Writing an AI disclosure policy for client work

Why this stopped being an ethics question

For a while, whether to use AI in client work was a conversation people had with themselves. It is now a commercial question, and it arrives in three forms: a procurement questionnaire with a question about AI on it, a client who has read something and is nervous, or a competitor who has started advertising "100% human written" and made your silence look like an admission.

The agencies that handle this well are not the ones with the purest position. They are the ones who answered before they were asked. A policy you can send in thirty seconds reads as competence. A week of internal debate followed by a carefully worded paragraph reads as exactly what it is.

The three positions, and which one you are actually in

There are really only three defensible positions. The trouble starts when an agency claims one and operates another.

Position 1 — No AI in deliverables. Nothing generated by a model reaches the client. Research tools, transcription and grammar checking may still be used, and you should say so, because most people count those differently and you do not want to be caught by a definition. This position is genuinely sellable and genuinely expensive to hold. Do not claim it unless you audit it.

Position 2 — AI-assisted, human-authored. Models are used for outlining, research summaries, first drafts, alternative phrasings. A named human rewrites, verifies and takes responsibility for what ships. This is where the overwhelming majority of agencies actually are, and it is a perfectly honest position — but only if the human work is real rather than a light pass over generated text.

Position 3 — AI-generated, human-reviewed, disclosed. Volume work where the model does the drafting and a person checks facts and brand fit. Legitimate at the right price for the right client, dishonest if billed as position 2.

The useful exercise is to look at last month's actual deliverables and decide which position they were produced under. Most agencies discover the answer varies by client and by deadline pressure, which is the real thing the policy has to fix.

What the client is actually worried about

"Do you use AI?" is almost never a question about technology. Underneath it is usually one of four concerns, and answering the surface question satisfies none of them:

• "Am I paying writer rates for something that took four minutes?" A pricing concern. It is answered by being clear about what you are charging for — judgement, accountability, revisions — not by denying AI.
• "Will this get me penalised by Google?" A ranking concern. Google's stated position is that it rewards helpful, reliable content however it is produced, and targets content produced primarily to manipulate rankings. Say that plainly; it is more reassuring than a promise you cannot keep.
• "Is my confidential material being fed into someone's training data?" A data concern, and the most legitimate of the four. It deserves a specific answer about which tools you use and what their terms say. There is a separate guide on that.
• "Will this sound like everyone else?" A quality concern, and usually the real one. It is answered with work, not policy.

A policy that addresses all four is a sales document. One that only says "we use AI responsibly" is noise.

The policy

Short is better. This fits on one page, which means it will actually be sent to clients rather than living in a folder.

HOW WE USE AI — [Agency name]
Last updated: [date]

Our position
Every deliverable we send you is written or rewritten by a named person
at [Agency], who is accountable for its accuracy and its fit with your
brand. We use AI tools in the process. We do not send you generated text
that nobody has worked on.

Where we use it
- Research summaries and background reading
- Outlines and structural options
- First drafts, which are then rewritten by a person
- Alternative phrasings and headline options
- Transcription, grammar and readability checking

Where we do not use it
- Facts, figures, statistics and quotations. Anything presented as a
  fact is checked against a primary source by a person, and we can show
  you the source.
- Anything attributed to a named individual, including quotes and
  bylined opinion, without that person's sign-off.
- [Add your own exclusions.]

Your data
Material you send us is processed using [list the tools]. [State plainly
whether any of them may retain or train on the content, and where in the
world it is processed.] We will not put your confidential material into
a tool whose terms allow it to be used for training. If you need us to
work under tighter constraints, tell us and we will price for it.

What we will not claim
We do not promise that content will pass any particular AI detector.
Those tools disagree with one another and produce false positives on
ordinary human writing. What we will stand behind is that a person
wrote, verified and approved what we sent you.

Disclosure
If you would like a note on any deliverable stating how it was produced,
we will add one. If your own policy requires disclosure to your
customers, tell us and we will work to it.

Questions about this policy: [email]

Two lines in there do the heavy lifting. The refusal to promise detector results protects you from a promise you cannot keep. The offer to work to the client's own policy turns a defensive document into a service.

The clause for the contract

The policy is a communication document. If AI use is genuinely material to what you are selling, it belongs in the agreement too, where it also protects you.

AI TOOLS AND AUTHORSHIP

1. The Supplier may use artificial intelligence tools in the preparation
   of Deliverables. Every Deliverable is reviewed, edited and approved
   by a natural person employed or engaged by the Supplier before
   delivery.

2. The Supplier warrants that factual statements, figures and quotations
   in the Deliverables have been verified against primary sources, and
   will provide those sources on request.

3. The Supplier gives no warranty that any Deliverable will or will not
   be identified as AI-generated by any detection tool. The Client
   acknowledges that such tools are probabilistic and produce false
   results in both directions.

4. The Supplier will not input Client Confidential Information into any
   third party tool whose terms permit that information to be used to
   train models, without the Client's prior written consent.

5. On request, the Supplier will disclose the categories of AI tools
   used in producing a Deliverable.

Clause 3 is the one to keep. Agencies have been asked to warrant that content will "pass" a detector, and agreeing to that is agreeing to be judged by a third party's black box that changes without notice.

This is a starting point, not legal advice. Have a solicitor look at anything that goes into a real contract, particularly if you work with regulated clients.

The claims that get you into trouble

Marketing your position is fine. Overstating it is a consumer protection problem in the UK, not just a reputational one.

Under the Digital Markets, Competition and Consumers Act 2024, whose consumer protection provisions came into force in April 2025, the Competition and Markets Authority can act directly against misleading commercial practices — including issuing fines without going to court first. A claim like "100% human written" on a page where some work is model-drafted is a misleading claim about the nature of a service.

The same Act bans fake and undisclosed incentivised reviews, which is worth knowing if you are also being tempted to pad a testimonials page.

Practical version: claim the position you actually hold, define your terms so "human written" cannot be read more strongly than you mean it, and if the position changes, change the page.

Common questions