An agency sent over a 14-page master services agreement for a project worth a few thousand dollars. Or the landlord wants the lease signed by tomorrow to hold the unit. Or the offer you want came with thirty pages of non-compete and IP assignment language nobody walked you through.
AI helps with that as preparation. Used well, AI translates dense clauses into plain language, pulls the terms buried on page nine, and turns all of that into the questions worth asking. It cannot tell you whether a contract holds up where you live, how a clause would be enforced, or whether to sign. It also misses things without telling you it missed them.
You are not trying to replace a lawyer. You are trying to work out what you would be agreeing to, and whether this is one of the times you actually need a lawyer.
Below are the four passes of a practical AI contract review, the confidentiality check to run first, and the point where you stop and call someone.
Note: This guide is for people checking their own contracts, not for lawyers or other legal professionals. If you review contracts as part of your job, AI legal document review comes with professional duties and liability, and these guides are a better fit:
What AI can safely check on your contract, and what it can’t
The split runs between reading tasks and judgment tasks. Reading is what these tools do well; judgment is what you are paying a lawyer for.
| Part of the contract | What AI can do | What needs a lawyer | Risk if you skip that |
|---|---|---|---|
| Dense or archaic language | Translate it clause by clause | Nothing, once you check it against the text | Low |
| Dates, notice periods, payment terms | List every one with its section reference | Nothing, once you verify each | Low |
| Auto-renewal and termination | Extract the trigger, window, and method | Whether cancellation meets your state’s rules | You renew for a term you did not want |
| Liability caps and indemnity | Explain what it exposes you to | Whether the cap holds up, and your real exposure | Uncapped risk you priced as capped |
| Non-competes, IP assignment | Explain scope, duration, coverage | Whether it is enforceable where you live | Money or future work left on the table |
| Anything that comes across as unfavorable or unusual | Flag it and say why | Whether it is a real problem in your state | You negotiate the wrong thing, or nothing |
| Whether to sign | Nothing | Everything | This is the whole reason lawyers exist |
Check three things before you upload the document anywhere
This comes first because it is the only step you cannot undo. Once a document is in a chat window it sits in the vendor’s systems on the vendor’s terms, and deleting the conversation does not necessarily reverse that.
Google says so directly. Its Gemini Apps privacy documentation, updated in August 2026, tells users not to enter confidential information they would not want a human reviewer to see, and adds that chats picked up for human review survive deletion of your activity for up to three years.
Anthropic sets different terms: deleted Claude conversations leave back-end storage within 30 days on consumer plans, though data may sit in training pipelines for up to five years if training is on.
Three checks before you paste anything.
- Does the contract restrict this? Read the confidentiality clause. Many agreements limit disclosure to named parties or advisers, and a chatbot is neither.
- Does your employer’s policy allow it? If the document belongs to your company or a client, this is not your call alone. Our guide to whether it’s safe to use AI tools at work covers how approved-tool lists work.
- What does the tool do with the file? Check the data controls inside your account, not the marketing page. Switch off training, use a temporary chat, use a business tier if you have one. The walkthrough on whether ChatGPT is safe to use covers those settings.
If any check comes back wrong, retype the clause without names or amounts, or ask about the mechanism rather than the document.
Pass one: make the contract readable in plain English
When you use AI to read a contract, start with comprehension, the one thing these models are reliably good at. Turning legal English into ordinary English is a language task, and language is what they were built for.
Work clause by clause rather than asking for a summary. A summary compresses, and compression is where things go missing. The guide on how to use AI to summarize a PDF applies the same discipline to any long file.
Go through this section one clause at a time and tell me what each clause means in plain English. Don’t summarize it, translate it. If a sentence could be read two ways, say so instead of picking one.
Read the translation next to the original before you accept it. This pass also sets the trap in every later one.
In March 2026, the researchers behind the CLAUSE benchmark presented their findings at the EACL conference. When models explained problems in a contract, the explanations were easy to read but often left things out. So a clear answer doesn’t mean a complete one.
Pass two: pull out the terms that carry consequences
Since clear writing tells you nothing about coverage, the next pass asks for specifics you can count. The clauses that hurt people are buried rather than hidden, and retrieval is what these tools are for.
List every date, deadline, notice period, and renewal trigger in this contract. For each one, tell me what happens if I miss it and which section it comes from.
Asking for section numbers lets you check the answer. Go to each section it names and make sure the contract says what the AI claims. If you haven’t read the section yourself, don’t rely on the claim.
Auto-renewal deserves its own look, partly because it is the classic quiet cost. Check the notice window in your contract, and check your state.
Pass three: ask what looks unusual, and treat the answer as a question
Extraction tells you what is in the document, not whether any of it is reasonable. This pass is what people usually mean by AI contract analysis, and it is where reliability starts to fall. The model has a statistical impression of its training data, not a survey of what is standard in your industry, market, or state.
Compare this against what you’d normally expect in a contract of this type. Flag anything that looks one-sided or unusual, and separately tell me what you’d expect to see here that’s missing.
