Can AI Draft a Contract? What It Does Well and Where It Fails

Yes, AI can draft a contract, and for a standard document like an NDA, a consulting agreement, or a basic service agreement, it can produce a usable first draft in under a minute. What it cannot reliably do is get Indian law right, hold your commercial positions consistently, or tell you when it has made something up. Those three gaps are the difference between a first draft and a contract you can sign. (Adira, which publishes this guide, builds contract drafting and review software, including tools that ground AI output in your own templates. This guide is written to be useful whether or not you ever use ours.)

Below is a plain look at what AI drafting is actually doing under the hood, where it earns its keep, where it fails in ways that matter in India specifically, and a test you can run on any AI-drafted clause before you trust it.

What an AI is actually doing when it "drafts" a contract

A large language model does not know contract law. It predicts the next most statistically likely word, given everything it was trained on and the prompt you gave it. Ask it to draft a non-compete clause and it is not reasoning from the Indian Contract Act, it is reproducing the pattern of non-compete clauses it saw most often in training, which is overwhelmingly US and UK material, because that is what exists in bulk on the public internet in English.

This one fact explains most of where AI drafting goes right and wrong. It is excellent at reproducing familiar structure and boilerplate. It is unreliable the moment the correct answer depends on facts specific to your deal, your company's positions, or a jurisdiction underrepresented in what it was trained on, which includes Indian contract law relative to US contract law.

Where AI drafting genuinely helps

Three things AI does well enough to use today, with a human reviewing the output:

Speed on a blank page. A structured first draft, sections in a sensible order (definitions, scope, payment, term, termination, boilerplate), in minutes instead of the hour or two it takes from scratch.

Standard, low-stakes documents. An NDA, a simple freelance agreement, a basic vendor purchase order. These are heavily templated categories where the gap between a good draft and a bad one is small.

Boilerplate. Notices, severability, entire agreement, and counterparts clauses are close to interchangeable across contracts, low risk even if the AI's version is generic.

None of this means the output is ready to sign. The AI has saved you the blank-page problem, not the judgment problem.

Where it fails: jurisdiction-specific law

This is the most consequential failure, and the clearest example is the non-compete clause. Ask a general-purpose AI model to draft a post-employment non-compete for an Indian employment contract, and it will often produce a US-style clause: enforceable-sounding language restraining the employee from joining a competitor for 12 to 24 months after they leave, because that is what "reasonable" non-competes look like in the US training data it learned from.

In India, that clause is void. Section 27 of the Indian Contract Act, 1872 says:

"Every agreement by which any one is restrained from exercising a lawful profession, trade or business of any kind, is to that extent void."

You can read the full text on Indian Kanoon. Unlike the US, where courts ask whether a non-compete is "reasonable" in time, geography, and scope, Indian law runs no such test for post-employment restraints. Section 27 voids them outright, with a narrow exception for the sale of business goodwill. Courts have applied this consistently, most recently in a Delhi High Court ruling that struck down an injunction sought against a departing employee under a post-termination restrictive covenant. We cover the statute, the case law, and how to spot a disguised non-compete in full in Are Non-Compete Clauses Enforceable in India?. The point here is narrower: an AI model asked to draft this clause with no further instruction hands you language that does not work in India, confidently, with no warning attached.

The same pattern shows up more subtly elsewhere: governing law defaulting to a US state, or an at-will termination clause assuming US employment law, which does not exist the same way under Indian labour law. This is not the AI being careless. It is doing exactly what it was built to do, predict the statistically likely pattern, without knowing that pattern belongs to a different legal system.

Where it fails: it can hallucinate clauses and citations, and Indian courts have started punishing this

A hallucination is the AI generating something that sounds plausible and specific, a section number, a case name, a defined term, that does not actually exist. This is a known, structural property of how these models generate text, and in 2026 it started showing up in real Indian litigation.

On 2 July 2026, the Supreme Court of India set aside orders of the National Company Law Tribunal and the National Company Law Appellate Tribunal in Pooja Ramesh Singh v. Jammu and Kashmir Bank Ltd. (2026 INSC 668), after finding that both forums had relied on AI-generated judgments that do not exist. The Court held that citing such unverified material is misconduct on the part of an advocate, and went further: a decision must be set aside even if "an iota" of fake or hallucinated material entered the reasoning, and it directed the Bar Council of India to frame guidelines on AI use in legal work. Around the same time, the Bombay High Court fined a litigant fifty thousand rupees for written submissions built around a judgment, Jyoti w/o Dinesh Tulsiani vs Elegant Associates, that neither the court nor its law clerks could trace anywhere.

