On October 6, 2026, OpenAI announced a research collaboration with AI contracting platform Ironclad to push computer use beyond isolated clicks. The teams are training and evaluating whether an agent can understand business rules, configure specialized software across multiple steps, and verify that the finished workflow still satisfies the original requirements.
Researchers worked with Ironclad employees and OpenAI users to select 11 legal, commercial, and procurement tasks. They included setting up nondisclosure agreements, creating procurement approval processes, and updating a reusable legal clause based on a requester’s jurisdiction. OpenAI estimates that an experienced user would take about 30 to 40 minutes per task. Each task was graded against 8 to 50 criteria depending on its complexity, making failures in particular rules or stages visible.
Ironclad provided hosted software environments, while the researchers created synthetic training tasks from public contract data and used reinforcement learning to improve the models through practice and feedback. OpenAI says the tasks were derived from public SEC EDGAR contracts filtered to remove personal information. OpenAI customer data, OpenAI internal contracts, and nonpublic Ironclad customer contracts were not used for training or evaluation.
Across the 11 research tasks, GPT-6 Astra achieved a mean rubric score of 55.0%, compared with 41.6% for GPT-5.6 Sol. OpenAI also reports that estimated time per attempt fell from 37.0 minutes to 19.2 minutes. In two demonstrations, Astra met about 94% of the criteria in an estimated 20 minutes, while Sol met about 85% in 32 minutes. OpenAI notes that these times are simulated estimates based on model processing and generation speeds, not measured customer time savings, and that the results cover only the 11 research tasks.
The hard part is handling exceptions and global rules. A contracting process may require Finance approval above a threshold, Security review for one request type, and Legal review for nonstandard terms. An agent can complete individual settings and still create a workflow that fails at the boundaries. OpenAI therefore defines success through checkable criteria rather than a screen that merely looks complete; Ironclad also stresses that human oversight and platform controls remain important.
For computer-use agents, the collaboration points to a combined design: training data, expert-defined rubrics, resettable software environments, and end-to-end verification. A vendor benchmark is evidence about a research direction, not a legal opinion, a promise of customer time savings, or a guarantee across every contracting platform. Production deployments still need permissions, escalation paths, audit history, and human approval.
