Skip to Content
What you can doContract review

Contract review by several AI models

For legal, purchasing and due-diligence teams, where one AI model’s reading isn’t enough. Several models from different companies each read the contract from their own angle, and a final “judge” model writes one risk memo from what they found. A missed auto-renewal or indemnity clause gets caught by the model that was looking for it.

For example: one model reads as your company, one as the other side, one checks the governing law. The memo lists the risks all of them agreed on.

What you need

  • At least two AI providers connected (for example one from OpenAI and one from Mistral), under Models, Connect a provider.
  • Optionally, a search tool over your clause library or past contracts.

How to build it

  1. Create the assistant. Open Agents, click Create agent and switch to Autonomous. In Instructions, write: “Review the contract and write a risk memo.”

  2. Set the work mode to Panel. In the Work mode section, pick Panel. You then fill in the main model, two or more panelists and, if you like, a judge. Give each panelist a different provider, and under its Advanced, an angle in its extra instructions: “your company’s risk”, “the other side’s leverage”, “compliance with the governing law”. A panelist can also be one of your other assistants, with its own tools.

    Work mode set to Panel, with panelists on different providers and a judge

  3. Add a judge. The judge reads what the panelists said and writes the one answer they agree on. Without a judge, the main model does it.

  4. Optionally give it a search tool under Tools and APIs, so its reading can point to earlier contracts, not just the text in front of it.

  5. Send it the contract. Add a Web chat channel on its Channels tab, publish it, and attach the contract as a PDF, or have your document system call it.

  6. Read the memo. Each run shows every panelist’s answer, which model gave it, and how far they agreed.

Ideas

  • The same setup works for research summaries, due diligence or incident reviews: anywhere one model isn’t enough.
  • For large volumes of low-stakes work, use Cascade instead: a cheaper model answers first, and the main model steps in only when a checker rejects the draft.
  • Have it ask a person before anything above a risk level goes out (see Approvals).
Match Stripe payoutsGoogle Drive and SharePoint