Skip to content
Implementa.
OpinionPlaybook··9 min

The AI use clause in your service contracts: what to disclose to clients and what to promise

An AI use clause in your service contracts with clients isn't optional anymore. If you don't write it, your client will, and theirs either bans everything or makes you liable for everything. What to disclose, what to promise about the process, and what never to sign about the machine.

Managing Partner

Implementa

Illustrative scenario. A fifteen-person marketing consultancy renews the contract with its biggest client. The draft from the client's legal team has a new line, quiet, on page eleven: “Provider shall not use artificial intelligence tools in performing the Services without Client's prior written authorization.” The account lead reads it, thinks “we don't use AI for anything important,” and signs. That same month, the team transcribes meetings with an automatic tool, writes first drafts with an assistant, and translates assets for three markets with a neural translator. Nobody asked for authorization. Technically, from the day of signature, the consultancy has been in breach.

The thesis in one line: the AI use clause isn't a courtesy notice. It's the line between what you can do without asking and what you answer for when it goes wrong. If you don't write it, your client will, and theirs usually bans everything or makes you liable for everything. What protects both sides: disclose the use precisely, promise on the process — review, data, notice of changes — and never promise on the machine. Before we go on: this isn't legal advice. For your contract and your jurisdiction, talk to your lawyer.

What is an AI use clause in a service contract with clients?

An AI use clause is the section of a service contract where the provider states whether it uses AI systems to deliver the service, for which tasks, with which client data and under which controls — and where both parties set what's allowed, what's excluded, and who answers for the output. It doesn't regulate technology. It regulates trust. It tells the client what they're actually buying when they buy your work.

Don't confuse it with its two cousins. The first is the contract with your own AI vendor, where you're the buyer and what's at stake is the intellectual property of what the tool generates. The second is the security questionnaire a large client sends you about AI: a due-diligence snapshot you fill in once a year. The clause is different. It's what stays signed and binds you every day the contract runs.

Why does your client suddenly want to know if you use AI?

Because today, in most cases, they don't know. The Association of Corporate Counsel, the international association of in-house lawyers, published in November 2025 the results of a survey run with Everlaw among in-house legal departments: 59% of respondents don't know whether the outside law firms they hire use generative AI on their matters. Scope: in-house legal departments and their relationship with law firms, not every kind of vendor. But the pattern shows up in any professional service: the client pays for work and doesn't know how it gets done.

The same survey gives the other half of the story: generative AI adoption inside those legal departments went from 23% in 2024 to 52% in 2025. In other words, the people hiring you already use AI and know exactly what to ask. The reasonable expectation — and this is our inference, not a survey finding — is that the question ends up written into the next contract.

And buyers already have templates. The European Commission, through its community of practice on public procurement of AI, published on March 5, 2025 an updated version of its model contractual clauses for procuring AI: a full version for high-risk systems, aligned with the EU AI Act, and a light version for everything else. They're built for public buyers, but the message to any provider is clear: the buy side already shows up with text ready. If you don't bring yours, you negotiate on theirs.

What to disclose: the minimum inventory

A good AI use clause fits on one page plus an annex. The page sets the rules; the annex describes actual use and gets updated with notice. What can't be missing:

  1. Which service tasks involve AI. By task, not by tool: “meeting transcription,” “first draft of reports,” “ticket triage.” Tools change every quarter; tasks almost never do. Name brands and your contract expires with the next market update.
  2. With which client data. Whether client data ever goes into an AI system, what kind it is, and whether it's minimized first. This is the line that reassures most when the answer is precise — and worries most when it's vague.
  3. Which third parties are involved. If an AI vendor processes the client's personal data on your behalf, it's a sub-processor. When that data falls under the GDPR — EU clients, for example — Article 28(2) of the GDPR is explicit: a processor can't engage another processor without the controller's prior specific or general written authorization; and if it's general, the processor must report any change of sub-processor so the controller can object. Not a courtesy. An obligation that already exists.
  4. Who reviews before delivery. Which person or role validates what the AI produced before it reaches the client, and against which criteria. It's the most valuable promise, because it's the only one that depends entirely on you.
  5. What isn't done with AI. The explicit exclusion is often the most reassuring sentence in the contract: “AI systems are not used for decisions about people,” “health data never leaves our systems.” What you won't do is worth as much as what you will.

