OpenAI and Ironclad evaluate agents for enterprise computer use

What is changing?

OpenAI said it is working with Ironclad to turn contracting workflows into training and evaluation tasks for computer-using agents. The teams identified 11 tasks across legal, commercial, and procurement work, including setting up nondisclosure agreements, building procurement approval flows, and updating a reusable clause according to the selected jurisdiction. Each task was evaluated against 8 to 50 criteria depending on complexity, while Ironclad provided hosted product environments in which the models could operate.

In OpenAI’s research evaluation, GPT‑6 Astra achieved an average score of 55.0%, compared with 41.6% for GPT‑5.6 Sol. Estimated time per attempt fell from 37.0 minutes to 19.2 minutes. These are simulated estimates rather than measured customer time savings and cover only 11 tasks. For production teams, the method highlights the importance of evaluating agents against end-to-end rules, exceptions, and verifiable outcomes instead of individual clicks.