Practical
How to use AI to understand a contract
A lease, a job offer, a freelance agreement, a terms-of-service update you're supposed to have read. Most people sign these having understood maybe half, because the alternative is paying a lawyer several hundred pounds to review a flat rental. AI closes a real part of that gap — but it closes the comprehension part, not the judgment part, and the difference is where people get hurt.
What it's genuinely good at
Three things, and they're worth more than they sound:
- Translation. "Explain this clause to me like I'm not a lawyer" works extremely well. Legal language is formulaic, that formula is heavily represented in training data, and plain-language restatement is squarely the kind of task these models do reliably.
- Structure. Turning eleven pages into "here is what you're agreeing to do, what they're agreeing to do, what happens if either side stops, and how this ends" is genuinely clarifying, and it's tedious enough that people skip it.
- Finding the absence. This is the underrated one. Ask what's missing compared to a standard agreement of this type. Contracts hurt you through the clause that isn't there — no cap on liability, no notice period, no mention of who owns the work — and absence is exactly what a human skim misses and a pattern-matcher catches.
What it's bad at, specifically
- Jurisdiction. Contract law differs by country and often by state or province. A model will happily explain a clause using assumptions from wherever most of its training text came from, without mentioning that. Notice periods, non-competes, and tenant rights vary enormously — non-competes in particular are near-unenforceable in some places and routine in others.
- Whether it's enforceable. "This clause is unenforceable" is a legal conclusion that depends on jurisdiction, case law, and how the rest of the document reads. Getting this wrong in either direction is expensive — either you sign something you shouldn't, or you ignore a term that binds you.
- Citing law. Ask for the statute or the precedent and you'll get one, formatted perfectly, sometimes invented. Invented legal citations have gotten actual lawyers sanctioned. Same rule as always: ask twice and see if the citation changes.
- Whether to sign. That depends on your leverage, your alternatives, how much you need this job or flat, and what you can afford if it goes wrong. None of that is in the document.
The prompt that keeps it in its lane
Strip identifying details first — names, addresses, account numbers, salary if you'd rather not — per the 10-second rule. The analysis works fine on an anonymised document. And if it's under an NDA, stop here and don't paste it at all.
I'm reviewing a [type of contract] as the [tenant /
employee / freelancer]. I'm in [country/state].
Do NOT tell me whether to sign, and do NOT cite any
statute or case law.
Do this instead:
1. Plain-English summary of what each side must do.
2. The five clauses most likely to cost me money or
freedom, and what each one means in practice.
3. What is MISSING that a standard agreement of this
type usually includes.
4. The specific questions I should ask the other side
before signing.
5. Flag anything where the answer depends on local law,
and say so rather than guessing.
Point 5 does most of the work. Left alone, a model fills a jurisdictional gap with a plausible guess. Told explicitly that "this depends on where you are" is an acceptable answer, it will often flag the thing you most needed a lawyer for.
The move that makes it worth real money
Run the output back through the question that matters: what would have to be true for this to be a bad deal for me?
Then take the two or three sharpest questions from step 4 and actually ask the other party — in writing, by email. This is the part people skip, and it's where the value is. A landlord or employer who answers clearly has told you something. One who gets evasive about a clause has told you considerably more, and you learned it before signing rather than after.
You've now used AI for the part it's good at — comprehension, structure, spotting gaps — and kept the part that needs a human, which is judgment about people and consequences. That's the same division as using AI to prepare for a doctor's appointment rather than to replace it.
When to just pay the lawyer
The honest boundary. Get a professional when: the money at stake is large relative to what you have; you're signing away ownership of something you made; there's a non-compete or an exclusivity term; it involves immigration or visa status; or you're being asked to sign quickly. Time pressure on a contract is a signal, not a coincidence. AI review makes you a much better-prepared client — it doesn't replace the person carrying professional liability for being wrong.
Keep your head:
Use AI to understand the document and to find what isn't in it. Don't use it to decide whether to sign, and don't trust it on local law or a citation. The most valuable output isn't the summary — it's the three questions you send the other side before you agree to anything.
Get one of these a week. Our free newsletter sends one genuinely useful AI habit and one judgment check every week — no hype, four-minute read. Subscribe on the home page.
Related: Is it safe to ask AI for financial advice? and What never to paste into AI.