How to Check a Contract for Risks with AI: A Pre-Lawyer Filter

Someone sends you a twelve-page contract and wants a signature by Friday. Your lawyer is busy, a consultation costs real money, and signing blind feels wrong. This is exactly where AI earns its keep: not as a replacement for the lawyer, but as a first-pass filter. In ten minutes it maps out where the unpleasant clauses hide, so you walk into the lawyer's office with specific questions instead of "please look at everything."

Let's be honest up front: AI does not replace a lawyer. Whatever the model finds is a draft analysis, not a legal opinion. But that draft saves hours and catches the obvious stuff. Here is the step-by-step process using Claude.

Step 1. Privacy: clean the file before uploading

The contract travels to a third-party server, so first strip out everything the risk analysis does not need:

• ID numbers, tax IDs, bank account and card numbers;
• addresses and phone numbers;
• people's names: replace them with "Client" and "Contractor";
• the other side's trade secrets, if any ended up in the appendices.

The analysis loses nothing: risks live in the wording, not in the payment details. Separately, check whether the contract itself, or an NDA you already signed, forbids sharing the text with third parties or services. If it does, don't upload it, period.

Step 2. Upload the contract and give Claude a role

Attach the file to the chat (Claude reads PDF and DOCX directly) and write a prompt with a role and a risk checklist. The role matters: "review this contract" produces a summary, while a lawyer role on your side produces an analysis. Example:

``` You are a contract lawyer working on my side. I am the contractor; this is a website development agreement. Review the text against the checklist below. For every item, quote the exact wording from the contract and cite the clause number:

1. Payment and prepayment: when do I get paid, is a delay built in.
2. Liability: penalties and fines, whether their size is capped.
3. Jurisdiction: which court resolves disputes.
4. Auto-renewal: does the contract renew itself, how do I exit.
5. Acceptance: how the client accepts the work, can they stall forever.
6. Unilateral termination: who can walk away and on what terms.

Finish with the three clauses most dangerous for me and explain exactly why. ```

Swap in your own role (client or contractor) and the subject of the deal. The demand for exact quotes is non-negotiable: without them you cannot verify a single claim against the original.

Step 3. Ask cross-examination questions

The first answer is a map; now dig into specific spots:

• "What happens if I miss the deadline by a week? Calculate the penalty using the formulas in the contract."
• "Which clause will be hardest for me to actually comply with?"
• "Do any clauses contradict each other?"
• "What is unusual in this contract for this kind of service?"

Cross-questions surface what a checklist misses: clause combinations. A 30-day acceptance window looks harmless on its own, but paired with "payment after the acceptance act is signed" it quietly becomes a one-month payment delay.

Step 4. Verify the wording by hand

Now open the original and check every quote from the answer:

• the quote actually exists in the text, word for word;
• clause numbers match;
• amounts, percentages, and deadlines match the original to the digit;
• the parties are not swapped: "the contractor pays the penalty" and "the client pays the penalty" look similar but cost you very different amounts.

This takes about fifteen minutes. Skip it and the whole exercise collapses: the document has to confirm the conclusions, not the confident tone of the answer.

Where AI gets contracts wrong and what to check manually

Failure modes worth knowing in advance:

• Swapped parties. Especially in contracts that say "Party 1" and "Party 2" instead of naming the roles.
• Invented legal references. A statute number in the answer can be wrong. Verify laws against official sources, not through the model.
• Lost appendices. If the penalties sit in Appendix 3, the model may analyze only the main body. Upload the whole package, appendices included.
• No feel for court practice. A clause can look scary on paper while courts stopped enforcing it years ago, or the reverse. This is precisely the territory of a licensed human.
• Smoothing things over. Ask "is this contract fine?" and the model leans toward "mostly yes." A checklist and pointed questions about specific risks work better than a request for a general verdict.

Contracts you should not filter this way

The AI filter suits routine deals: services, contracting work, supply, office leases, software licenses. Some categories need a lawyer from step one, not after:

• real estate and anything that requires registration;
• loans, collateral, personal guarantees;
• investment and corporate agreements, equity, options;
• any contract where a mistake would genuinely hurt you, whatever its type.

Simple rule: the more painful it would be to lose money on this deal, the earlier a licensed professional should enter the process.

Wrapping up

The routine: clean the data, upload, set the role and checklist, cross-examine, verify quotes against the original. You end up with a list of specific risks to bring to your lawyer or to the negotiation table. That beats reading twelve pages yourself and hoping it "looks standard."

If you review contracts regularly, save the role and checklist into a project so you never paste them twice: the mechanics are covered in our guide to Projects and context. More things a business can hand over to Claude, from reports to spreadsheets, are collected in Claude for business.

And if you want to build this skill hands-on, AGINE Academy teaches Claude as a game: you run missions in the real Claude and leave every lesson with a working result, with an AI mentor beside you. The full program is 81 lessons across 12 blocks, from setup to Projects, skills, and Claude Code. The starter block of 4 lessons is free with no signup: take your own contract and run this article's playbook today.