What to promise and what never to sign

The underlying rule is simple: promise on what you control — your process — and never on what you don't — the behavior of a third-party model. Applied to the promises that show up most in drafts:

PromiseSignable?Why
“Every deliverable is reviewed by an accountable person before it reaches Client”YesDepends only on you, and you can prove it with a log.
“Client data is not used to train models”Yes, if your vendor signs it to youIt's an inherited promise: you can only give the one you have in writing from your AI vendor.
“We'll give advance notice before using AI on a new task or switching vendors”YesIt's process, not technology. Set a notice period you can actually meet.
“The AI won't make mistakes” or “99% accuracy”NoA model is probabilistic. You promise control over the error, not its absence.
“No AI tools will be used”Almost neverIt's false the moment someone uses a spellchecker, a transcriber, or a translator. Silent breach from day one.
“Unlimited indemnity for any generated content”NoYou take on a risk your own AI vendor doesn't cover for you.

The review promise deserves one more line, because it's what turns the clause into a system instead of paperwork. Promising “human review” without saying who, when, and against what criteria is theatre. A well-designed human in the loop sets where in the workflow the review happens, what counts as an error, and what happens when one is caught. That's what you sign. The rest is set dressing.

Same with data. “We don't train on your data” doesn't answer whether it's stored, for how long, and where — the difference we break down in AI vendor data retention. Before you promise anything about your client's data, check what your vendor promises you. You can't pass on more guarantee than you hold.

What do you do when the client brings their own ban clause?

Don't sign it as is, and don't sign it planning to ignore it. A blanket ban is the contract version of the internal memo that bans AI: looks good on paper, pushes usage into the shadows — the same mechanism we describe in shadow AI: banning it doesn't work. The alternative a reasonable client usually accepts has three parts:

  • Replace the ban with an authorization plus annex. A list of authorized uses, by task, that the client approves at signature. Anything not on the list requires notice and consent.
  • Scope by data, not by technology. If what worries the client is their health data, financial information, or trade secrets, write the clause around that data: what can never go into any external AI system. It's more precise, and you can actually comply with it.
  • Define what AI means for this contract. Without a definition, the spellchecker counts, the spam filter counts, and so does the summarize feature your office suite ships turned on by default. A useful definition focuses on systems that generate or transform service content, or make decisions about it, using client data.

Who keeps the clause alive after signing?

A clause that describes March's uses is false by June if nobody touches it. The real risk isn't signing badly. It's signing well and letting operations drift away from the paper. Keeping what you disclosed true takes two things that aren't legal but operational: an inventory of which AI systems touch each service, with an owner, and a log of which output went through which system and who approved it. That's what AI automation governance and control covers: permissions, logs, and the ability to roll back.

The day the client asks “did anyone review this deliverable?”, the answer has to come from a log, not from the memory of whoever ran the account. Building that traceability of AI decisions is what turns the clause into something you can prove. And if the service also puts an AI assistant in front of your client's customers, there's a separate obligation that lives in the product, not the contract: your chatbot has to disclose that it's an AI.

When don't you need a specific AI use clause?

When AI only touches your internal admin, no client data goes in, and nothing about the deliverable changes. If you use an assistant to sort your calendar or draft your own sales proposals, your general confidentiality clause and your data processing agreement already cover what needs covering. Adding an AI clause there creates expectations of control over something that doesn't affect the client. One caveat: if they ask, tell the truth. Not needing a clause isn't the same as not using AI.

What the industry does and what works

Professional services have split into two reflexes. Some hide AI because they fear the client will ask for a discount; others turn it into a slogan — “AI-powered” — without saying where, with what data, or under what control. Both are theatre. What works is writing down your AI use with the same precision you write price and deadline: what it does, with what, who reviews it, and what happens when it fails. The client who reads that doesn't see a risk. They see a provider who knows what they're doing.

Note: this article isn't legal advice. References are Article 28 of Regulation (EU) 2016/679 (GDPR), the European Commission's model contractual clauses for procuring AI (updated version of March 5, 2025, built for public buyers), and the ACC survey with Everlaw published November 5, 2025 (in-house legal departments and outside law firms). The scenario is illustrative. To draft your clause, talk to your lawyer.

Shall we get it shipping?

If this resonated, 30-minute conversation with no commitment. We tell you what fits, what doesn't and the approximate price.

See cases
The AI use clause in your service contracts: what to disclose to clients and what to promise · Implementa