How Sullivan & Cromwell's Agreement Analyzer, Built With OpenAI, Fits Into Its AI Strategy

Sullivan & Cromwell's OpenAI-Built Tool Shows "Proprietary" Now Means the Prompts, Not the Model

Sullivan & Cromwell, an Am Law 100 firm, announced on Sept. 17, 2026 that it had built a bespoke Agreement Analyzer with OpenAI, Law.com's Benjamin Joyner reported on Sept. 25. The tool produces memoranda that list issues in a proposed agreement, give the rationale for each issue, and offer proposed treatments to address them, drawing on the firm's own review process and precedent set. The firm frames it as one piece of a broader push to build proprietary tools alongside its OpenAI relationship. No pricing, volume, or accuracy figures were disclosed in the announcement or the report.

Is this really different from off-the-shelf OpenAI contract tools?

Mostly in packaging: Sullivan & Cromwell describes its edge as its own precedent data and review methodology layered on OpenAI's models, not a new engine.

That distinction matters commercially but not technically. A firm that builds "with OpenAI" is still routing client agreement language through a third-party foundation model; its own contribution is prompt design, precedent curation, and workflow. Buyers evaluating any BigLaw-branded AI tool should ask what, specifically, is proprietary — the model, the training data, or just the interface wrapped around a general-purpose API.

What happens to confidential deal language once it reaches OpenAI's models?

Neither the firm's announcement nor Law.com's report says, which is the gap buyers should press on before trusting any AI-drafted issue memo.

This site has already flagged matter economics as the live pressure point pushing firms toward generative tools. A tool co-built with a foundation-model vendor raises the same question e-discovery buyers already ask their own vendors: does the provider train on client data, is the deployment single-tenant, and who holds the keys. None of that is addressed in the coverage of the Analyzer, and it's the first thing worth asking Sullivan & Cromwell or any firm citing this as a model.

Can you check the Analyzer's issues list against the actual contract language?

Unclear — the reported output describes issue lists, rationale, and proposed fixes, with no mention of citations back to specific clauses.

For a tool making legal judgment calls, that's the detail worth chasing. Platforms built for defensibility tie every AI-generated statement to a source passage so a lawyer can verify it in seconds rather than re-reading the whole document. If the Agreement Analyzer's memoranda summarize risk without pointing back to the clause that created it, associates end up re-reading the agreement anyway — which undercuts the efficiency case before anyone has measured it.

Should other firms read this as a signal to build their own tools?

Only if they have the precedent data and review process to differentiate on — otherwise they are paying to rebuild what vendors already sell.

Sullivan & Cromwell's framing — proprietary tools built on what Law.com calls an "ongoing partnership" with OpenAI — fits a pattern this site has tracked among Am Law firms getting serious about matter economics. But bespoke builds carry maintenance cost and vendor dependency that a licensed platform doesn't. Buyers should weigh whether a firm-branded tool is actually more defensible than a purpose-built one, or just harder to audit because fewer outside eyes have tested it.

Frequently asked questions

Did Sullivan & Cromwell disclose how accurate the Analyzer is?

No. The Sept. 17 announcement and Law.com's report describe the tool's outputs — issue memoranda, rationale, and proposed treatments — but include no accuracy, error-rate, or adoption figures.

Is the Agreement Analyzer an e-discovery tool?

No. It reviews and drafts transactional agreements rather than processing litigation evidence, though the data-handling questions it raises match the ones e-discovery buyers already put to their vendors.

Source: Law.com, reporting by Benjamin Joyner, Sept. 25, 2026.