What OpenAI's Ropes & Gray diligence build actually automates
OpenAI says the law firm Ropes & Gray is building AI for transactional due diligence. A look at what that work really involves, where models help, and where a human lawyer still has to sign off before any client gets advice.
TL;DR: OpenAI says Ropes & Gray is building AI to speed up transactional due diligence, and the useful detail is structural: the model groups and summarizes documents, but a lawyer still reviews before anyone gets advised.
The primary source here is thin. It’s a single OpenAI customer post, “How Ropes & Gray is developing AI for transactional diligence,” and it describes a law firm building AI tooling for the document grind that precedes a deal. No model names, no eval numbers, no pricing. So I’m not going to pretend it’s a benchmark study. What it is: a clear picture of where generative AI slots into a real, expensive, high-stakes legal workflow. That’s worth looking at carefully, because diligence is one of the better-shaped jobs for this technology, and most of the lessons transfer to any document-heavy operation.
What is transactional diligence, and why is it a good AI target?
When a company buys another company, or raises a big round, or signs a major financing, lawyers read everything first. Contracts, leases, employment agreements, IP assignments, litigation records, corporate filings. The job is to find what could blow up the deal: a change-of-control clause, an unassigned patent, a lawsuit nobody disclosed. It is enormous, repetitive, and judgment-heavy all at once.
That combination is exactly why it attracts AI. The volume is high enough that marginal speed matters a lot. The documents are structured enough that a model can extract and summarize. And the stakes are high enough that a firm will actually pay for tooling instead of eating associate hours. OpenAI’s framing of the Ropes & Gray work, per its own post, is that lawyers “review grouped documents, summaries, and reports before advising clients.” That sentence is doing more work than it looks.
Break it down. Grouped documents: the model clusters related material so a human isn’t hunting across a data room. Summaries: it compresses each cluster into something readable. Reports: it rolls those up into the diligence memo structure lawyers already work from. The human step comes last and stays mandatory. Nobody gets advice straight from the model.
Where does the model actually add value here?
The honest answer is retrieval and compression, not judgment. The hardest part of diligence is not reading a single contract, it’s reading two thousand of them and remembering that clause 14.3 in the Delaware lease contradicts a rep in the purchase agreement. Models are genuinely good at the first-pass surface: pull every change-of-control provision, flag every non-standard indemnity, list every agreement missing a counterparty signature. That is hours of associate time collapsed into a query.
What models are not reliable at, and what the Ropes & Gray structure quietly admits, is deciding whether a flagged item matters. A non-standard indemnity might be fine or might be the whole negotiation. That is a call that depends on the deal, the client’s risk tolerance, and a hundred things not written in the data room. So the sensible design is: let the model do exhaustive extraction, let the lawyer do selective judgment. The post describes exactly that split.
I’d push one caution on anyone reading this as a template. OpenAI’s customer posts are, by nature, favorable. We get the shape of the workflow, not the failure rate. We don’t know how often the summaries miss something, how the firm evals accuracy, or what the review burden looks like in practice. A summary a lawyer still has to fully re-read saves nothing. A summary that lets them skip 80 percent and spot-check the rest saves a lot. The post doesn’t tell us which it is, and that gap is the whole ballgame.
How should an operator copy this pattern?
You don’t need a law firm to steal this. The pattern is: high-volume documents, model does extraction and clustering, human does the judgment at the end. It maps to vendor contract review, compliance audits, insurance claims, grant applications, any pile of semi-structured text where someone currently reads all of it to find the few things that matter.
Three things make or break it. First, grouping quality. If the model clusters documents badly, every downstream summary inherits the mess, and the human spends their saved time untangling it. Build and test the retrieval layer before you fall in love with the summaries. Second, traceability. Every summary line needs to link back to the source passage, or your reviewer can’t verify fast, and an unverifiable summary in a legal context is worse than no summary. Third, an eval you actually run. Take fifty documents a human already diligenced, run your pipeline, measure what it missed. Do that before you trust it on live work, and keep doing it as documents drift.
The catch most people skip: the economics only work if the human review step gets genuinely shorter, not just relocated. It is easy to build a system that produces beautiful summaries that a cautious professional re-reads in full anyway, because they can’t trust it. Then you’ve added a step and saved nothing. The win is in the review time, and you measure the win by timing the human, not admiring the output.
Does this change the business of legal work?
Modestly, and slowly. Diligence has long been the training ground where junior associates learn how deals actually fit together by reading everything. Automate the reading and you change what juniors do and learn. Firms will have to decide whether the model is a leverage tool for the same people or a reason to need fewer of them. The OpenAI post doesn’t touch that question, because customer posts never do, but it’s the real one hanging over every build like this.
My read: the near-term reality is augmentation. A senior lawyer with good tooling covers more ground, catches more, and bills the judgment rather than the reading. That’s a better product for clients and a better job for the lawyer. The displacement worry is real but further out and depends entirely on whether these systems earn trust at the extraction layer, which, again, we can’t verify from a single favorable post.
Practitioner’s take: treat the Ropes & Gray structure as a blueprint, not a result. Build the same three-stage shape, group, summarize, report, with a mandatory human gate at the end, for whatever document pile is eating your team’s hours. Then do the one thing the announcement can’t do for you: measure the human review time before and after. If your reviewer still reads everything, you built a demo. If they verify instead of read, you built leverage. The accuracy of your grouping and the traceability of every summary line are what decide which one you get, so spend your engineering there, not on prettier output.