OpenAI and Ironclad Train Computer‑Operation Agents for Enterprise Contract Work

On 6 October 2026, OpenAI and the contract‑management firm Ironclad publicly disclosed a collaborative research effort aimed at building computer‑operation agents capable of performing complex business processes inside enterprise‑grade software platforms. The partnership defined eleven distinct challenges that span the legal, commercial, and procurement domains, and it introduced a systematic evaluation framework that checks whether an agent can go beyond simple mouse clicks to satisfy the full set of workflow requirements expected by end users. Ironclad contributed a sandbox‑style training environment, while OpenAI applied synthetic task generation together with reinforcement‑learning techniques to train the agents. The initiative is positioned as the first large‑scale experiment that tests whether artificial intelligence can reliably execute real‑world operational steps at the level required by modern corporations, and the results will be released as an industry‑standard benchmark that other AI developers can reference.

Research Background

Legal departments and procurement teams within corporations routinely confront labor‑intensive activities such as drafting contract documents, managing approval flows, and tailoring clause language to meet jurisdiction‑specific regulations. These manual processes are prone to human error and procedural delays, which in turn can slow overall business velocity. If artificial intelligence can not only mimic screen clicks and data entry but also interpret document content and automatically customize contractual terms, companies could achieve substantial reductions in processing time while simultaneously improving the quality and consistency of output. OpenAI and Ironclad therefore set out to address these pain points by training a computer‑operation agent that operates at a practical, work‑level standard.

Task Design

The research team constructed eleven separate tasks that reflect typical activities across the legal, commercial, and procurement functions. Each task incorporates between eight and fifty explicit criteria that go beyond mere click completion, requiring the agent to verify that the final workflow meets pre‑defined requirements. For instance, the creation of a non‑disclosure agreement includes criteria such as defining confidential information, specifying signing authority, and setting a retention period. In the procurement approval process, the agent must respect budget caps, evaluate vendors, and confirm that all approval steps have been completed. Jurisdiction‑specific clause‑modification tasks embed legal requirements for the target region, localized clause wording, and compliance checks as part of the evaluation rubric.

• Creation of a non‑disclosure agreement
• Execution of a procurement approval process
• Modification of jurisdiction‑specific clauses

Evaluation Method and Results

Agent performance was measured using representative synthetic tasks combined with reinforcement‑learning‑based evaluation metrics. According to the published report, the model named GPT‑6 Astra achieved an average success rate of 55.0 % and reduced the estimated time required to complete tasks from 37.0 minutes to 19.2 minutes, effectively halving the duration. The GPT‑5.6 Sol model recorded a 41.6 % success rate, with a comparable reduction in execution time. An internally developed prototype model surpassed both public offerings, reaching a 63.7 % success rate, thereby demonstrating superior performance relative to the released models. These figures illustrate that the agents can orchestrate multi‑step click sequences and conditional logic to fulfill complex workflow requirements.

The internal model’s 63.7 % success rate was especially notable in tasks involving customized legal clauses and multi‑stage procurement approvals, where it showed markedly higher efficiency. GPT‑6 Astra’s 55.0 % success rate was evaluated as the highest among general‑purpose models, suggesting viable applicability in real‑world business settings. In contrast, GPT‑5.6 Sol’s 41.6 % success rate highlighted how task difficulty and conditional complexity can cause performance to vary. Collectively, these outcomes underscore the direct influence of model architecture and the design of reinforcement‑learning reward structures on overall success rates.

Nevertheless, the study found that 55 % of the eleven tasks were deemed impossible to fully automate, limiting the research to controlled environments. Deploying such agents in production would require additional governance mechanisms, including permission management, audit logging, and validation of configuration settings. The agents’ actions presuppose strict access controls and comprehensive change‑history records to mitigate risks associated with erroneous clause modifications or inappropriate approvals. Consequently, moving the research findings directly into live corporate systems would necessitate further security assessments and user training before practical adoption.

The concrete impact of the research can be summarized as follows: a reduction in the time needed to draft initial contract versions, cost savings through automation of procurement approval flows, and a decrease in compliance risk thanks to dynamic generation of jurisdiction‑specific clauses. Companies can leverage Ironclad’s sandbox environment to create custom tasks that mirror their internal workflows and then test OpenAI’s agents against those tasks. For tasks that demonstrate high success rates, organizations can progressively roll the agents into production, operating them under human‑in‑the‑loop supervision to ensure safety and reliability. This approach enables legal teams to focus on higher‑order analytical work while allowing procurement groups to halve approval cycle times.

Overall, the joint research effort by OpenAI and Ironclad represents a significant step toward the practical deployment of computer‑operation agents for enterprise use. Building on the findings, both companies have announced plans to develop implementation guidelines and a security framework, and they intend to broaden the partnership to cover additional use cases. The anticipated outcome is an accelerated digital transformation of AI‑driven contract management and procurement processes, delivering a balanced combination of operational efficiency and regulatory compliance for businesses worldwide.