How to Review a Contract With AI: A Workflow That Survives Legal Review

Most people who say they "reviewed a contract with AI" did one of two things: pasted the document into ChatGPT and asked "is this okay to sign?", or ran it through a legal-AI tool and trusted whatever came back because it looked thorough. Both produce an answer. Neither produces a review you could defend if someone later asked "how did you check this?" This guide is the workflow in between, structured enough to catch what matters, honest enough to tell you when to stop trusting the machine. (Adira, which publishes this guide, sells contract review and CLM software, so we have a commercial reason to want you comfortable with AI-assisted review. The workflow below is tool-agnostic and works with Adira, a general chatbot, or nothing paid at all.)

Step 1: Ask clause by clause, never "is this contract okay?"

"Is this contract okay?" is not a request an AI model can answer responsibly, because "okay" depends on facts it does not have: your risk appetite, your leverage, whether this is a one-off vendor or a five-year lock-in. Asked that way, most models answer anyway, with false confidence, because refusing feels unhelpful.

A structured ask fixes this. Go clause by clause, or category by category (parties, term, payment, liability, indemnity, IP, termination, dispute resolution), and for each ask three things: what does it say in plain English, what does it require of each side, and does it deviate from what is normal for a contract of this type. "Normal" is doing real work in that third question, which is why Step 2 matters.

Step 2: Give the AI your position, so it flags deviations, not generic issues

A generic AI review of a services agreement will tell you, correctly but uselessly, that "the indemnity clause is one-sided" or "the liability cap could be higher." You already suspected that. What you need is: one-sided compared to what, and by how much.

The fix is to give the AI your playbook before asking it to review anything, your standard position on the clauses that matter, stated as a rule. If you are the vendor and your position is "liability capped at 12 months' fees, indemnity excludes indirect and consequential loss, payment terms no longer than 30 days," state that up front and ask it to check the draft against those positions and flag every deviation, not every clause that merely exists. A model told "flag anything longer than a 30-day payment term" will catch a buried "60 days from receipt of a correctly issued invoice" far more reliably than one asked to generally assess the payment clause, because you have turned a judgement call into a comparison it can run.

Step 3: Demand an exact-sentence citation for every flag

This is the single highest-leverage instruction in the workflow. For every issue the AI raises, require it to quote the exact sentence the flag is based on, with the clause number if there is one, and say "not stated" rather than paraphrase when it cannot find a specific sentence.

This changes the failure mode from silent to visible. A vague flag ("the termination clause seems to favour the other party") gives you nothing to check. A quoted one ("Clause 11.2: 'Client may terminate for convenience on 90 days' notice; Vendor may terminate only for uncured material breach.'") gives you a sentence you can Ctrl+F in the source document in five seconds. If the quote is not there, you have caught a hallucination before it cost you anything.

Step 4: Run a human verification pass on the clauses that can actually hurt you

Not every clause deserves the same scrutiny. After the AI pass, spend your own time on the handful where being wrong is expensive: indemnity and its cap, limitation of liability, IP ownership and assignment, termination triggers and notice periods, and, for India specifically, governing law, dispute resolution, and whether the document is correctly stamped. Read the AI's quoted sentence against the actual clause, not just its summary. A model can quote a real sentence correctly and still get the significance wrong, especially where Indian statute changes what a clause means, which the next section covers.

Step 5: Never let AI redline the final version alone, for anything with real stakes

There is a difference between AI helping you understand and mark up a draft, which it is genuinely good at, and AI generating the redline that goes back to the other side on a contract with real money attached. For a low-stakes agreement, a personally-checked AI redline is usually fine. For anything with meaningful liability, IP, employment, or a long lock-in, a lawyer should see the redline first, because the AI does not know what you would actually concede in negotiation, and does not know Indian law as well as it sounds like it does.

What AI reliably misses about Indian contracts

General-purpose AI models are trained overwhelmingly on US and UK contracts and commentary. That produces a predictable blind spot: the model applies common-law defaults that are wrong once an Indian statute overrides them, and it will not warn you, since it does not know the override exists unless told. Three examples are worth checking every time, each a place where "the clause is silent" means something different in India.

