OpenAI has partnered with legal technology company Ironclad to train and evaluate AI agents on complex, multi-step enterprise software workflows. The collaboration targets advanced computer-use capabilities, teaching models to execute business rules, configure workflow automations, and verify outputs across multi-step administrative tasks in specialized corporate environments.

Using real-world workflows identified by Ironclad domain experts across legal, procurement, and commercial functions, OpenAI created synthetic training environments and applied reinforcement learning. The benchmark suite comprises 11 complex tasks evaluated against up to 50 individual criteria. OpenAI reported that its latest model, GPT-6 Astra, achieved a 32% higher average score on these Ironclad research evaluations compared to GPT-5.6 Sol, while reducing estimated attempt time by 48%.

The partnership represents OpenAI’s broader strategy to refine agent capabilities by collaborating directly with domain-specific software vendors. By providing hosted, sandboxed product environments for reinforcement learning, enterprise partners help frontier labs train models on realistic professional tasks that extend beyond simple web browsing or text generation.

Why it matters

  • Demonstrates a clear methodology for training enterprise AI agents using domain-specific software sandboxes and reinforcement learning.

  • Shows measurable performance gains in agentic computer-use models, reducing execution time while increasing task accuracy in software interfaces.

  • Highlights a strategic playbook for vertical SaaS vendors to partner with frontier labs to embed native automation capabilities.

Source: openai.com