OpenAI and Ironclad Train AI Agents to Handle Complex Contracting Work

OpenAI is teaming up with contract management company Ironclad to train and evaluate AI agents on the kind of complicated workflows that legal, procurement and commercial teams deal with every day.

Announced on October 6, the research collaboration is focused on a problem that goes well beyond asking an AI model to read or summarize a contract. The goal is to make AI agents capable of understanding business rules, navigating specialized software, completing several steps in sequence and checking whether the final result actually meets the original requirements.

For OpenAI, contracting provides a useful test case for a broader ambition: getting AI agents to perform real professional work inside the software businesses already use.

OpenAI Is Turning Contracting Workflows Into AI Training Tasks

The collaboration uses realistic contracting scenarios as research tasks for OpenAI’s models. Ironclad brings expertise in contract lifecycle management, helping define workflows that require more than a single instruction or isolated computer action.

The tasks can involve configuring agreements, setting up approval processes, managing legal terms and handling other pieces of a contracting workflow. An agent must understand what the business wants, determine which steps are required and complete them in the correct order.

That distinction matters. A model might know how to click a button, but professional software often requires it to understand why that button needs to be clicked and what should happen afterward.

GPT-6 Astra Shows a Measurable Improvement on the Tasks

OpenAI evaluated GPT-6 Astra against its predecessor, GPT-5.6 Sol, across 11 contracting research tasks. Astra recorded an average score of 55.0%, compared with 41.6% for GPT-5.6 Sol.

The estimated average time per task also fell from 37.0 minutes to 19.2 minutes. OpenAI describes those figures as simulated estimates based on assumed model processing and generation speeds, rather than measured customer time savings.

The numbers are interesting, but they should not be mistaken for proof that AI can now independently run a legal department. The evaluation measures performance on a defined research set, not every contracting workflow used by customers.

Ironclad Helps Define What Good AI Performance Looks Like

Ironclad’s role goes beyond providing software. Its employees helped identify difficult, high-value contracting tasks and define what a successful outcome should look like.

Each task can require an agent to satisfy multiple criteria rather than simply completing one action. That gives OpenAI a way to evaluate whether the model understood the entire workflow instead of rewarding it for reaching the right screen or performing a handful of correct clicks.

This approach is particularly relevant to enterprise AI. Real business processes contain exceptions, approvals, dependencies and internal rules that generic computer-use benchmarks often fail to capture.

OpenAI Says Customer Data Was Not Used for the Research

Data privacy is an obvious issue when AI companies work with legal software. Contracts can contain sensitive commercial information, personal data and confidential terms.

OpenAI says the research did not use OpenAI customer data, OpenAI internal contracts or nonpublic Ironclad customer data or contracts for training or evaluation. Instead, researchers created simulated tasks using contracts available through the SEC’s public EDGAR database after applying filters designed to remove personal information.

That distinction is important because the project is trying to improve AI’s ability to work with professional software without turning confidential customer documents into training material.

The Partnership Builds on a Longer OpenAI-Ironclad Relationship

The latest research effort is not the first connection between the two companies. Ironclad previously incorporated OpenAI technology into its products, including AI Assist, which helps legal teams work with contracts.

Ironclad has also connected contract information and workflows with ChatGPT, while OpenAI has increasingly focused on bringing AI deeper into specialised professional environments.

The relationship gained another connection in 2026 when Ironclad co-founder Jason Boehmig joined OpenAI and took a leadership role in its legal programme.

OpenAI Wants to Repeat the Model With Other Software Companies

Ironclad appears to be the first example of a broader research strategy. OpenAI says it is inviting a small number of software companies to work directly with its research and engineering teams on difficult professional tasks that current AI agents cannot reliably complete.

The model is fairly straightforward: the software company brings knowledge of its industry and workflows, a secure environment for testing and concrete examples of where current AI systems fail. OpenAI can then turn those problems into training and evaluation tasks.

That could become a significant part of how future AI agents are developed. Instead of testing models only against generic computer tasks, researchers can measure them against the messy workflows businesses actually care about.

Contracting Is Becoming a Test Case for Agentic AI

Legal contracting is a particularly useful environment for this kind of research because the work involves structured processes but also requires judgment, permissions and business-specific rules.

A contracting workflow may require different approvals depending on the value of a deal, the jurisdiction involved or the terms included in an agreement. An AI agent that handles such a process needs to keep track of those conditions throughout the task.

That is a harder problem than document summarisation. It also points toward where agentic AI could become more valuable: not just generating information, but moving work through enterprise systems while maintaining the rules that govern it.

Blockchain and Enterprise Software Could See More AI Agents Inside Workflows

Although the Ironclad project is focused on contracting, the underlying technology has implications beyond legal departments. Financial services, procurement, compliance, insurance and other industries rely on software with similarly complex rules and multi-step processes.

For blockchain companies in particular, the development of capable computer-using agents could eventually intersect with areas such as digital asset compliance, transaction monitoring, smart-contract operations and enterprise workflows.

The technology still has plenty to prove. A faster agent is useful only if it also performs the correct actions, follows business constraints and knows when a task requires human intervention.

OpenAI and Ironclad Are Testing What Comes After AI Assistants

The more important story here may not be contracts themselves. It is the move from AI that helps with a task to AI that can potentially carry a workflow through specialised software.

OpenAI’s work with Ironclad gives researchers a controlled environment for testing that transition. GPT-6 Astra’s results show improvement, but the 55% average score also makes clear that these systems are not finished products for fully autonomous professional work.

The next challenge is obvious: make agents more reliable without losing the speed gains. In legal and enterprise environments, getting 80% of a workflow right is not enough if the remaining 20% contains the approval, clause or business rule that matters most.

That is the real test OpenAI is now taking on.

Sources

Related Posts