.png)
Agentic Contracting: What OpenAI and Ironclad's Research Means for Legal Ops
AI agents are moving from reading contracts to running the workflows behind them. That shift is the core of agentic contracting, and new research from OpenAI and Ironclad shows how fast it is arriving.
On October 6, 2026, OpenAI published Advancing computer use with Ironclad. It describes how the two companies trained and tested AI agents on real contracting work inside Ironclad. For legal ops leaders, the lesson is practical: the teams that win with agents will be the ones whose workflows are ready for them.
What OpenAI and Ironclad actually tested
The research did not ask models to write clauses. It asked them to configure the contracting system itself. Ironclad employees and Ironclad users at OpenAI helped define 11 tasks across legal, commercial, and procurement work.
The examples will look familiar to any legal ops team:
Setting up a nondisclosure agreement process.
Building procurement approval logic.
Updating a reusable clause so it reflects the jurisdiction a requester selects.
Each task was scored against 8 to 50 criteria, depending on complexity. Ironclad provided hosted software environments for practice, and the tasks were built from public SEC EDGAR contracts. OpenAI states that no customer data or nonpublic Ironclad customer data was used.
The results, read carefully
The headline numbers are strong. Across the 11 tasks, OpenAI reports that its new model, GPT‑6 Astra, averaged a 55.0% rubric score, compared with 41.6% for GPT‑5.6 Sol. Estimated time per attempt fell from 37.0 minutes to 19.2 minutes.
The caveats matter just as much. OpenAI notes that the tasks were simulated and do not cover all Ironclad workflows. The time figures are modeled estimates, not measured customer savings.
So the right reading is optimistic and disciplined. Agents are getting better at multi-step configuration work quickly. A 55% average score also means the best released model still missed many criteria. That gap is where your operating model does the heavy lifting.
Why agentic contracting is a workflow problem
Contracting is not a string of isolated clicks. A single software purchase may need Finance approval above a spending threshold, Security review for certain data, and Legal review for nonstandard terms. OpenAI uses almost exactly this scenario in its research.
An agent can finish nine of ten steps and still fail the process. It only takes one missed condition or one misrouted exception. OpenAI makes the same point: an agent that loses a rule halfway through a task limits what a business can safely ask it to do.
That is why agentic contracting depends less on the model and more on the system around it. Clear rules, clean data, and connected platforms give agents a fair chance to succeed. They also give your team a way to prove the agent did the job right.
A worked example: one purchase, three reviewers
Picture a request to buy a new analytics tool. The workflow should send it to Finance because it is above the spend threshold. It should add Security because the vendor will process customer data. It should add Legal only if the vendor rejects your standard terms.
Now write the checks a reviewer would use. Does a request above the threshold reach Finance? Does a request below it skip Finance? Does a customer-data flag always trigger Security? Each check is simple, and together they define "correct." That is the same structure the OpenAI research used to score agents, and it works just as well for people.
We made a related case in How AI Agents Enhance CLM (Not Replace It). This research adds hard evidence to that view.
Five readiness moves for legal ops leaders
You do not need to wait for perfect models to prepare. These five moves make your contracting stack agent-ready and improve it for people today.
1. Write your business rules down
Agents follow rules they can read. Many approval rules still live in someone's head or a stale policy PDF. Document thresholds, reviewers, and fallback paths in one place your CLM can enforce.
2. Turn every exception into a named path
Exceptions are where processes break. List your common deviations, such as nonstandard payment terms or new data types. Give each one an owner, a route, and a clear trigger.
3. Define "done" as testable criteria
The research scored each task against detailed criteria. Borrow that idea. For every workflow, write the checks a reviewer would use to confirm it works, then test against them before and after any change.
4. Keep people at the control points
OpenAI states that human oversight still matters as agents improve. Decide where a person must approve, such as spend above a threshold or high-risk terms. Build those gates into the workflow so they hold no matter who, or what, does the setup.
5. Connect the systems your rules depend on
Approval logic often needs data from Salesforce, Coupa, or NetSuite. If that data is missing or out of sync, no agent can route correctly. A sound integration layer, built with tools like Workato or native connectors, keeps every decision grounded in current data. Our guide to CLM system architecture design covers how to plan that layer.
What this means for each executive
Agentic contracting touches every team that depends on the contract process. Each leader has a different first question to answer.
General Counsel: Which decisions must always stay with a lawyer, and are those gates enforced in the system today?
Legal Ops: Are your workflows documented well enough that a new hire, or an agent, could configure them without tribal knowledge?
Procurement: Do approval thresholds and supplier risk flags flow from Coupa or NetSuite into the CLM, or are they re-keyed by hand?
RevOps and Deal Desk: Does Salesforce pass clean deal data into contract requests, so routing rules have what they need?
Clear answers to these questions are a strong sign your stack is ready for agents. Gaps show you where to invest first.
How to pilot an AI agent in your CLM
Start small and measure everything. A focused pilot builds trust faster than a broad rollout.
Pick one high-volume, rules-based workflow, such as NDA setup.
Write the acceptance criteria before the agent touches anything.
Run the agent in a sandbox and compare its output to your criteria.
Add human approval at every control point you defined.
Expand only after the workflow passes consistently.
Not sure where your team stands? Our CLM Readiness Assessment is a quick way to find the gaps that would trip up an agent.
Frequently asked questions
What is agentic contracting?
Agentic contracting is the use of AI agents to carry out multi-step contracting work, such as configuring approval workflows or updating clause libraries, inside a CLM platform. People still set the rules and approve key decisions.
Can AI agents run CLM workflows on their own today?
Not reliably. In OpenAI's research with Ironclad, the best released model averaged a 55.0% rubric score on 11 simulated tasks. Agents are improving fast, but human review at key control points is still essential.
How should legal ops prepare for AI agents?
Document your business rules, define clear paths for exceptions, write testable acceptance criteria, keep human approval at control points, and connect the systems your rules rely on.
Make your contracting stack agent-ready
Agentic contracting will reward teams with clear rules, connected systems, and strong governance. We help legal, procurement, and RevOps leaders build that foundation in Ironclad and the platforms around it. If you want a practical plan for your stack, let's talk.
Comments