This is not an India-only problem. In the US case Mata v. Avianca, Inc. (S.D.N.Y., 2023), two lawyers were sanctioned under Rule 11 after filing a brief citing cases ChatGPT had invented outright, complete with fabricated quotations, then compounding the error with fake excerpts when the other side could not find the cases. The mechanism is identical wherever it happens: the model generates a confident, specific, checkable-looking fact, and nobody checks it before it goes into a real document.

A drafting AI is exactly as capable of this as a legal-research AI, inventing a section number that does not say what it claims, or a stamp duty rate stated with total confidence and no basis. The output reads as authoritative, which is what makes it dangerous to a non-lawyer with no independent way to catch the error.

A test you can run right now: before you rely on any specific case name, section number, or cited "standard practice" in an AI-drafted contract, search for it directly, in quotes, on Indian Kanoon or a general search engine. If a case name returns nothing, or a section number does not say what the AI claims, treat the whole surrounding clause as unverified.

Where it fails: your commercial positions and defined-term consistency

Even when an AI draft is legally sound, it does not know your deal. A generic prompt produces a generic contract: a liability cap the model considers "market," a payment term of "net 30" because that is common in training data, neither favouring your side. If you are the vendor and always push for a liability cap equal to fees paid, and the AI hands you an uncapped clause instead, you will not notice unless you already know to look for it.

Defined-term consistency is a smaller but real failure mode. A long AI-generated draft will sometimes call the same concept by two names in different sections, "Confidential Information" in clause 4 and "Proprietary Information" in clause 9, because each section was generated somewhat independently. In a signed contract, that creates genuine ambiguity about whether the two terms mean the same thing, exactly the kind of dispute a court resolves by interpretation rather than by reading the plain words.

A test you can run: search your AI-drafted document for every capitalised defined term and count how many different phrasings describe what looks like the same concept. If "Confidential Information" appears next to "Proprietary Information" or "Deliverables" sits next to "Work Product" for what is clearly the same thing, that is a consistency pass you need to do before anything else.

How grounding the AI in your own documents helps, but does not solve everything

Most of what goes wrong above comes from one root cause: a generic AI, given a generic prompt, produces a generic draft. The fix that moves the needle is grounding, feeding the AI your own executed contracts, clause library, and negotiating positions, so it drafts from your approved wording instead of the statistical average of the internet. We cover how this works, and its own limits, in What Is a Company Legal Persona?

Grounding meaningfully fixes generic commercial positions and defined-term drift, if your source documents are themselves consistent. It does not fully fix jurisdiction-specific law unless your source corpus was already India-correct, and it does not eliminate hallucination risk, a model can still generate a citation that was never in your source documents. Grounding narrows the gap. It does not close it.

Red flags in an AI-drafted contract

Normal	Red flag	Why it matters
Clause references a section number you can verify on Indian Kanoon or India Code	Clause cites a specific case name or judgment you cannot find anywhere	Fabricated citations read as confidently as real ones; unverifiable case names are the clearest hallucination signal
Non-compete, if present, is scoped to the employment or engagement period only	Non-compete restrains the person for months "following termination"	Void under Section 27 of the Indian Contract Act, regardless of how reasonable the time period sounds
Governing law and courts named match where the parties actually operate	Governing law defaults to Delaware, California, or another US state with no connection to either party	AI models default to the jurisdiction most common in their training data, not your deal
A defined term is used the same way every time it appears	The same concept is called two different names in different clauses	Creates real interpretive ambiguity in a signed contract; courts have to guess which term controls
Numbers (fees, notice periods, caps) trace back to something you told the AI	A specific number appears that you never gave the AI and cannot explain	The model filled a gap with a statistically common value, not your actual deal term
Stamp duty or filing fees are flagged as "confirm for your state"	A specific stamp duty rate or government fee is stated with confidence	These vary by state and instrument in India; a confident wrong number is worse than an honest gap
The draft is clearly labelled as a first draft needing review	The draft looks final and polished, ready to sign	Polished formatting invites false confidence; a rough-looking draft gets reviewed harder, which is what it needs

A clause AI commonly gets wrong, rewritten correctly

What an AI often produces, unprompted, for an Indian consulting agreement:

"For a period of eighteen (18) months following the termination of this Agreement for any reason, the Consultant shall not, directly or indirectly, provide services to, or engage in, any business that competes with the Company within the territory in which the Company operates."