Non-compete clauses. A US-trained model will often assess a post-employment non-compete on reasonableness, duration, geography, scope, because that is how US courts (in states that allow them at all) analyse them. Under Indian law, that analysis mostly does not apply. Section 27 of the Indian Contract Act, 1872 is direct:

"Every agreement by which any one is restrained from exercising a lawful profession, trade or business of any kind, is to that extent void." Source: Section 27, Indian Contract Act, 1872

The only real carve-out is for someone who sells the goodwill of a business agreeing not to compete with the buyer, within reasonable limits. A post-employment non-compete is, on the weight of Indian authority, void to that extent. An AI reviewing an employment contract without being told this will happily debate whether a two-year, pan-India non-compete is "reasonable," when the more useful flag is that it is very likely unenforceable as drafted.

IP assignment clauses silent on duration or territory. This catches people who assume silence means "forever, everywhere," which is roughly how a US-trained model reads it. Indian copyright law has an explicit default for this exact gap. Sections 19(5) and 19(6) of the Copyright Act, 1957 state:

"If the period of assignment is not stated, it shall be deemed to be five years from the date of assignment." "If the territorial extent of assignment of the rights is not specified, it shall be presumed to extend within India." Source: Section 19, Copyright Act, 1957

An IP assignment that simply says "the Contractor assigns all rights in the Work to the Client," with no stated period or territory, does not assign those rights permanently and globally by default under Indian law. It reverts to a five-year, India-only assignment. For a client commissioning software or content meant to be owned outright and worldwide, that gap is a real defect, and exactly what a generic AI pass will not catch unless asked to check for both a period and a territorial extent.

Stamping. This has no real equivalent in Western contract practice, so a general model usually has no reason to raise it at all. Under Section 35 of the Indian Stamp Act, 1899, an instrument chargeable with duty that is not duly stamped cannot be used as evidence:

"No instrument chargeable with duty shall be admitted in evidence for any purpose by any person having by law or consent of parties authority to receive evidence... unless such instrument is duly stamped." Source: Section 35, Indian Stamp Act, 1899

This matters even for arbitration clauses. In In Re: Interplay Between Arbitration Agreements under the Arbitration and Conciliation Act, 1996 and the Indian Stamp Act, 1899 (2023 INSC 1066), a seven-judge Supreme Court bench held that an unstamped arbitration agreement is not void, but the defect is curable, and the instrument stays inadmissible until the duty is paid. A generic AI review will read your dispute-resolution clause and call it fine. It will not tell you the document might be under-stamped for the state it was signed in, since that is not a question a US or UK contract ever raises.

The test that catches a fabricated flag: ask it to find something you know isn't there

Here is a test worth running once on any AI tool before you rely on it, and worth repeating occasionally after. Pick a clause you are certain is genuinely absent, a specific IP indemnity, a most-favoured-customer clause, anything you have confirmed by reading the document yourself is simply not there. Then ask the AI, directly, "does this contract contain a [that clause]? Quote the relevant section."

Two outcomes matter. If it correctly says the clause is not present, that is a good sign for this session. If it "finds" one anyway, quoting something plausible-sounding that does not address what you asked, or inventing a clause number and sentence not in the document, you have caught it hallucinating on command. Treat everything else from that session with suspicion: a model that fabricates on demand will do the same unprompted, just less predictably.

Signs your AI review pass is trustworthy, and signs it isn't

Normal	Red flag	Why it matters
Every flag comes with a quoted sentence and clause number	Flags are stated as summaries, with no locatable quote	You cannot verify a paraphrase; only a quote can be checked
It says "not stated" when a playbook item has no matching clause	It states a specific number or period for something the document is silent on	A confident specific for a genuine silence is the classic shape of a hallucination
Flags reference your stated position ("deviates from your 30-day term")	Flags are generic ("this clause could be improved")	Generic flags were never actually checked against a standard
Indian-specific issues (non-compete, IP assignment defaults, stamping) get raised unprompted	These never come up unless explicitly asked for, even on a contract for use in India	The model is applying US or UK defaults by default
Passes the "find something absent" test	Fails it, by inventing a clause that is not there	Fabricating on a direct test means it fabricates elsewhere, less visibly

A bad review prompt vs a better one

Bad: "Here is our vendor agreement. Please review it and tell me if there are any issues."

