What can AI contract review actually do, and what can it not?

What it does well

These are tasks where reading every word consistently beats reading selectively, which is where software has a real advantage over a tired human at 5pm.

Finding clauses that are present

Locating the liability cap, the termination provisions, the renewal mechanism and the indemnities across a long document, including where they are scattered or defined elsewhere.

Noticing what is missing

Absence is harder for a person to spot than presence. No termination for convenience clause, no liability cap, no data deletion obligation. Consistent checking is where software is strongest.

Comparing against a reference point

Telling you a 90 day cure period is longer than the market position, and by how much. This only works if the reference point is real, which is why sample sizes matter.

Triage

Sorting a contract into the five to twelve things worth a human's attention and the rest that is standard. This is the honest core of the value: not replacing review, but deciding what gets reviewed.

What it does badly, or cannot do

A vendor who will not tell you this list is not being straight with you.

Judging whether a term is right for you

A liability cap that is fine for a $5,000 subscription may be unacceptable for a system your operations depend on. The market position is an input to that decision, not the decision.

Understanding your leverage

Whether you can actually get a change depends on deal size, timing, competitive alternatives and the relationship. None of that is in the document.

Giving legal advice

Applying law to your specific circumstances and telling you what to do is legal advice. Software can tell you what a clause says and how it compares. It cannot tell you whether to sign.

Reading a badly scanned document reliably

Text extraction from poor scans, photographs of pages, or heavily formatted tables is where errors enter quietly. Ask any vendor how they handle it and what happens when extraction fails.

Knowing about the side agreement

Email exchanges, verbal assurances and separate order forms change the deal and are frequently not in the file you upload.

Why accuracy claims are hard to compare

Vendors quote accuracy percentages. They are close to meaningless without knowing what was measured, on what documents, against whose judgement, and how disagreement was resolved.

A more useful question is what the tool shows you when it is uncertain. A system that surfaces confidence honestly, and lets you confirm or reject a finding, is more trustworthy than one that presents everything with equal certainty.

The other useful question is whether every finding is traceable. If the tool says a clause is unusual, you should be able to see the clause, the comparison, and where the comparison came from, in one step. If you cannot check it, you are being asked to take it on faith.

Common questions

Is AI contract review accurate?

It depends entirely on the task. Locating clauses, spotting missing provisions and comparing terms against a reference set are tasks where automated review is reliable. Judging whether a term is acceptable for your situation, or what leverage you have to change it, is not something it can do. Treat vendor accuracy percentages with caution unless you know what was measured and against whose judgement.

Can AI replace a lawyer for contract review?

No. Applying law to your circumstances and advising whether to sign is legal advice. Automated review is best understood as triage: it decides what deserves a human's attention and gives you the comparison to argue with, which is a different job from advising you.

What should I ask a contract review vendor?

Ask what happens when the tool is uncertain, whether every finding links back to the clause and the source of its comparison, how text extraction failures are handled, where your documents are processed, and whether your contracts are used to train models.