This is a textbook US-pattern non-compete. It restrains the Consultant after the engagement ends, which puts it squarely inside what Section 27 voids, whatever the eighteen months or the territory wording implies.

A version corrected for India:

"During the term of this Agreement, the Consultant shall not, without the Company's prior written consent, provide services to any business that directly competes with the Company. This restriction ends on the termination or expiry of this Agreement and does not apply afterward. The Consultant's confidentiality obligations under Clause 9 survive termination for a period of 24 months and are unaffected by this Clause."

What changed: the restraint now operates only during the engagement, which Indian courts treat as ordinary exclusivity, not restraint of trade. The genuine post-termination protection the company needs, confidentiality, is pulled into its own clause with its own survival period, instead of being smuggled inside a non-compete that cannot legally do that job in India.

When AI drafting is enough, and when you need a lawyer

A useful test: ask whether this is a template deal or a decision. A standard NDA on your own paper, a low-value freelance agreement, a vendor purchase order under a policy you already have, these are template deals, where a reviewed AI draft is usually enough. A first-time joint venture, a shareholders' agreement, anything in a regulated sector, or a deal above the size where a mistake would hurt, these are decisions, and an AI draft is a starting point for a lawyer, not a substitute for one.

If you want to sanity-check a clause yourself first, you can mark it up for free in Weave, Adira's free browser-based contract tool, and flag exactly the kind of post-termination or unverifiable-citation language covered above.

US and global contrast

The gap between AI drafting and Indian law is sharper than the equivalent gap in the US, precisely because training data over-represents US contract norms. A US-trained model drafting a US non-compete draws on the correct legal tradition, even though individual states vary and California bans most post-employment non-competes outright. The same model drafting for India applies the wrong tradition by default, unless instructed otherwise.

Hallucination, by contrast, is not a jurisdiction problem, it is a model problem. Mata v. Avianca in the US and the Pooja Ramesh Singh ruling in India are the same failure mode in two legal systems: a fabricated citation, caught only because someone checked. Wherever you draft, verify every specific, checkable claim an AI hands you before it goes into a document with your name on it.

FAQ

Can AI draft a legally valid contract in India? It can draft something signable, but "valid" and "enforceable exactly as drafted" are different questions. A draft can contain clauses, like a post-termination non-compete, that are void under Indian law even though the document's form is fine. Have a human check enforceability before signing anything with real stakes.

Is it safe to use ChatGPT or a similar tool to draft a contract? Safe as a first-draft tool, the same way a template is a starting point. Not safe as a final source of truth, particularly for citations, section numbers, or India-specific enforceability, without independently checking its claims.

Will an AI-drafted non-compete clause hold up in India? If it restrains the other party after the relationship ends, generally no. Section 27 of the Indian Contract Act voids post-employment and post-engagement non-competes, and an AI trained mostly on US patterns will often draft exactly this kind of unenforceable clause unless told not to.

Can AI cite real Indian case law accurately? Sometimes, but not reliably enough to trust without checking. Indian courts, including the Supreme Court in the July 2026 Pooja Ramesh Singh ruling, have now formally addressed AI-generated citations that turned out not to exist. Treat every case name an AI gives you as a lead to verify, not a fact to cite.

How does grounding AI in my own templates fix this? It fixes the parts caused by genericness, since the AI drafts from your actual clause wording instead of an internet average. It does not by itself fix jurisdiction-specific law unless your source templates were already correct for India, and it does not remove the need to verify any citation the AI produces.

---

This guide explains what AI contract drafting generally does well and where it generally fails. It does not tell you whether a specific AI-drafted clause in your specific contract is enforceable, or safe to sign. For that, especially for anything above a low-stakes, standard document, have a lawyer review the draft before you sign it.