Evidence at a glance
Contract Review Is More Than Reading the Clauses
On October 15, 2026, OpenAI introduced Agreement Analyzer, an AI tool it developed with Sullivan & Cromwell (S&C) to help lawyers review draft agreements and identify potential issues. The tool is aimed at more than proofreading. Lawyers must interpret clauses in their legal and transactional context, compare them with relevant precedents, account for firm experience and client priorities, and notice how different parts of an agreement affect one another.
That is why technical leaders should read this kind of case closely. Finding a passage in a document is not the same as supporting a legal judgment that someone can act on. Agreement Analyzer is designed to put precedents, negotiating positions, and a draft into one workflow, making the model’s output an input to a lawyer’s work rather than the final answer. The shift is not that AI takes over contracts. It is that institutional knowledge is being organized into an operational review process.
The Workflow Turns Firm Knowledge into Review Inputs
After submitting a draft, a lawyer can add selected precedents, a negotiating playbook that records firm and client positions, and information about market practice. The tool combines S&C’s workflows, curated precedents, negotiating playbooks, and ChatGPT Enterprise, then conducts multiple rounds of review from different perspectives. It compares corresponding provisions across precedents, tracks interactions between clauses, and organizes its analysis into a memorandum and proposed tracked changes for a lawyer to inspect, adapt, or reject.
This structure means the tool’s value depends on more than the underlying model. Whether a clause is acceptable may vary with the client’s objectives, the transaction, and the precedents that actually apply. Precedents are not generic background material that can be added without selection, and a negotiating playbook is not a rule that can be applied without regard to the matter. For a technical team, the important questions are whether the inputs fit the review at hand and whether each finding can be clearly connected to the material supporting it.
Traceability Says More About Trust Than Agent Count
S&C partner John Savva emphasizes that a claim about market practice should come with a visible source. The public account also describes an adversarial review process in which some agents propose conclusions and others challenge them. Results distinguish findings that remain supported after challenge, findings that are dismissed, and unresolved issues retained with qualifications. Lawyers can inspect the explanations and relevant precedents, then decide which findings should make it into the final markup.
That is a useful output design because it avoids presenting every alert as equally certain. But agents challenging one another do not automatically provide independent fact-checking, and they cannot replace judgment about the quality of the evidence. The material describes parallel checks for market deviations, risk allocation, clause interactions, and ambiguities that could affect litigation as potential use cases. It does not establish that all of these checks are already active in routine reviews. Nor does the public account explain where market-practice information comes from, how representative it is, or how the system handles precedents that do not fit the transaction at hand.
Moving from a Solo Prototype to Shared Use Tests the Workflow
Agreement Analyzer did not begin as a complete firm application. Savva first built a single-user proof of concept with Codex. He then worked with OpenAI to improve output quality and parallel processing and expand the tool for use by multiple lawyers. The timeline shows a move from personal experiment toward an internal firm tool: the application was demonstrated at OpenAI’s legal industry summit on September 17, 2026, S&C announced the collaboration the following day, and OpenAI published its account on October 15.
That path points to a question often obscured by discussions of model capability: a working prototype does not mean that a review process is reliable inside an organization. Shared use also requires the tool to fit existing work practices so that different lawyers can understand the inputs, trace the reasoning, and handle proposed changes. Public materials describe the tool as being for internal S&C use and mention plans to expand deployment, but do not say that it is sold to other firms or used by them. “Available to multiple lawyers” should not be mistaken for validation across a broader market.
Business Benefits and Governance Still Need Evidence
The available material explains how the tool works but provides no figures for review accuracy, time saved, number of users, contract volume, cost, or return on investment. It also offers no independent comparison with lawyers reviewing the same material. The account therefore cannot establish whether the tool is faster or more accurate, or whether it has improved client outcomes. S&C’s stated aim is to help lawyers conduct deeper analysis and focus on the work that requires their judgment. That is a business proposition, not a conclusion validated by published performance data.
Teams considering a similar system can start by turning its claims into outcomes they can check: whether lawyers can trace every finding to its source, which issues are missed or incorrectly flagged, whether review time changes, and how often lawyers adopt, revise, or reject suggestions. They should also establish how client contracts are retained, who can access them, how they are isolated, and where market information comes from and when it applies. The public account does not disclose those governance details. Agreement Analyzer is a design example for embedding institutional knowledge in review, but lawyers still make the judgments and trade-offs, and any real benefit must be demonstrated through evaluation.