Why your AI policy has to reach the people you pay
An AI use policy sets the uses a team allows by risk tier, a data rule about what may go into any tool, the review each tier needs and when customers and platforms are told. Written for staff, it covers the work staff do. Yet in many marketing teams the agencies and contractors produce the adverts, the social posts, the press material and the research, and they do it in their own accounts, on their own plans, under terms the business has never read.
Three things follow. The business's data can end up in tools it has not approved, because the agency's tools are outside the team's approval route. Work can arrive with generated elements nobody declared, so the team's disclosure rules never get applied. And when something goes wrong, the team finds out late, because its register lists its own uses and none of the agency's. The disclosure and data privacy guide notes that customers react to surprise more than to harm; the same is true inside a business that discovers how its campaign was made after it has run.
The fix is not a separate policy for suppliers. It is the same policy, reaching further. The guide to working with freelance writers settles AI use with writers by having them work to the team's policy, agreed in writing before the first commission. This page applies that idea to every agency and contractor, with what larger suppliers add: account access and platform settings. The NIST AI Risk Management Framework, as the site applies it, asks where AI is used and who is accountable; for most marketing teams, part of the honest answer is "at our agencies".
Which agency work AI touches
Start with the work, not the agency. Each type of supplier uses AI in different places, and the risks follow the work. The table maps the common ones; it describes general practice, not what any particular agency does.
| Supplier | Where AI may enter the work | What the business most needs to know |
|---|---|---|
| Creative agency | Concepts, copy drafts, generated or edited images, video, voice, resizing | Which delivered assets contain generated elements, and from which tools |
| Media or search agency | Platform bidding and targeting, generated ad assets and enhancements, reporting summaries | Which AI settings are on in the business's accounts, and who switched them on |
| Paid social contractor | Ad variations, platform creative enhancements, comment replies | What the platform changed in each ad, and whether replies are drafted or sent automatically |
| PR freelancer or agency | Release drafts, media lists, monitoring summaries, spokesperson quotes | Whether quotes attributed to people were written by a tool, and who approved them |
| Content or SEO supplier | Articles, briefs, keyword research, product descriptions | Review standard and bylines, as the freelance writers guide sets them |
| Research supplier | Survey drafting, open-text coding, summaries, synthetic respondents | Whether findings rest on real respondents, and what data went into which tools |
Platform AI deserves its own line. When a media agency or contractor runs the business's Google Ads or Meta accounts, the AI enhancements and generated assets those platforms offer are switched on or off by whoever builds the campaign. The tutorial on reviewing AI enhancements in ad accounts builds a register of those settings and warns against leaving agency-run accounts out; the agreement should make the agency part of that review rather than a gap in it.
What the agreement should settle
The agreement does not need new legal language from the marketing team; it needs the policy attached and a few practical points made explicit. Most of these fit in the agency scope of work, which already sets deliverables, responsibilities, approvals and exit, or in the schedule a contract refers to. Whether a clause belongs in the contract itself, and how it should be worded, is a question for your legal adviser.
| Point to settle | What to agree | Who on the business side decides |
|---|---|---|
| The policy applies | The agency works to the business's AI use policy for this work, and is told when it changes | The policy owner |
| Permitted and excluded uses | Uses allowed by tier, and any the business rules out, such as generated people or voices | The policy owner, with the brand lead |
| Tools | Which tools the agency will use for which work, and on what kind of plan | The policy owner, against the evaluation route |
| Data | What business and customer data the agency may put into AI tools, and on what terms | The data protection lead |
| Declaration | What the agency tells the business about AI use in each deliverable, and when | The approver of the work |
| Platform settings | Which AI features the agency may switch on in the business's accounts without asking | The account owner |
| Records | What the agency keeps and shares: tools, prompts for key work, generated source files | The policy owner |
| Ownership and exit | Questions on rights in generated work, and what comes back at the end | Legal adviser, then the account owner |
A RACI matrix helps here. For each row, one person on the business side is accountable, the agency is responsible for doing it, and others are consulted. Without that, the agency's account lead and the business's marketing lead each assume the other is checking. A RetainerA standing fee paid regularly to an agency or supplier for continuous work with a stable scope, as distinct from project fees, time and materials or commission. or project fee was set for a way of working; if AI changes the work, revisit scope and fees at the next review, as the agencies guide describes. For media agencies, the guide to setting up a media agency agreement covers ownership of accounts, data and tools, to which the AI points add a layer rather than a replacement.
