If the tool cannot take you there in under a minute, it has moved your night instead of shortening it.