To review a contract revision with an AI agent, do not paste both versions into the chat and ask what differs. Put both files in a folder, let the GroupDocs.Comparison.Mcp server find the changes locally, and ask Claude Desktop, Claude Code, Cursor or GitHub Copilot to judge them:

Compare contract-v1.docx with contract-v2.docx. List every change, and flag anything touching payment terms, liability, or termination.

The step-by-step version with config and troubleshooting is in the documentation: How to review contract revisions with an AI agent.

What is wrong with asking a chatbot to spot the edits?

The usual workflow has a gap in it. A counterparty sends back “a few small edits” as a new file with no track changes. You paste both versions into an assistant and ask what is different. The assistant answers fluently, and you cannot tell whether the answer is complete, because finding differences is a measurement and the model is producing text. A reworded sentence or a number inside a table cell is exactly the kind of change that a confident summary can leave out.

The evidence you can check is the shape of the output. A model’s summary is prose. An engine’s result is a list: each change has a type (inserted, deleted, style change), a page, a component, the changed fragment, the text around it, and the table cell when the change sits in a table. You can count that list and filter it. You cannot count a paragraph.

What should the agent decide, and what should the engine decide?

Question Decided by
Which text differs between v1 and v2? The engine, through analyze_changes
On which page, in which table cell? The engine, in each change entry
Does this change touch payment, liability or termination? The agent, reading the engine’s list
Is a 45-day term acceptable to us? You
What does the marked-up file look like? The engine, through compare

The agent never invents a difference. It only interprets the ones the engine returned.

How do you run the review?

Step 1: Ask what changed. Put contract-v1.docx and contract-v2.docx in the folder that GROUPDOCS_MCP_STORAGE_PATH points to, then use the prompt from the start of this post. The agent calls analyze_changes and reads the list against your three criteria. Because each entry carries its page and table cell, the answer can say that a payment schedule in the table on page 4 moved from 30 to 45 days, not only that something changed on page 4.

Step 2: Get the marked-up file.

Now give me the marked-up version I can send to legal.

The agent calls compare, which saves contract-v1_compared.docx in your output folder. Reviewers open it in Word with the differences highlighted. They need nothing else.

Step 3: Interrogate the list. The change list is still in the conversation, so follow-ups need no new comparison:

Did anything change in section 7? Were any clauses deleted outright? Show me only the style changes. I want to know if emphasis moved.

A style change and a wording change are separate entries, so the agent can tell “this became bold” from “this became 45 days”.

Example session

Abridged and illustrative; the real wording depends on your client and model.

You:    Compare contract-v1.docx with contract-v2.docx. List every
        change, and flag anything touching payment terms, liability,
        or termination.

Agent:  [calls analyze_changes: contract-v1.docx, contract-v2.docx]

        Flagged: the payment schedule in the table on page 4 changed
        from 30 to 45 days. Other changes are listed below with their
        page and surrounding text.

What are the limits of this review?

  • Check the license mode before you trust a result. In evaluation mode only the first 2 pages are compared and the result carries a watermark. On a twenty-page contract that produces a short, reassuring change list, which is the worst failure for this workflow. Ask “what is the license status of the comparison server?”; the agent calls get_license_status, and "mode": "evaluation" means you are seeing two pages’ worth of the contract.
  • Your prompt is not local. The documents stay on your machine. The text of your prompt, and any fragment the agent quotes back, goes to the model your client uses. For fully local operation, pair the server with a locally hosted model.

FAQ

Is there an AI contract review tool that keeps documents private? This one does: the comparison runs in a child process on your machine against files in a folder you choose. Only prompt text and quoted fragments reach your model provider.

Can I compare two contract versions with AI if there are no track changes? Yes. The engine compares the two files themselves, so it does not need track changes in either one.

Can it catch an edit hidden inside a table? Yes. Each change entry carries the table cell when the change sits in a table, so the agent can point to the cell instead of only the page.

Go deeper