Contract software is becoming a training ground for professional AI agents

A professional software agent needs to remember more than where the buttons are. If a procurement request exceeds a threshold, finance may need to approve it. A security review may depend on the product category. A reusable clause may need different wording for a different jurisdiction. Completing the visible clicks while losing one rule can leave the business with a clean-looking but invalid workflow.

OpenAI and Ironclad have turned that problem into a model-training programme. Their research collaboration defines specialised contract tasks inside hosted Ironclad environments, gives each task a detailed rubric and lets models practise through reinforcement learning. OpenAI describes GPT-6 Astra as its first frontier model trained on these Ironclad tasks.[1]

The task is the complete workflow, not the path through the screen

Ironclad employees and OpenAI staff who use the product selected 11 legal, commercial and procurement tasks. Examples include configuring a non-disclosure agreement, creating a purchasing approval process and updating a reusable legal clause to match a requester’s selected jurisdiction. OpenAI estimates that an experienced user would need about 30–40 minutes per task.[1]

Each task was evaluated against 8–50 criteria. That allows the system to receive credit for correctly configuring some requirements while exposing the exact rule it missed. A single “completed” label would hide the difference between a workflow that is almost finished and one that routes the wrong contract around a necessary approval.

The score increase is 13.4 percentage points, or about 32% relative to Sol’s score. The estimated time falls by 17.8 minutes, or about 48%. Those are meaningful changes within this task set. They do not say that Astra completes every contract workflow correctly: an average rubric score of 55% also shows how much professional work remains unresolved.[1]

The training data boundary is part of the result

OpenAI says the simulated tasks were built from contracts available publicly in the SEC’s EDGAR database after filters designed to remove personal information. It says training and evaluation did not use OpenAI customer data, OpenAI’s internal contracts, or non-public Ironclad customer contracts. This is an important design choice: the researchers needed realistic structures without quietly turning confidential agreements into training material.[1]

A hosted software environment adds another kind of realism. The model must preserve state across a sequence of actions and leave the final application in a condition the rubric can inspect. That is more informative than a screenshot benchmark where the agent receives credit for clicking the likely next control.

Ironclad is also adding agents to its commercial product

On 8 October, Ironclad expanded its own agent platform. Ironclad Agent acts as a conversational entry point, routing requests to specialised agents. A Contract Knowledge Graph connects clauses, obligations, related agreements and past decisions. Contract Family Agent can organise as many as 150 contracts per entity and route uncertain relationships into a human verification queue. Analytics Agent produces tables, charts and CSV files; Workflow Designer Agent can create or audit configuration changes.[2][3]

The access design matters as much as the agent catalogue. Ironclad says its policy-based access control applies organisational policies instead of hundreds of individual permissions. Conversational Intake only suggests workflows already available to the user and requires confirmation before launch. That is the practical difference between “the model knows a workflow exists” and “this person is authorised to start it”.[2][4]

The next benchmark is whether the agent understands the institution

Professional work is full of exceptions that cannot be inferred from the interface alone. The same contract value can require different approval in another business unit. A standard clause can become wrong when the governing law changes. Training against software actions, expert rules and finished-state tests gives the model a chance to learn that structure.

Contract management is therefore an early example of a broader shift. Enterprise applications are becoming training grounds in which models learn the logic of professional work. The model still needs the interface, but the durable capability comes from something deeper: understanding the rules, carrying them through a long sequence and checking that the completed system behaves as intended.

Report an error or suggest a correction