Agree the rules before the work starts, as part of onboarding. The agency onboarding checklist already covers access through the business's own accounts, conventions and the approval route; adding the AI use policy to what is shared and agreed in those first weeks is the natural place for it. Once the work is running, the AI governance operations playbook reviews agency rules and the AI settings in the business's accounts alongside the rest of the use register, as part of a standing quarterly walk.
Your data in their tools
The data rule in the AI use policy, as the policy tutorial writes it from the disclosure guide's classes, has three parts: public and non-personal data may go into any approved tool; Personal DataAny information relating to an identified or identifiable living person, such as a name, email address, location data or an online identifier. only into tools whose contractual terms forbid training on inputs, limit retention and set a region, and only when the task needs it; special-category data not into a prompt for marketing purposes. The rule should apply to the business's data wherever it goes, including into a supplier's tools.
Agencies receive briefs, product plans, sales figures, customer lists for audience building and survey responses, and their AI tools may sit on plans whose terms on training and retention the business has never seen. The disclosure and data privacy guide lists what to read in any AI vendor's terms: training use, retention, sub-processors, region, vendor access and ownership of outputs. Ask the agency for those answers for the tools it will use on your work, or require that only tools on your approved list touch your data.
Where an agency processes personal data on the business's behalf, the guide to evaluating AI tools describes what the GDPR expects of a contract with a processor, among other things acting on documented instructions and using sub-processors only with authorisation. An AI tool the agency puts that data into joins the chain. Whether a given agency is a processor, and whether its terms and its tools' terms are sufficient, are questions for your data protection adviser; the marketing lead's job is to make sure the question is asked before customer data is handed over. The simplest protection is the one the data minimisation tutorial teaches: send only the fields the task needs, pseudonymised or aggregated where possible, so less is at stake whichever tool it reaches.
Review, disclosure and who answers for errors
Work from an agency passes the same review as work from staff. The accuracy and review workflows guide sizes review to risk, and the business's Approval MatrixA table that says which reviews each type of marketing work needs, such as legal, brand or budget, and who gives each, matched to the risk of the work. says who signs off what. An agency's internal review is welcome but does not replace the business's, because the business is the one that will answer for the result. The site's guide to AI claims and consumer protection, drawing on the FTC's business guidance, treats the standard as unchanged by the tool: claims must be truthful and capable of SubstantiationThe evidence a business holds, before a claim is published, that the claim is true; advertising rules in most markets expect claims to be capable of it. whether a person or a model wrote them, and endorsements must be genuine. The same logic applies to who held the tool. An agency's generated headline is the advertiser's claim once it runs.
Declaration is what makes review possible. Ask the agency to state, at the point of approval, which parts of a deliverable were generated or materially altered by AI and which platform enhancements are switched on for the ads being approved. That lets the approver apply the team's disclosure rules, which the disclosure guide sets out: tell customers when they are interacting with an automated system, label realistic synthetic media where platforms require it, and never present a generated person as a real customer, employee or expert. Without the declaration, the approver cannot know that a rule applies.
When an error reaches customers, the guide to handling an AI mistake sets out the response: contain, correct, tell, trace and log. Agree in advance that the agency reports an AI-related error to the business promptly and helps trace it. How liability is shared is a contract matter for your adviser; the operational point is that the business hears first from its agency, not from a customer. The academy checkpoint on AI incidents, claims and agency rules tests what an agency agreement should settle about the business's data in the agency's tools, who answers for an error and which AI settings the agency may switch on in the business's ad accounts.
