AI Contract Review, Explained: What It Actually Does

The four things it's actually doing

Strip away the marketing language and "AI contract review" breaks down into a handful of concrete tasks. Worth naming them separately, because they're different jobs with different reliability:

• Clause extraction. Pulling out the sections that matter - termination, indemnification, liability caps, assignment, governing law - so you're not hunting for them in a thirty-page document. This is the part models are genuinely good at: contracts follow enough structural convention that a model can locate "the indemnification clause" reliably, even when it's not labeled that way.
• Risk flagging. Pointing out language that deviates from what's typical for this contract type - an uncapped liability provision, a one-sided termination right, an unusually broad non-compete. This is pattern-matching against what "normal" looks like, not a legal judgment that the clause is a problem. A flag means "look at this," not "this is bad."
• Summarization. Condensing the document, or a section of it, into plain language you can scan in thirty seconds instead of five minutes. Useful for triage - deciding what to read closely first - and for briefing a client on what a contract says without handing them the whole document.
• Q&A against the document. Asking a direct question - "what's the notice period for termination," "is there a cap on liability," "does this renew automatically" - and getting an answer sourced from the actual text, instead of re-reading the whole thing to find it.

None of these four things involve the model deciding anything. They involve the model finding and organizing things faster than you'd find and organize them by hand.

Where this actually saves time

The honest case for AI contract review isn't "it reviews the contract for you." It's that a solo attorney reading a routine NDA or vendor agreement spends most of the time on things that are mechanical - locating the relevant clauses, checking them against what's standard, confirming nothing's missing - before spending any time on the part that requires judgment. Moving the mechanical part from fifteen minutes to two minutes doesn't change the judgment part at all. It just means more of a limited afternoon goes to matters that need it.

This matters more for contract types you review often than ones you review rarely. If you review commercial leases every week, a tool that reliably surfaces the renewal clause, the assignment clause, and the operating-expense pass-through in ten seconds is a real efficiency gain, because you already know what "normal" looks like and can evaluate the flag quickly. If you're reviewing a contract type you've never seen before, the tool is less useful as a shortcut and more useful as a first orientation - which is a different, more limited job.

What it can't do, and why that's not a flaw

The honest limits are worth stating plainly rather than glossing over:

• It can't tell you whether a clause is acceptable for this client, in this matter. That's a judgment call that depends on context the model doesn't have - the client's risk tolerance, the leverage in the negotiation, what else is in the deal. A flagged clause is a prompt to think, not an answer.
• It can hallucinate. A model can state something about the document that isn't actually there, especially on ambiguous or poorly scanned text. Every output needs to be checked against the actual contract language, not accepted at face value.
• It doesn't know your jurisdiction's specific requirements unless you tell it, and even then, it's not a substitute for knowing the law that applies to the matter. A clause that's standard in one jurisdiction can be unenforceable in another, and a general-purpose review tool has no way to know which jurisdiction's rules should govern its output.
• It won't catch what isn't in the document. A missing clause - no assignment provision, no indemnification language at all - is sometimes the actual risk, and that requires knowing what should be there, not just reading what is.

None of this makes the category not useful. It means the category is a research assistant for reading, not a decision-maker. Treated that way, it's genuinely useful. Treated as a substitute for legal judgment, it's a liability.

The cloud-vs-local question, for this specific task

Contract review is a use case where the underlying architecture matters more than most, because the input is the entire document, not a paraphrased excerpt. A cloud-based review tool needs the full contract to do clause extraction and Q&A well, which means the full contract - counterparty names, deal terms, pricing, sometimes tracked-change history - gets transmitted to a third-party server before the review even starts. A tool that runs the same review locally, on your own machine, never has that transmission step at all: the document is read, analyzed, and answered from, without leaving the device.

This is the architecture Privileged uses. It's a local-first document analysis and Q&A tool for solo attorneys, running on your own Mac via Ollama - no cloud calls, organized by matter, with a contract-review workflow template built around exactly the mechanics described above: extracting key clauses, flagging language worth a second look, summarizing, and answering direct questions against the document. It doesn't do legal research, and it doesn't draft anything - it's built around reading the contract you already have, not producing new documents. For the specific job of getting through a routine contract review faster without sending the document anywhere, that's the whole point.