When a Client Puts a No-AI Clause in the Contract

Do not sign it as written, and do not refuse it either. A clause reading "Supplier shall not use artificial intelligence in the performance of the Services" is unenforceable in spirit and dangerous in letter, because on a strict reading your IDE autocomplete, your spam filter and your spell checker all breach it. Find out which of four worries produced the sentence, then write a clause that addresses that worry specifically.

The clause is almost never about AI. It is about one of four things, and they have different answers.

The four worries behind the clause

Confidentiality. They do not want their data in someone else's training set. This is the most common and the most reasonable, and it has nothing to do with whether you use AI. It is about which AI, configured how.

Ownership. They are worried that AI-generated work cannot be copyrighted, or that its provenance is unclear, and they are buying something they need to own cleanly.

Quality. Someone in their organization believes AI-assisted work is worse work, possibly because they have seen a bad example. This is a judgment about output, dressed as a rule about process.

Policy compliance. Their own regulator, insurer, or enterprise client imposes a rule, and it flows down to you. The person you are talking to may not agree with it and cannot change it.

The reason to identify which one applies is that three of the four have a version you can accept. The fourth, quality, usually dissolves once you agree to be judged on output.

Ask the diagnostic question

One question does most of the work, ideally on a call rather than in email:

"Happy to work within that. So I write it correctly, is the concern about your data going to a third party, about who owns the deliverable, or is this a policy you have to apply?"

You are offering to comply and asking which compliance. Almost nobody refuses to answer that, and the answer tells you what to write. It is the same move as any good discovery conversation: find the requirement under the request.

If the answer is confidentiality, you are two sentences from agreement. If it is policy, ask to see the actual policy, which is often narrower than the clause someone drafted from it.

What to propose instead

Replace the blanket prohibition with commitments you can genuinely meet.

For confidentiality:

Supplier shall not submit Client Confidential Information to any third-party AI service that retains inputs for model training. Supplier may use AI tools operating under zero-retention or enterprise terms, and shall disclose on request which tools are used.

This gives them the protection they want and leaves you able to work. It is also verifiable, which a blanket ban is not.

For ownership:

Supplier warrants that all Deliverables are original, that Supplier holds or can transfer all rights necessary for Client to own them outright, and that no Deliverable infringes third-party rights.

Note that this says nothing about AI. It commits you to the outcome they care about, and outcome warranties are the correct instrument here. Your existing agreement probably already contains something like it.

For quality:

Deliverables shall meet the acceptance criteria in Schedule A. Client may reject work not meeting those criteria at no cost.

Again, no mention of AI. You have offered to be judged on the work, which is what they actually want.

For policy compliance: ask for the policy text and mirror its scope precisely. If the policy prohibits generative AI for customer-facing copy, the clause should say that and not more.

Where the danger actually is

Two traps are worth naming.

The definition is too broad. "Artificial intelligence" in a 2026 contract covers nearly every tool you own. If you sign a blanket ban, you are technically in breach from the moment you open your editor, which gives a client in a dispute a free lever. Always narrow the definition to something specific, such as generative AI systems that produce substantive portions of the Deliverables.

Warranting no AI in the supply chain. You cannot make this promise about your subcontractors, your stock asset providers, or the libraries you use. Scope any commitment to your own direct use.

If they hold firm on a genuine blanket ban

Some clients, usually for regulatory reasons, cannot move. Then it becomes a pricing and scoping question rather than a negotiation.

Work without AI assistance takes longer. Say so plainly, price it accordingly, and let them decide whether the constraint is worth the delta. This is the same conversation as any other constraint that costs time, and it goes better when framed as a schedule and cost consequence rather than a complaint. If they have already anchored on a number, the dynamics look a lot like a client asking for a discount because you used AI, from the opposite direction.

Get the acceptance criteria unusually tight in this case. Without AI in the loop, iteration is slower, and you want scope pinned.

Put it in your standard terms first

The best time to handle this is before it comes up. A short paragraph in your own agreement describing how you use AI tools, what data never leaves your environment, and what you warrant about the deliverable heads off most versions of this clause, because you have already answered the question.

Our template guidance on writing an AI usage policy for clients covers what that paragraph contains, and it pairs with the related provisions in an NDA for an AI project and the model change clause that handles what happens when a provider deprecates something mid-engagement.

Nothing here is legal advice. Run any clause you draft past someone qualified in your jurisdiction before it goes in a signed agreement. More on the commercial side in our guide to making money with AI.

FAQ

Should I just sign a no-AI clause to win the work?

Not as written. A broad definition makes routine tools a breach, which hands a client leverage in any later dispute. Narrow the definition and commit to what they actually care about.

What if the client will not tell me why they want it?

Propose the confidentiality version, since that is the most common driver, and say you are happy to adjust if the concern is different. A concrete proposal usually surfaces the real reason faster than another question.

Can I promise no AI was used anywhere in the work?

Only for your own direct use. You cannot warrant it for subcontractors, stock assets, or the tooling inside libraries you depend on, so scope any commitment to yourself.

Does using AI affect who owns the deliverable?

Ownership rules vary by jurisdiction and are still moving. The practical answer is to warrant the outcome, that the client owns the deliverable outright and it infringes nobody, rather than arguing about the process that produced it.