OpenAI title card reading Advancing computer use with Ironclad — OpenAI announced its Ironclad research collaboration on October 6, 2026. Image: OpenAI.

AI contract management just got one of its most honest tests so far. On October 6, 2026, OpenAI shared results from a research project with Ironclad, the contract software company. The goal was to see how well AI agents handle real contracting work inside real business software.

The short answer: better than before, but not ready to run alone. OpenAI's newest model, GPT-6 Astra, scored 55.0% on average across 11 contracting tasks. The older GPT-5.6 Sol scored 41.6%.

I'm Nitin from Debugged Pro. Our team builds software and workflow tools for clients in India and the USA, and contracts touch almost every project we run. So I read the full OpenAI announcement closely. Here is what it says, what it does not say, and what I think teams should do with it.

What OpenAI and Ironclad Actually Tested

This was not a chatbot writing contract summaries. It was an agent clicking through Ironclad's product and setting things up, the way a legal operations person would.

OpenAI says it is now working with a small number of software companies to train agents on specialized business workflows. Ironclad is the first partner. Ironclad employees, along with people who use Ironclad inside OpenAI, helped pick 11 tasks across legal, commercial and procurement work.

Examples from the announcement include:

Setting up nondisclosure agreements.
Creating procurement approval processes.
Updating a reusable legal clause so it matches the jurisdiction a requester selects.

OpenAI estimates each task takes an experienced user about 30 to 40 minutes. These are fiddly, multi-step jobs, not one-click actions.

A realistic example from the research

OpenAI described a legal ops team setting up a process for buying software. Finance may need to approve purchases above a set amount. Security may need to review some requests. Legal may need to check nonstandard terms.

The person doing the setup has to turn those rules into an intake form, document templates, approval rules and a record of the final agreement. An agent has to do the same, and keep every rule in mind while it moves through the screens. Getting single steps right is not enough. The finished process has to work for every situation it was built for.

How the AI Contract Management Agents Performed

Each task was graded against 8 to 50 criteria, depending on how complex it was. That let researchers see exactly which parts a model got right and where it slipped.

Metric	GPT-5.6 Sol (High reasoning)	GPT-6 Astra (Max reasoning)
Average rubric score	41.6%	55.0%
Estimated time per attempt	37.0 minutes	19.2 minutes

OpenAI says that works out to a 32% higher score and 48% less time per attempt. An internal model used while building Astra did even better, at 63.7%.

In one side-by-side example, Astra met about 94% of a task's criteria in an estimated 20 minutes. Sol met about 85% of the same task in an estimated 32 minutes.

Read the fine print before you quote these numbers

OpenAI's own footnotes add three caveats that matter:

The results cover 11 research tasks only, not all Ironclad workflows.
The times are simulated estimates. They are based on assumed model speeds, not measured time saved by real customers.
A 55% average still means a lot of missed criteria. On a contract, one missed approval rule can be the one that matters.

I would treat this as a progress report, not a product launch. It shows the direction of travel. It does not prove an agent can manage your contracts without a person checking the work.

Ironclad is the first software partner in OpenAI's computer use research program. Image: Ironclad.

How the Agents Were Trained

The training setup is the most useful part of this story for software teams. It had three pieces:

Tasks chosen with people who know the work. Ironclad staff helped define what success looks like.
Detailed grading criteria. Every task had a clear checklist, so progress could be measured.
A safe place to practice. Ironclad provided hosted copies of its product where models could try tasks again and again.

OpenAI's researchers built synthetic training tasks around common workflows. They then used reinforcement learning, which means the model improves through practice and feedback.

The data choice is worth noting too. OpenAI says the practice tasks were built from contracts that are public in the SEC's EDGAR database, after filters to remove personal information. It says it did not use OpenAI customer data, its own internal contracts, or private Ironclad customer contracts.

Why Ironclad, and Why Now

The two companies already have close ties. In June 2026, Ironclad co-founder Jason Boehmig joined OpenAI to lead product for its legal vertical. LawSites reported the move at the time and noted he remains Ironclad's executive chairman.