What is wrong: no stated position, no requirement to cite the source sentence, and "any issues" invites the model to surface whatever is easiest to describe rather than what is actually risky for your side.

Better: "You are reviewing this contract for [our role, e.g. the vendor]. Our position: liability capped at 12 months' fees, indemnity excludes indirect and consequential loss, payment terms of 30 days. Go clause by clause. For each: (1) what it says in plain English, (2) whether it matches or deviates from our position, quoting the exact sentence, (3) if there is no matching clause, say 'not present,' do not infer one. Separately, check whether this contract, used in India, contains a post-employment non-compete, and whether any IP assignment states both a period and a territorial extent."

What changed: a stated position to compare against, a required citation for every point, an instruction to say "not present" instead of guessing, and a direct prompt for the two Indian-specific gaps a generic pass would miss.

Keeping a confidential contract out of AI you don't control

Before any of the above, a separate question needs an honest answer: should this specific contract go into this specific AI tool at all? Under the Digital Personal Data Protection Act, 2023, a company that decides why and how personal data is processed is a "Data Fiduciary," and Section 8(2) requires any processor it uses to be engaged "only under a valid contract":

"A Data Fiduciary may engage, appoint, use or otherwise involve a Data Processor to process personal data on its behalf for any activity related to offering of goods or services to Data Principals only under a valid contract." Source: Section 8, Digital Personal Data Protection Act, 2023

Most commercial contracts carry personal data (signatory names, emails, sometimes salary details), and a fair few carry your own playbook too, the pricing floors you fed the AI in Step 2. Pasting either into a consumer AI account with no data processing agreement, without checking the document's own confidentiality clause, is a real exposure. Before uploading anything sensitive, check the tool's data controls (training opt-out, a non-retained session mode), and for a genuinely confidential draft, do the Step 4 verification locally instead: you can mark up and flag clauses in the browser using Weave without uploading anything. For the fuller privacy answer, see Is It Safe to Upload a Contract to AI?

For the structured prompt from Steps 1 to 3 as a copy-paste block, with variations for buyer, seller, employee, and investor positions, see our master prompt template for reviewing a contract with AI. For how reliable this approach is across tools, see Is AI Contract Review Accurate?

FAQ

Can AI actually replace a lawyer for contract review? No. Treat it as fast, structured triage that tells you what to look at closely, not as the review itself, and act on its raw output directly only if the stakes are genuinely low.

Is it better to use a general AI tool like ChatGPT or a dedicated contract-review tool? Either works if you follow this workflow. A dedicated tool may make Steps 1 and 3 faster, but does not remove the need for Step 4, and neither kind reliably knows Indian statutory overrides unless prompted.

Why does the AI keep missing that non-competes are largely unenforceable in India? Its training is mostly US and UK contract law, where non-competes are judged on reasonableness rather than presumptively void. Tell it Section 27 of the Indian Contract Act, 1872 applies, or it defaults to the wrong test.

What should I do if the AI quotes a sentence that isn't in my document? Treat the flag as unverified, and the rest of that session's output with more suspicion. Run the "find something absent" test above; if that fails too, start a fresh session.

Does giving the AI my negotiating position, like a liability cap floor, create a privacy risk? Yes, more than a plain summary request, since you are feeding it your commercial limits, not just the document. Use a tool with training switched off, and see the privacy section above first.

How do I know when a clause needs a lawyer rather than another AI pass? As a rule: anything touching indemnity caps, IP ownership, non-compete terms, or governing law and dispute forum, on a contract with real money or a long lock-in, should get a lawyer's eyes first, however clean the AI review looked.

This workflow gets you a structured, checkable AI review, one where every flag traces to a real sentence and the common India-specific gaps get checked deliberately instead of by luck. It does not tell you whether a clause in your contract is enforceable, or whether the contract is safe to sign, that depends on facts a lawyer needs to see, and none of this is legal advice. Talk to a lawyer before relying on an AI-assisted review for a decision with real money attached.