OpenAI and Ironclad turn contract workflows into agent evals

OpenAI and Ironclad published a case study on using production contract workflows as both training data and evaluation benchmarks for computer-use agents. Ironclad's platform provides stateful multi-step workflows with approval chains, redlining and conditional branching, with customer data replaced by synthetic contracts.

• Ironclad opened its workflow engine as an agent training and eval ground
• Real approval chains, redlining and conditional logic form the basis
• Customer contracts are replaced with synthetic, sanitized data
• Evals are versioned to survive product UI updates