When a client contract forbids AI tools
Blanket no-AI clauses are usually aimed at a risk you are not creating. A decision tree for complying, negotiating a scoped exception, or declining.
A no-AI clause arrives and the instinct is to treat it as a verdict on your production method. It almost never is. It is usually a procurement template written by someone protecting against a specific incident they read about, applied to your statement of work because it is applied to every statement of work.
The useful move is to read the clause literally before deciding anything, because four very different prohibitions get written in language that looks identical from a distance, and three of them have a workable answer.
Four clauses that look like one clause
| Shape | Typical wording | What it actually restricts |
|---|---|---|
| Tool ban | "Supplier shall not use generative AI in performance of the Services" | Your process, everywhere, including drafting and research |
| Output ban | "No Deliverable shall contain AI-generated content" | The final asset only |
| Data ban | "Client Confidential Information shall not be submitted to any AI system" | Where client material goes, not what you make |
| Consent gate | "Supplier shall not use AI without prior written approval" | Nothing yet; it is a process requirement |
These need four different responses, and mixing them up is how a solvable clause becomes a lost engagement.
A data ban is the easiest and the most common in regulated sectors. It says nothing about generative production and everything about where the client's material travels. If you can generate original material without ingesting client confidential information, you are already compliant, and the correct response is a short note confirming that rather than a negotiation.
An output ban is narrower than it reads. It restricts the delivered asset, not your research, ideation, storyboarding or internal drafting. Whether you can work within it depends on what fraction of your value is generated pixels versus everything around them.
A consent gate is not a prohibition at all. It is an invitation to write the scoped exception yourself, and clients using this drafting are usually the easiest to agree with, because they have already decided the answer is conditional rather than no.
A tool ban is the only genuinely hard one, and even then the question is whether the drafter meant it as broadly as they wrote it. Very few clients intend to prohibit a transcript summariser. They mean the pixels.
What the ban is aimed at
Ask what incident the clause is guarding against, because the answer determines whether you can address it.
Confidential material leaking into training data. The most common real concern, and the most tractable. It is about data handling, not creative method, and it is answered by a written training exclusion plus a description of where material actually goes. If the client's worry is unreleased product photography ending up in someone's training corpus, a clause saying that cannot happen is more responsive than a promise to avoid AI entirely.
Provenance and regulatory exposure. Larger clients with EU exposure are managing machine-readable disclosure obligations and would rather not have assets in the pipeline that complicate it. This is answerable with a provenance and disclosure commitment. It is a documentation problem, not a prohibition problem.
Audience backlash. This is the one that gets least airtime in the clause and most airtime in the room. Procurement rarely writes this one down, but the clause often exists because a senior stakeholder watched a large brand take public criticism for how a campaign was made rather than for the campaign itself. The data underneath the worry is real: IAB and Sonata Insights research published in January 2026 put advertising executives 37 points apart from Gen Z and millennial consumers on AI-generated ads, a gap that widened from 32 points in 2024. A brand-safety clause aimed at that is not irrational and it is not really about your tooling. It is about whether the finished work reads as synthetic.
That last one is the most important distinction here, because it is the one the clause cannot actually fix. If the concern is perception, the answer is craft and human authorship in the finish, not tool avoidance in the process. Which means it is answerable inside a contract: a human-authorship commitment plus a review gate addresses the real worry, where a blanket prohibition only addresses the appearance of it. "We don't use AI" is only defensible if it is true.
The decision tree
Work through it in order. Most engagements resolve at step two or three.
Does the clause reach this work? Check the definitions. Many prohibitions attach only to "Deliverables" or "Creative Materials published under the Client's marks," which excludes internal drafts, research and pitch material. If it does not reach your work, note it in writing and proceed.
Which of the four shapes is it? Data ban and consent gate resolve without renegotiation in most cases. Output ban and tool ban continue to step three.
Can you deliver at an acceptable margin under the clause as written? Be honest before negotiating anything. If yes, comply, say so in the proposal, and price accordingly, because a compliant version of the same scope is usually a different price. If no, continue.
Is the underlying concern addressable? Data handling, provenance and disclosure all are. Perception risk is addressable with a human-finish commitment and a review gate. If the concern is a regulated-sector policy the client cannot vary, it is not addressable by you, and you should stop trying.
Propose a scoped exception, once, in writing. One proposal, specific, with the controls attached. Not a debate.
If the exception is declined, decline or requote. Both are respectable. What is not respectable is signing the clause and working around it.
That last line is the whole point of the exercise. The failure mode that ends relationships is not a no-AI clause. It is a supplier who signed one and then used the tools anyway, discovered later by an ad platform flagging generated creative or a synthetic-media label appearing on the client's own post. What happens when ad review flags generated creative is the mechanism by which that discovery usually happens, and it happens after the asset is live.
The scoped exception that gets signed
A scoped exception works when it trades a broad prohibition for narrow, verifiable controls. Vague reassurance does not clear procurement; specifics do.
Permitted use. Supplier may use generative tools for [ideation, reference imagery, background elements, voice-over drafting, and rough-cut assembly]. Supplier will not use generative tools to produce [final talent likeness, spokesperson audio, claims copy, or product depiction].
Client material. No Client Confidential Information will be submitted to any third-party model, and no Client material will be used to train, fine-tune or improve any third-party model.
Human authorship. Every Deliverable includes substantive human authorship in concept, selection, direction and finish. Supplier maintains a record of that contribution per Deliverable.
Provenance and disclosure. Deliverables are supplied with provenance metadata and the disclosure wording set out in the Delivery Specification.
Review. Client may request the production record for any Deliverable on [10] business days' notice.
Two design notes. The permitted list and the prohibited list should both be specific, because a client reading only a permitted list assumes it is exhaustive of your ambitions and a client reading only a prohibition list assumes everything else is a fight. Naming both closes it.
The review right at the bottom does more work than it looks like. Offering inspection voluntarily converts a policy objection into a process, and a client with an inspection right rarely uses it. Governance from the client's side of the table is mostly this same control set, which is why offering it reads as fluency rather than concession.
When to decline
Three situations where the right answer is no.
The client's prohibition comes from a regulator or a parent-company policy they cannot vary. You are negotiating with someone who has no authority to agree, and the time cost is unbounded.
The scope only works at your price because of generative production, and the compliant version is a job you would not have bid for. Requote at the honest number and let them decide. If the honest number loses the work, it was already lost.
The client wants the speed and the price of generative production alongside a contractual representation that none was used. That is not a negotiation, it is a request to sign something untrue, and it is the one case worth walking away from without a counter-offer.
FAQ
Should I disclose AI use if the contract does not require it?
Yes, at the commercial level, in the proposal. The account-losing event is almost never a client learning that you use AI. It is a client learning it from somewhere else. What to tell clients about using AI in delivery separates the commercial conversation from the compliance obligation, which are different things with different owners.
Does a no-AI clause cover tools that were already in my stack?
Read literally, most tool bans would prohibit the machine-learning features inside ordinary editing and transcription software, which no client intends. That over-breadth is a reason to ask for a definition rather than a reason to assume you are covered. Get the answer in writing before you rely on it.
Can I use open-weight models to sidestep the clause?
Not as a matter of contract. A prohibition on generative AI does not become inapplicable because the weights are self-hosted, and arguing otherwise is exactly the workaround that destroys trust when it surfaces. Where model licensing genuinely does change what you can ship is a separate question covered in open-weight video licences and what you can ship.
How do I price a compliant version honestly?
Scope it as the work it actually is and quote it against your normal production economics rather than your generative ones. The comparison in in-house versus agency versus AI content production is a reasonable sanity check on where that number should land before you send it.