In most agencies the studio started using AI before anyone wrote a rule: moodboards, first-pass copy, a quick image to sell an idea in a pitch. The contracts say nothing about it, and the next enterprise client's questionnaire has a section asking how you use it. An agency's AI policy is the document that lets you answer, and it covers client material, ownership and likeness as well as the usual rules.
The first rule: nothing of a client's goes into a tool they have not agreed to
Unreleased products, embargoed campaigns, results before announcement and every brief a client trusted you with stay out of any tool the client has not agreed to. In most client contracts it is already a term. Put it first, in one sentence, so a junior designer at a late finish knows it without looking it up.
Five parts that carry more weight
The structure is the same as any AI policy, set out in what an AI policy should contain. These are the parts where an agency's version differs.
Who owns the output
Client contracts usually assign the intellectual property in the work and promise it is original, and AI-generated material makes both promises harder to keep.
In the United States, the Copyright Office's 2025 report on copyrightability said prompts alone do not make someone the author of the output, while human selection, arrangement and changes can be protected. In the UK, the government's March 2026 report on copyright and AI deferred a decision on protection for computer-generated works, while original work made with AI's help remains protected.
The position is still moving and none of this is legal advice, so your advisers decide it for your contracts. The policy should say where generated material may appear in final work, that it needs the client's agreement and how the human contribution is recorded.
Which generators, for what
Image, video, voice and music tools have very different terms on commercial use, on what they were trained on and on whether the vendor stands behind you if a claim arrives. Split the register in two: tools approved for ideas and internal work, and the smaller list approved for anything a client will publish.
Likeness, voice and style
Decide and write down what the agency will not do. For most agencies, the lines are no real person's face or voice without their written consent, and no prompting for the style of a named living artist in client work. Each is a choice the agency makes for itself, whether or not the law requires it, and clients increasingly ask what yours are.
Telling clients
Decide once, at leadership level, whether proposals and statements of work say how AI is used, and whether a client can opt out. Some client contracts now set this for you. The answer you settle on is also the answer to the AI section of a client's security questionnaire.
Freelancers
Because agencies run on freelancers, a freelancer's own AI account holding your client's material is the most likely route for it to leave. The policy covers everyone who touches client work, the freelance contract points to it and freelancers work on your accounts and tools. Logins end when the job does, which is why onboarding and offboarding get automated before the studio is connected.
The pitch is where it gets tested
The week of a pitch is when a shortcut looks most reasonable: a generated image to sell a route, a competitor's campaign pasted in to be picked apart. Make the approved tools good enough to use under that pressure, and make the line on generated material in pitches explicit, because a pitch image the client loves becomes the brief.
Where this fits
The policy is written in a half-day with the agency's leadership as the first output of Govern, and the readiness snapshot shows which tools the studio and the freelancers are already using. How the route plays out for an agency, including the candidates where a plain workflow does the job, is on the creative agencies page.