LawSites also framed legal tech as a growing fight between the big AI labs. It pointed out that Anthropic had released more than 20 legal software connectors and 12 legal plugins for Claude less than a month earlier.

Ironclad's chief technology officer, Sunita Verma, explained the company's view in the announcement. "For AI to be genuinely useful in complex areas of contracting, it has to do more than complete individual actions," she said. "Agents need to understand the full contracting lifecycle, including how business workflows connect while preserving the controls teams rely on."

Ironclad was also careful to say that human oversight still matters, and that a full contracting platform is still needed. That is a fair position for a software vendor. It is also good advice for buyers.

What AI Contract Management Means for Agencies in India and the USA

Most agencies do not run a legal department. But we handle a steady flow of NDAs, statements of work, master service agreements and vendor terms. Here is where this research touches daily work.

1. Your own contract admin

Setting up approval rules and templates is exactly the type of work tested here. A well-trained agent could one day handle the first draft of that setup. Today, the safe use is narrow: let AI prepare, and let a person approve.

2. Cross-border jurisdiction clauses

One test task was updating a clause based on the jurisdiction a requester picks. That is a real need for Indian agencies with US clients, and for US firms hiring Indian vendors. It is also where an error is costly. Always have a qualified lawyer review governing law and dispute clauses.

3. Client SaaS products

If you build B2B software for clients, this is the bigger signal. OpenAI is now asking software companies to bring hard tasks that agents fail at today. That means agents will increasingly be expected to work inside business apps, not just chat windows.

We covered the wider risks of the new model in our post on GPT-6 Astra agent risks for agencies. The same lessons apply here.

Three Ironclad team members working together on laptops at a meeting table — Ironclad staff helped define the tasks and success criteria used in the research. Image: Ironclad.

How to Make Your Software Ready for Computer Use Agents

OpenAI listed what it wants from partner companies. The list doubles as a good checklist for any team building business software:

A concrete task. Describe exactly what the agent should do.
Evidence of failure. Show where current agents get it wrong.
A clear way to judge success. Write the grading checklist before you test.
People who know the work. Domain experts must help define "done."
A secure test environment. Agents should practice on a sandbox, never on live client data.
Data that is safe for research. Strip personal and confidential details first.

Even if you never partner with OpenAI, this structure works for your own agent pilots. Our guide to computer use tips for delivery teams covers the day-to-day side of running agents on real screens.

A Simple Pilot Plan for AI Contract Management

If you want to try AI contract management in your own business, start small and measure everything. Here is the order I would follow:

Pick one repeatable workflow. NDA setup or a standard vendor approval is a good start.
Write down the rules. List every approval limit, reviewer and exception.
Build a checklist score. Use 10 to 20 yes or no checks, like OpenAI's criteria.
Test in a sandbox first. Use dummy contracts, not real client agreements.
Keep a human approver on every output. No agent should send or sign anything on its own.
Log every action. You need a record of what the agent changed and when.
Review results monthly. Expand only when scores stay high over several weeks.

For governance, the NIST AI Risk Management Framework is a practical, free reference. It helps you map risks, set controls and decide who is accountable.

Questions to ask any AI contract tool vendor

Which tasks has the agent been tested on, and what were the scores?
Were the tests run on your workflows or on sample data?
Is our contract data used to train any model?
Can we set hard limits on what the agent is allowed to change?
What happens when the agent is unsure?

Key Takeaways

OpenAI and Ironclad tested AI contract management agents on 11 real workflows, such as NDA setup and procurement approvals.
GPT-6 Astra scored 55.0%, up from 41.6% for GPT-5.6 Sol, with estimated task time nearly halved.
The times are simulated and the tasks are a small sample, so treat the numbers as research results.
Human review is still required. Even Ironclad stresses oversight and strong controls.
Software teams should prepare now with clear rules, test sandboxes and scored checklists.

Want help planning an agent pilot or making your product ready for AI workflows? Our team offers free software development consulting for agencies and businesses in India and the USA.

About the Author

Nitin Kumar is the Lead Project Coordinator at Debugged Pro. He works with development, QA and delivery teams across India and the USA to keep client projects on track. He writes about the AI and software news that changes how agencies plan, build and ship work.