On October 29, 2026 OpenAI published a customer story on Ropes & Gray's proprietary transactional diligence tool, built with OpenAI. It follows the September 17 launch of Astra for Law, a legal configuration of GPT-6 Astra offered to selected firms through a Trusted Access program, where Ropes & Gray was named among firms whose workflows OpenAI's forward-deployed engineers helped build.
The newsroom flagged that the story alone lacked detail on workflow, scope, controls and results. The question for builders is how much of this is a working, measured system versus a described design.
What we checked
- What specific diligence workflow, models, controls, deployment scope, and measured outcomes does Ropes & Gray describe publicly? partly confirmed. Workflow (data-room inventory, theme grouping, summaries, firm-format report), scope (material commercial contracts), controls (lawyer review with traceable sources; Trusted Access, ZDR) and the build method (compare to past human reports) are public. The exact model binding is only implied via GPT-6 Astra, and no measured outcomes are published; hours-not-weeks is stated as an ambition. [1][2][3]
What we found
- The workflow is a data-room first pass. Per OpenAI's October 29 case study, the tool reads thousands of pages, picks out relevant documents, groups them by theme and issue, inventories and categorizes the whole data room so relationships are reviewed rather than isolated files, and drafts summaries plus a report in the firm's own format. [1]
- Scope is narrow so far. Initial development covered material commercial contracts only, and the evaluation method described is qualitative: lawyers compared test outputs against human-written diligence reports from completed deals and then refined the rubric and report format. No accuracy, recall or time-saved figures are published. [1]
- The headline speed claim is an ambition, not a result. OpenAI's page says the firm has publicly described an ambition of producing an issue-level diligence report in hours rather than weeks, which is a goal statement rather than a measured outcome. [1]
- The control model is human review with traceability. A Ropes & Gray partner says lawyers still review everything and the tool is built so they can check interim summaries and the underlying documents; OpenAI's Astra for Law launch adds that the system helps lawyers trace findings back to the source, for example whether key customer contracts require notice or consent. [1][2]
- Model and deployment context come from the Astra for Law launch on September 17, 2026. Access is through a Trusted Access program for selected firms in ChatGPT and Codex, with API access coming; eligible firms get Zero Data Retention on the API and ChatGPT Enterprise usage excluded from human review by default. The case study names GPT-6 Astra alongside the proprietary tool but does not state exactly which model or interface the diligence tool runs on. [2][1]
- The strategic motive is differentiation. Law.com reports the firm wants a sustainable competitive edge by encoding its own institutional review process into proprietary tools rather than relying only on off-the-shelf legal AI. [3]
What it means
For builders of document-heavy agent workflows, the pattern is clear and copyable: index the whole corpus first, cluster by issue, generate a structured report in the customer's house format, and make every finding traceable to its source so a human can verify it. The evaluation loop described, comparing drafts to past human reports and iterating the rubric, is the practical way firms are building trust in these tools.
Treat the speed claim as a target. Nothing public yet shows accuracy, recall on missed issues, or real time saved, and the scope is limited to commercial contracts. Vendors selling into law firms should expect buyers to ask for exactly those numbers, and firms building in-house should plan to measure them before claiming them.
Competing readings
- A real but early, narrowly scoped tool: a reviewed first pass on commercial contracts, still being tuned against past deals, with no measured results yet (likely, about 65%). What would tell it apart: Scope limited to material commercial contracts and tuned against completed-deal reports [1]; Speed is framed as an ambition, not a measured outcome [1].
- Mainly a marketing showcase for Astra for Law, with the firm's tool serving as a launch reference more than a production system (unlikely, about 25%). What would tell it apart: The tool was announced as part of OpenAI's Astra for Law launch with Trusted Access gating [2][3].
- A mature production system already delivering hours-not-weeks diligence on live deals (very unlikely, about 10%). What would tell it apart: Only an ambition statement exists for speed; no outcome data [1].
Base rate
The New Runtime archive holds only 0 earlier case-study stories about companies or organisations from the last 12 months that are at least 3 months old: too few (minimum 5) for a base rate. BASE_RATE was not supplied, so the archive has 0 cases and 0 matches; too few cases to draw on, and this analysis does not lean on it.
What would change this view
- Ropes & Gray or OpenAI publishes accuracy, recall or time-saved numbers from live deals (signal): The tool is past the pilot stage and the speed claim is a measured result.
- gpt-6-astra-law becomes generally available on the API and Harvey or Legora ship comparable diligence features (model-listing signal): The firm's edge narrows and the tool is closer to a reference build than a moat.
Still open
- What accuracy or recall did the tool reach against the human-written reports used in testing?
- Which exact model and interface does the tool run on (ChatGPT Enterprise, Codex, or the API)?
- Has the tool been used on live client deals, and at what volume?
- When will scope extend beyond material commercial contracts?