Sullivan & Cromwell brings legal expertise into agreement review
- ID
- 36099
- Status
- summarized
- Published
- 15 Oct 2026, 8:00 AM
- Fetched
- 16 Oct 2026, 9:04 AM
- Provider
- OpenAI News
- Category
- ai-labs
- Original URL
- https://openai.com/index/sullivan-cromwell
- Source URL
- https://openai.com/news/rss.xml
Summary
- Score
- 3.5
- Created
- 16 Oct 2026, 9:05 AM
- Tags
- Audience
- developersai_ml_learnerssaas_foundersvibe_coders
What happened
OpenAI published a customer story on Sullivan & Cromwell's Agreement Analyzer, a legal review tool for draft agreements. Partner John Savva built it first as a single-user proof of concept using Codex, then worked with OpenAI to improve output quality, add parallel processing, and make it usable by multiple lawyers at once. The tool combines firm precedents, negotiating playbooks, and ChatGPT Enterprise, runs multiple review cycles from different perspectives, and returns a memorandum plus proposed tracked changes that a lawyer can review, adapt, or reject. The article is a vendor case study with no accuracy numbers, pricing, latency, or independent evaluation, and the text is cut off mid-sentence in the section about making sources and uncertainty visible.
Why it matters
There is nothing decision-grade here: no benchmark, no cost, no accuracy figure, so it should not be used to justify buying or building legal AI. The one transferable detail is the described workflow shape — a domain expert ships a single-user Codex PoC, then productionizes by adding parallel processing, multi-user access, and a multi-perspective review pass that synthesizes into a human-reviewed diff. If you are building a document-review agent for any regulated domain, that PoC-to-parallel-multi-user path is the concrete thing to copy; the ChatGPT Enterprise and firm-precedent layers are specific to this customer and are not available to you.
Discussion angle
Is 'run several review passes from different perspectives, then synthesize and hand the lawyer a tracked-changes diff' a generalizable agent architecture, or does it only work because the firm's own precedents and playbooks supply the grounding? Ask what the group would need to see — accuracy on a fixed contract set, per-document cost, latency — before trusting this pattern in their own domain.