According to an OpenAI announcement dated October 15, 2026, law firm Sullivan & Cromwell developed an AI tool called Agreement Analyzer in collaboration with OpenAI to identify issues in draft agreements. Firm partner John Savva first built the tool as a single-user proof of concept using Codex, then worked with OpenAI to improve output quality, add parallel processing and turn it into a multi-user application. The tool integrates the firm's workflows, curated precedents, negotiating playbooks and ChatGPT Enterprise. A lawyer inputs a draft agreement and can add selected precedents and a playbook covering firm and client positions and market practice. The tool runs multiple review cycles from different perspectives, synthesizes the results, and returns a memorandum and proposed tracked changes for a lawyer to review, adapt or reject. Savva describes an adversarial process in which some agents develop conclusions and others challenge them; results distinguish findings supported after challenge, dismissed findings, and unverified issues retained with qualifications. Savva emphasized source transparency, saying that when a tool makes a claim about market practice, "you need to be able to see the source of that information." He credited reasoning improvements in GPT-6 Astra with moving the work forward. This account comes from OpenAI, which is promoting its own products; no independent evaluation of the tool's performance was provided.
- Announced by OpenAI on October 15, 2026
- Agreement Analyzer was built by a S&C partner using Codex, then developed with OpenAI into a multi-user tool
- It integrates firm precedents, negotiating playbooks and ChatGPT Enterprise, returning a memo and tracked changes for lawyer review
- Uses an adversarial agent process; lawyers decide which findings warrant changes
What it means for you
A large law firm turned its own review process and precedent library into an AI tool, rather than buying a generic contract-review product. The lesson for operators is in the sequence: a lawyer who already understood the review workflow built the first version, and the firm kept a person on every final decision — the tool proposes changes, lawyers accept or reject. The output is only trustworthy because it shows its sources and flags what it couldn't verify.
Try this
If your team reviews recurring documents — contracts, applications, claims — write down the actual steps one reviewer takes and what they check against. That map, not a tool, is what makes anything automatable later.
Who should care
Operations and knowledge leads at firms that do high-volume document review against internal standards or precedent — legal, insurance, procurement, compliance.
Skip this if
You don't have a documented review process with reusable reference material, or document review isn't a meaningful part of your work. This is a case study of a bespoke build at a major firm, not a product you can buy.
Sources: OpenAI — read the original