Records, ownership questions and exit
The business's register should show the AI uses in work made for it, not only the uses by its own staff. The AI use register template records one row per use with the tool, the data it sees, the review route and the disclosure. Add rows for material agency uses, with the agency named as the party doing the work and a business owner named as accountable. Ask the agency to keep, for key work, the tools used, the prompts or instructions behind generated assets and the generated source files, so that a question months later can be answered.
Ownership is the open question. The copyright and training data guide frames rights in AI-generated work, and what contracts with agencies and freelancers should say about AI use, as questions for counsel rather than settled answers. Agencies and contractors may also use stock, music or talent with Usage RightsThe terms in a creator or influencer agreement that set how, where and for how long a brand may reuse the content, organically or in paid ads. limited by channel and time, and generated elements add another layer of uncertainty. Record which delivered assets contain generated elements, so the adviser can answer for the assets that matter, such as logos, mascots and signature imagery.
At the end of the relationship, the agency offboarding checklist collects creative, history, tracking documentation and reports, confirms account ownership and removes access. The AI layer adds a few items: the register rows for the agency's uses, the prompts and generated source files for work the business will keep using, the current state of AI settings in the accounts the agency ran, and confirmation, as far as the agency's tools allow, of what happens to the business's data held in them. The incoming agency then starts from the same register rather than discovering the settings for itself. The AI governance hub sets out the risk tiers and the other governance guides these rules draw on.
A brewery sets AI rules with its agencies (illustrative)
Mabrushka Brewing extends its AI policy to three suppliers
Mabrushka Brewing, an invented independent brewery, works with 3 suppliers: a creative agency on a monthly retainer, a paid social contractor running its Meta account and a PR freelancer. Its marketing lead wrote an AI use policy for the 4-person in-house team the previous year and had never sent it to any of them.
A review of the last quarter's work finds 3 different practices. The creative agency used generated backgrounds in 5 of 12 can-launch visuals and did not mention it. The social contractor had uploaded the brewery's mailing list of customers who opted in to marketing into a free AI tool to draft audience descriptions. The PR freelancer drafted a quote for the head brewer with a chat tool, which the head brewer approved without changes.
| Supplier | Rule agreed | Where it is recorded |
|---|---|---|
| Creative agency | Declare generated or AI-altered elements in every asset at approval; no generated people | Scope of work, approval form |
| Social contractor | No customer lists in any AI tool; Meta enhancement settings listed with each ad for approval | Scope of work, enhancement register |
| PR freelancer | Quotes are drafted from an interview and approved by the person quoted | Commission note |
| All three | Work to the brewery's AI use policy; report AI-related errors within 1 working day | Policy attached to each agreement |
The marketing lead adds 6 rows to the register for the suppliers' uses, each with a brewery owner. The social contractor deletes the uploaded list as far as the tool allows and confirms it in writing; the data protection adviser reviews what happened. The offboarding list for each supplier gains 2 items: prompts and generated source files for assets still in use, and the AI settings in any account they ran.
Common mistakes
- Writing a policy for staff only. If most customer-facing work is made by suppliers, a staff-only policy covers the smaller part of the risk.
- Banning AI instead of governing it. A ban nobody can check is followed in name; extending the policy, with declarations at approval, is followed in practice.
- Handing over customer lists without asking where they go. The data rule applies to the business's data in a supplier's tools as much as in its own.
- Relying on the agency's review. The business answers for what runs under its name; its own review still applies.
- Leaving agency-run ad accounts out of the enhancement review. Platform AI settings in those accounts are the business's settings.
- Treating ownership of generated work as settled. It is a question for counsel; record which assets contain generated elements so the question can be answered.
- Forgetting the AI layer at exit. Prompts, generated source files, register rows and account settings belong on the offboarding list.
