OpenAI published a customer story about Sullivan & Cromwell on October 15, 2026, and the subject is a tool rather than a model. Agreement Analyzer is a contract review workflow the New York law firm built with Codex and runs on ChatGPT Enterprise: a lawyer loads a draft agreement, adds the precedents and the negotiating playbook that fit the deal, and the tool returns a memorandum with proposed tracked changes to review, adapt or reject.
What was actually released
No model, no product and no launch. The page is a customer story OpenAI published on its own site, and what it describes is an internal build at one firm, named in public for the first time: Agreement Analyzer, developed by Sullivan & Cromwell in collaboration with OpenAI.
The workflow is specific enough to follow. A lawyer starts from a draft agreement, then attaches selected precedents, a playbook carrying the firm’s own positions and the client’s, and material on market practice. The tool runs several review cycles from different perspectives, pulls the results together, identifies potential issues, and writes up a memorandum and proposed tracked changes.
John Savva, a partner at the firm, built the first version as a single-user proof of concept with Codex. He then worked with OpenAI on the quality of the output, on parallel processing, and on making the tool usable by several lawyers at once. That is the only engineering detail the page offers about the build itself.
The part worth reading twice is how the tool handles doubt. Specialized agents review the same agreement in parallel, on market deviations and risk allocation, on how clauses interact, and on ambiguities that could matter if the deal ended in litigation. Another set of agents challenges what the first set found. What comes back separates findings that survived the challenge, findings that were dismissed, and issues that stayed unverified and are kept with qualifications. A lawyer can open any issue, read the explanation and see the precedents behind it before deciding what goes into the final markup.
What a buyer can take from this
- Provenance and self-doubt sit inside the workflow rather than bolted on afterwards: each finding carries its source and its status.
- The build is a general coding agent plus the enterprise chat product, not a vertical legal tool. That is an architecture choice any team can copy, and the page puts no price on it.
What the claim rests on
Every statement on the page belongs to Sullivan & Cromwell or to OpenAI, and it reads as a sales document: a named customer, two executive quotes, a described workflow, and no measurement of any kind.
There is no count of lawyers using the tool, no number of agreements reviewed, no time saved, no accuracy figure and no independent check. The page does state that the customer is an enterprise account in North America in legal services, and it names ChatGPT as the product involved, but none of that is a result.
The absence is easier to see next to the rest of OpenAI’s October stream. We have read and published the other customer stories in it, and five of the six lead with a number. This one leads with none.
| Story, as published | The figure it leads with | What the figure measures |
|---|---|---|
| Oracle, October 8 | 130,000 ChatGPT users, 95,000 on Codex | adoption inside one company |
| Pollo AI, October 8 | more than 26 million users | the audience of a platform, not the models |
| Sophos, October 9 | 38 minutes cut to 89 seconds | one workflow, as reported by the customer |
| Block, October 14 | 18 minutes to a proposed patch | a red team exercise that stops at a proposal |
| Amp, October 14 | 10 billion tokens a week | spend on OpenAI models, not outcomes |
| Sullivan & Cromwell, October 15 | no figure at all | an architecture with no published result |
What is not published
No price, no seat count, no number of lawyers and no volume of agreements reviewed. The page gives no rollout date, no accuracy rate, and no measure of how often a lawyer kept a proposed change. ChatGPT Enterprise is named as the runtime, and GPT-6 Astra appears inside a quote about reasoning quality, without saying which model does the work in production. What the build cost is missing too, and that is the figure a firm reading this page wants most.
On this page
What it changes for a team weighing build versus buy
Two ideas here transfer to a team that will never hire a lawyer or build a contract tool. The first is that a finding needs a source attached. Savva’s test for the tool is the test a buyer should apply to any assistant that cites market practice or precedent: the claim is only usable if you can open what it rests on. An assistant that cannot show the source behind a proposed change is an assistant someone else has to redo.
The second is that uncertainty is a field, not a defect. Findings arrive as kept, dismissed or unverified, and the lawyer sees which is which. A team comparing vertical legal products can ask for exactly that: how the tool separates what it knows from what it guessed, and whether that distinction reaches the reviewer or disappears into a summary.
The build path is the third thing, and it cuts both ways. A firm took a general coding agent and the enterprise chat product, then built the tool around its own precedents instead of renting a vertical product. That is the cheapest-looking line in the story, and the page publishes no cost for either path, so the comparison cannot be made from here.
The Neural Minute take
The figure to keep in this story is that there is none. Six customer stories in eight days, five of them wearing a number, and the one about the law firm arrives with a workflow and nothing to count. Read it for the pattern, not the proof.
If an AI tool makes a claim about market practice, you need to be able to see the source of that information.
That line is the useful part, and it is a specification rather than a result. It applies to the assistant a solo consultant pays for as much as to a tool a firm builds: a claim about a market, a precedent or a policy is worth nothing without the passage it came from. The findings that carry a status, kept, dismissed or unverified, are the same idea turned into a screen.
Our reservation is about the stream rather than this page. This is the sixth OpenAI customer story we have read and published since October 8, after Oracle and Pollo AI on October 8, Sophos on October 9, and Amp and Block on October 14. A story with a number can be argued with; a story with an architecture can only be copied. For a reader choosing a tool, the copyable part is the standard, and the standard here is that a draft only moves when a person can see why.
Ask any vendor for the passage behind a claim, and mark it unverified if it cannot produce one.
Sources
- Sullivan & Cromwell brings legal expertise into agreement review, OpenAI, published October 15, 2026: https://openai.com/index/sullivan-cromwell/
- The Agreement Analyzer description, the Savva quotes, the parallel review and challenge pattern, the finding statuses, the account metadata and the mention of GPT-6 Astra are read on that page on October 16, 2026, as the firm’s and OpenAI’s own statements, not as independent measurements
- Our earlier readings of the same stream: Oracle, Sophos and Block