How to review a contract with AI before you sign it (and exactly where to stop)

TL;DR: Use AI to translate the contract into plain English, surface the asymmetries, and build a list of questions. Do not use it to decide whether to sign, whether a clause is enforceable where you live, or what a court would do. It is a preparation tool. For anything high-value, hard to reverse, or already in dispute, that preparation goes to a lawyer.

What this is good for

Most people signing a contract are not choosing between AI and a lawyer. They are choosing between AI and skimming it. That is the honest comparison, and against that baseline this is a large improvement: AI reliably catches plain-English meaning, one-sided clauses, auto-renewals and deadlines, and the terms that are confusing on purpose.

Where it fails is equally well documented. It does not know whether a clause is enforceable in your jurisdiction, it frequently fails to notice what is missing from an agreement, and it can sound authoritative while being wrong.

So: use it to understand and prepare. Then decide whether the stakes justify a professional.

Step 1: Get clean text

Photographed pages and scanned PDFs produce transcription errors, and a misread “not” reverses a clause. Use the original digital file where possible. If you only have a scan, run OCR first and spot-check the defined terms and any numbers — dates, amounts, notice periods.

Upload the whole agreement including schedules, annexes and anything incorporated by reference. Schedules are where the unpleasant terms usually live, precisely because nobody reads them.

Step 2: Plain-English translation, clause by clause

“Translate this contract into plain English, clause by clause. For each clause give me: the clause number, what it means in one or two sentences a non-lawyer would understand, and who it protects — me, them, or both. Do not summarise the document as a whole. Do not skip clauses that seem boilerplate. Where a clause refers to a defined term, expand the definition inline the first time.”

Read all of it. This is the least glamorous step and the one that produces the most “wait, what?” moments.

Step 3: The asymmetry pass

This is where the real value is:

“Identify every provision in this contract that is asymmetric — where one party has a right, remedy, or protection that the other does not. For each: quote the clause, explain the imbalance concretely, and describe a specific realistic scenario in which it would hurt me. Then tell me what a balanced version of that clause would typically look like.

Pay particular attention to: termination rights, notice periods, liability caps and exclusions, indemnities, auto-renewal, price escalation, IP ownership, confidentiality duration, dispute resolution and governing law, assignment, and anything about data.”

Ask for scenarios rather than abstractions. “The liability cap is one month’s fees” means little; “if their error takes your service down for a week and costs you $40,000, you can recover $2,000” is a decision-grade sentence.

Step 4: Explicitly hunt for what is absent

Models are much better at analysing text in front of them than noticing gaps, so you have to ask directly:

“For a contract of this type, list the protections and provisions that are commonly included but are ABSENT here. For each, explain what its absence means for me in practice, and how likely it is to matter. Be specific about the type of agreement this is.”

Then treat the output as a checklist to verify, not as fact — this is exactly the area where a confident answer can be wrong, since “commonly included” varies enormously by jurisdiction and industry.

Step 5: Dates, numbers and obligations

“Extract every date, deadline, notice period, monetary amount, and recurring obligation in this contract into a table: what it is, the exact clause, the deadline or amount, what happens if it is missed, and whether the obligation is mine or theirs. Include anything that triggers automatically without notice.”

Then verify each one against the original text yourself and put the deadlines in your calendar with a reminder well before the date. Auto-renewal that required 90 days’ notice, discovered on day 60, is one of the most common and most avoidable commercial injuries there is.

Step 6: Build your question list

“Based on everything above, give me the ten questions I should ask the other party before signing, ordered by how much money or risk each represents. For each, phrase it as I would say it in a normal business conversation — collaborative, not adversarial — and tell me what a reasonable answer sounds like versus an evasive one.”

Most negotiation leverage is simply asking. A surprising number of one-sided clauses are standard-template artefacts that the other side will soften without much argument, because they never expected anyone to read that far.

A worked example: the SaaS agreement that renews itself

A short software contract, the kind a small business signs without much thought. The plain-English pass returns twenty-odd unremarkable clauses. The asymmetry pass returns this:

Clause 11.2 (Term). “This Agreement shall automatically renew for successive periods of twelve (12) months unless either party provides written notice of non-renewal not less than ninety (90) days prior to the end of the then-current term.”

Plain English: it renews for another full year automatically, and your window to stop that closes three months before the anniversary. Miss it by a day and you owe another twelve months.

Asymmetry: none on its face — it binds both parties equally. But read it alongside clause 6.4, which lets them increase fees on 30 days’ notice. Combined, the position is: they can raise the price 60 days after your cancellation window has already shut, and you are contractually committed for a year at the new rate.

Neither clause is unusual and neither is hidden. The problem is that the interaction between them is where the risk lives, and reading clause by clause — which is how humans read contracts when they read them at all — does not surface interactions.

The specific prompt that finds this:

“Identify pairs or groups of clauses in this contract that interact in ways that are not obvious from reading either one alone. For each interaction: quote both clauses, explain the combined effect in plain English, and give a concrete scenario showing how it plays out. Focus on combinations involving notice periods, price changes, termination rights and liability.”

The practical remedy is unglamorous: put a calendar reminder 120 days before the anniversary, and ask for a cap on mid-term increases before signing. Both are easy to get if you ask before signature and impossible after.

Where to stop — and this part is not optional

Take the output to a qualified lawyer, not a chatbot, when any of these apply:

• The value is material to you. The threshold is personal, but “material” means an amount whose loss would genuinely hurt.
• It is hard to reverse — multi-year terms, exclusivity, IP assignment, personal guarantees, anything touching your home or your equity.
• There is an existing dispute. Once there is conflict, you need advice about your specific position, not a general reading of the document.
• Employment terms with restrictions. Non-competes and similar clauses vary wildly by jurisdiction — the same words can be routine in one place and unenforceable in another. A model that does not know where you live cannot tell you which.
• Regulated activity — anything involving financial services, healthcare, immigration status, or licensing.

Bringing a lawyer a marked-up contract and ten specific questions is dramatically cheaper than bringing them a PDF and asking what they think. That is the actual economic benefit here: not replacing the professional, but arriving prepared enough to use an hour of their time well.

Four failure modes to expect

1. Confident wrongness on enforceability. The single most dangerous output. Any sentence about what a court would do should be treated as unverified.
2. Hallucinated clause references. Models cite clause numbers that do not exist or attribute language to the wrong section — a well-known failure pattern, see the hallucination explainer. Check every citation against the document.
3. Missing incorporated documents. If the contract references a separate policy, standard terms, or a schedule you did not upload, the analysis is incomplete and the model will not tell you.
4. Confidentiality. Many contracts include confidentiality provisions covering the contract itself. Use a service tier with training disabled, and consider whether uploading is permitted at all.

None of this is legal advice — I am not a lawyer, and neither is the model. It is a method for understanding a document well enough to ask better questions. For contracts you review repeatedly with the same concerns, building a reusable checklist as a system prompt makes each subsequent review considerably faster.