What automated contract review catches, and what it structurally cannot

Everything it does well is a property of the text

A contract review model reads a document and reports on it. That framing sounds trivial and it is the entire limit. Completeness, consistency, arithmetic and symmetry are properties of the file, and a system that reads every line at page-one attention beats a person who read the first eight pages carefully and skimmed the rest. Value, norm, enforceability and counterparty behaviour are not in the file at all.

What it catches reliably, and why

This is not a small category. Most bad contract outcomes come from clauses nobody read rather than clauses nobody understood, and the failures above are precisely what tired reading produces. A missing limitation of liability in a service agreement, an auto-renewal notice date buried in clause 14.3, a defined term that changed meaning between drafting rounds — these are the things that cost money, and they are all catchable.

Four things it structurally cannot do

• Know what the deal is worth to you. A liability cap at one month's fees is fine if the work is a logo and ruinous if you are hosting someone's payroll data. The clause reads identically in both contracts. The difference lives in your business, not in the document.
• Know the market norm in your sector. "Unusual" is a comparison, and the comparison set that matters is what firms like yours actually sign. Ninety-day payment terms are a red flag in freelance design and the standard in automotive supply. A model trained on the internet's contracts averages across sectors that should not be averaged.
• Judge whether a clause is enforceable where you are. Enforceability is law applied to facts, varies by state and country, and changes. A non-compete that holds in one jurisdiction is void in the next. A machine reading the four corners of the page has no way to run that test.
• Tell you the counterparty never performs. The most valuable fact about a contract is frequently the other side's payment history, and it appears nowhere in the file. Ten minutes asking two people who have worked with them beats any review of the document.

Why enforceability is the hardest question to automate

It is worth being specific about how these systems fail on legal questions, because the failure is not random noise. Researchers at Stanford tested the purpose-built legal research tools sold by the major providers — systems grounded in real case databases rather than open-web models — and published the results in the Journal of Empirical Legal Studies in 2025. They found hallucination rates of roughly one in six for one product and about a third for another, against vendor marketing that had described the citations as hallucination-free.

Those are research tools rather than contract review tools, so the comparison is not exact. The transferable point is the direction of the error: grounding reduces fabrication without removing it, and what remains clusters where the question is about authority. A review saying a clause "may be unenforceable in your jurisdiction" is a prompt to check, not a finding. Treat any statute, section number or case it names as unverified until you have looked it up.

A division of labour that works

The sequence below is deliberate. The first step is the one people skip, and skipping it is what turns a useful report into forty undifferentiated notes.

1. 1

   Write the deal down in one line first

   Before uploading anything: what you are getting, what you are giving, over what period, and what the worst realistic outcome is. Everything the machine cannot see is in that sentence. Having written it changes which flags you care about, and it takes ninety seconds.

2. 2

   Run the machine pass over the whole document

   Completeness, dates, definitions, cross-references and one-sidedness, on every page rather than the pages you would have read. Set the governing law before it starts — the framing of half the findings depends on it.

3. 3

   Triage the findings against your one line

   Most flags will be real and irrelevant. Sort them into what changes the outcome for this deal and what does not, and be ruthless: a negotiation with twelve asks gets none of them, and a negotiation with three usually gets two.

4. 4

   Do the human pass on the four blind spots

   Price the exposure against your own numbers. Ask someone in your sector what is normal. Check the two or three clauses that would decide a dispute against local law, with a lawyer if the sums justify it. Ask two people about the counterparty.

5. 5

   Verify anything specific before you quote it

   Statutes, section numbers, cases and regulatory thresholds named in the report. If you cannot find the authority in a minute, drop the point rather than putting it in an email to the other side.

The failure mode is reading the report as a verdict

A risk score is an instrument reading. A low one means the document contains no unusual structural risks, which is a statement about drafting and not about the deal. That distinction is easy to lose when the output arrives as a number with a colour attached.

The mirror-image mistake is treating every flag as a problem. Contracts are negotiated documents and a degree of asymmetry is normal — the party carrying the delivery risk usually gets the termination right, and objecting to that reads as inexperience. Knowing which flags are ordinary is what our guide to contract red flags is for.

When to skip the machine pass entirely

• A short document you will actually read. A two-page order form does not need a report; it needs four minutes and the approach in how to read a contract.
• Anything already in dispute. The question there is conduct, correspondence and remedies, not the wording — and the wording has usually already been read three times by both sides.
• Where the whole question sits outside the text. Whether this supplier is solvent, whether the licence you are buying covers the use you have in mind, whether the person signing has authority.
• Where you cannot upload it. Confidentiality obligations frequently prohibit putting a counterparty's draft into a third-party tool, which is a real constraint rather than a formality — see AI clauses in client contracts.

What the machine is actually buying you

Not judgement, and any page selling it as judgement is selling something it cannot deliver. What it buys is uniform attention: forty pages read at the same level of care, so the completeness and consistency failures are off the table before a person starts. The scarce resource was never the reading. It was the willingness to think hard about three specific clauses — and that is much easier to spend well once you know which three they are. A saas agreement reviewed this way takes an hour rather than an afternoon, and the hour lands where it matters.