OpenAI announced a research collaboration with Ironclad to train and evaluate AI agents on multi-step contracting and procurement work. The project turns real-world legal and commercial workflows into structured research tasks so models can practice configuring agreements, approvals and reusable legal terms inside hosted product environments.
Turning contracting workflows into research tasks
OpenAI and Ironclad identified 11 representative tasks across legal, commercial and procurement work, including setting up nondisclosure agreements, creating procurement approval processes and updating reusable clauses to reflect jurisdiction choices. The teams evaluated each task against between 8 and 50 rubric criteria depending on complexity and estimated that an experienced user would take about 30 to 40 minutes per task on average.
Ironclad provided hosted instances of its software where models could execute tasks. OpenAI researchers generated synthetic training variations from public contracts in the SEC’s EDGAR database (after applying filters intended to remove personal information) and used reinforcement learning with practice and feedback to improve model behavior. OpenAI said it did not use OpenAI customer data, OpenAI’s internal contracts, or nonpublic Ironclad customer data for training or evaluation.
Performance results
On the 11 research tasks, OpenAI reported that GPT‑6 Astra, run with a Max reasoning setting, achieved a mean rubric score of 55.0% compared with 41.6% for GPT‑5.6 Sol run at High reasoning. Estimated average time per attempt fell from 37.0 minutes for GPT‑5.6 Sol to 19.2 minutes for GPT‑6 Astra. An internal model used during Astra’s development scored 63.7% on the same tasks, OpenAI added. In example clips cited by OpenAI, Astra met roughly 94% of task criteria in an estimated 20 minutes versus about 85% in 32 minutes for GPT‑5.6 Sol. OpenAI noted that the time figures are simulated estimates based on assumed model processing and generation speeds, not measured customer time savings.
Implications and partnership request
OpenAI described Ironclad’s role as defining success criteria and bringing customer-facing exceptions and controls into model research. Sunita Verma, Chief Technology Officer at Ironclad, said: “For AI to be genuinely useful in complex areas of contracting, it has to do more than complete individual actions. Agents need to understand the full contracting lifecycle, including how business workflows connect while preserving the controls teams rely on.”
OpenAI is inviting a small number of software companies to propose concrete tasks for collaboration. Applicants should provide a clear example of what an agent must do, evidence of current failures, success criteria, people who deeply know the work, a secure testing environment and data that can be safely used for research.
Original source: OpenAI News