The second half of that prompt matters more than the first. Missing protections are harder to notice than bad ones, because there is nothing on the page to react to.
A clean result here is not evidence that the contract is clean. So check every flag against the clause it came from, then take it to the other side as a question rather than a verdict. “Why is the liability cap one-directional?” is a fair thing to ask, and costs you nothing if the answer is that it is standard for them.
Pass four: turn everything into your question list
End with a list of questions rather than conclusions, because questions are safer when the AI gets something wrong. If it flags a clause that turns out to be fine, the other side answers your question and you move on. If you had treated that flag as a conclusion, you might have signed, or refused to sign based on a mistake.
Based on everything in this document, write me the list of questions I should ask before I sign. Put the ones that affect money or how long I’m tied in at the top.
Then edit it yourself. Cut the questions you already know the answer to, and check the rest against the clauses they came from.
That list turns vague unease into specific asks, which are easier to say yes to than “I’m not comfortable with this.” It is also what makes a lawyer’s hour affordable, by replacing an open-ended review with a targeted conversation.
Four contracts people check themselves, and what to look for in each
Run the same four passes on every contract. What changes from one contract to the next is which terms can cost you money or tie you in, so that’s where to look closest.
Freelance or client agreement. Payment timing, kill fees and their triggers, who owns the work before you are paid, and whether liability is capped in both directions or only yours.
Residential lease. For the AI lease review, start with the deposit: the amount, the deductions allowed, and the deadline for its return. Then early-termination costs, repair obligations, and the notice period before renewal. All of that is regulated at the state level, so pass three produces questions for the landlord or a local tenant service, not conclusions.
Job offer. The vesting schedule, the cliff, what happens to unvested equity if you leave, and the notice period. Then the non-compete and the IP assignment, particularly whether the assignment reaches work you do outside the job.
Supplier or vendor terms. The renewal date, cancellation window, price escalation, and exit terms. Ask what happens to your data on the way out.
Where AI review goes wrong, and why you won’t notice
Use AI to review contracts a few times and you’ll see the same four problems. They all come from one habit: the model writes whatever sounds right, not what it has checked. And its answers sound most convincing on the legal points you’re least able to check yourself.
It misses things silently. A model that finds nothing in a section reports nothing, in the tone it uses when a section is genuinely clean. Silence is the failure you will never catch.
It invents references. The CLAUSE researchers found that even their best model matched legal citations correctly less than 14% of the time. The complaint in Nippon Life Insurance Company of America v. OpenAI, filed in March 2026, alleges that filings produced with ChatGPT cited a case that does not exist, Carr v. Gateway, Inc.
It does not know what has changed or where you are. Click-to-cancel and the FTC non-compete ban both moved between 2024 and 2026, so a model answering from an older snapshot describes protections you do not have.
It agrees with you. MIT and Penn State researchers reported in February 2026 that models grow more agreeable as a conversation runs longer and as they learn more about the user. Shomik Jain, who led the work, put it plainly: “If you are talking to a model for an extended period of time and start to outsource your thinking to it, you may find yourself in an echo chamber that you can’t escape.”
The last problem, agreeing with you, can do the most damage because it pushes you toward the answer you already wanted. If you disagree with something the AI flagged, it will often back down. Asking the same chat to double-check its answer doesn’t give you a real second opinion.
Get a lawyer when these factors show up
No dollar figure makes this decision for you. A $2,000 contract with a three-year lock-in and personal liability is a worse problem than a $40,000 project with clean termination terms. Look at the shape of the risk instead.
These are the factors I would use, drawn from the points above where general answers stop working.
- The term is long or hard to exit. Multi-year, auto-renewing, or with a notice window you could miss.
- Your personal assets are reachable. Personal guarantees, uncapped liability, a one-way indemnity.
- The other side wrote it and will not move. That says something about the relationship, not just the paperwork.
- The stakes are your livelihood. Your main client, your home, your next job.
If you’re skipping a lawyer because of the cost, don’t use AI to fill that gap. Look for low-cost or free legal help instead. The government site usa.gov/legal-aid lets you search by state for free and low-cost options, including the Legal Services Corporation, LawHelp.org, pro bono clinics at law schools, and the American Bar Association’s free legal answers service.
Paying a lawyer also costs less when you bring your question list. An hour spent on specific questions is cheaper than asking them to review the whole contract from scratch.
Build a repeatable AI reading workflow
Start with the contract in front of you. Run the confidentiality check, work through the four passes in order, and finish with a written question list. Each pass gives AI one job and produces an output you can check against the original document.
The same method applies beyond contracts. You can use it to translate dense language, extract facts, spot details that need attention, and prepare questions about an insurance renewal, policy document, or another set of terms.
If the contract matches any factor in the checklist above, take the original document and your question list to a lawyer. The lawyer can interpret the terms in the context of your situation and advise you on what to do next.
If you want to make AI part of how you handle routine work, Coursiv’s AI Certificate Program teaches structured, critical AI use through short lessons that fit into a busy schedule and include built-in practice. You finish with a certificate of completion.
The program develops AI workflow skills. It does not teach law or replace legal advice.