What Copilot Can (and Can't) Actually Review in a Contract

Ask Copilot in Word to "review this contract" and you'll get an answer that sounds like a review: organized, confident, reasonable-sounding paragraphs about risk. Whether it's actually caught the thing you needed caught is a separate question, and the honest answer is that it depends heavily on what you asked for and how you asked for it. For anyone using a personal Microsoft 365 subscription rather than a firm's enterprise Copilot deployment with legal-specific tooling, that gap matters even more, because the base capability is general-purpose document AI, not a legal review product. The Beginner's Guide to Copilot covers the general fundamentals this article assumes.

Scanning for a specific clause type, which it does reasonably well

Copilot's most reliable use in contract review is a targeted, named search: find every instance of a specific clause type and summarize what it says. This plays to a real strength, reading a full document and pulling out every place a pattern appears, faster than doing it by eye across a fifty-page agreement.

That last part, checking whether notice periods differ between the two parties, is a genuinely useful thing to ask for directly, because an asymmetric termination clause (the counterparty needs 90 days' notice, you only get 15) is exactly the kind of detail that's easy to miss reading a dense section start to finish, and exactly the kind of thing a clause-type-specific prompt is well suited to surface.

Liability caps, indemnification triggers, assignment restrictions, and exclusivity language all work well as this kind of named, structured request. The pattern that works is always the same: name the specific clause type, ask for specific attributes of it (who, what triggers it, what limits it), and ask Copilot to quote the actual language rather than paraphrase it, so you can check its reading against the real text yourself.

What "review this contract" actually gets you, and why it's not enough

A broad, unscoped request for a general review produces a broad, unscoped answer: a reasonable-sounding list of "areas of potential concern" that reads like due diligence but wasn't built against any specific standard, any specific deal history, or any specific client risk tolerance. It's not that the answer is wrong, exactly. It's that "concerning" isn't a legal judgment Copilot is positioned to make responsibly; it's a pattern-matched guess at what a contract review often flags.

A worked example: catching what a broad ask misses

Take a hypothetical services agreement with this clause in Section 14:

Section 9's own definition of "Confidential Information" in this hypothetical draft is broad enough to cover most of the technical documentation the parties exchange, which matters, because it means the exception is not a narrow edge case, it's the category most likely to actually get litigated.

A broad, unscoped prompt run against that document produces something like this:

That response isn't false. It mentions the cap exists. It never mentions the carve-out that removes the cap for confidentiality claims, which is the one detail that changes what the clause actually protects against. A narrower, targeted prompt catches it because it's built to look for exactly that structure:

This is the shape worth reusing for any clause where the real risk lives in the interaction between two separate sections rather than in either one alone: name the primary clause, then explicitly ask it to hunt for carve-outs or exceptions elsewhere in the document.

That contrast is also where the line between what Copilot can and can't do actually sits. Finding the carve-out is defined-pattern matching: you named two clause types and asked whether one modifies the other, which is a checkable, mechanical task against the document in front of it. What Copilot still can't tell you is whether a 12-month fee cap is generous or thin for a services agreement of this size and industry, whether this particular counterparty's negotiating history suggests the carve-out is a line they'd actually hold on, or how a court in the governing jurisdiction has treated similar carve-out language when it's been litigated. Those are judgment calls built on context that lives outside the four corners of the document, and no amount of prompt precision gets Copilot access to context it was never given.

What a personal Microsoft 365 subscription doesn't include

It's worth being explicit about the gating here: a personal Copilot subscription (Personal, Family, or Premium) is general-purpose document AI running inside Word, not a legal-specific contract review product with citation checking against case law, clause libraries benchmarked against market standards, or redlining workflows built for outside counsel. Firms using dedicated legal AI tools, or an enterprise Microsoft 365 Copilot deployment with legal-specific extensions, are working with a materially different tool than what's described here. One more point matters for client work: a personal account and a work account fall under different data-protection terms. Before pasting confidential client material into any Copilot, confirm which terms apply to the account you're signed in with and what your firm's policy and professional obligations require. For an individual using a personal subscription, the honest framing is: useful as a fast first pass that surfaces named clause types for you to verify, not a substitute for the judgment a qualified reviewer brings to a contract that actually matters.

Official sources

• Copilot in Microsoft 365 Personal, Family, Premium, and Pro

Checked on September 21, 2026. Features, plans and names change often, so the vendor's own pages are the final word.