What the policy must decide
A policy is a set of decisions written down so that nobody has to make them again under deadline. For a marketing team the decisions are few, and each one corresponds to a failure the team has probably already seen in some form: copy that cited a feature the product lacks, a customer list pasted into a consumer chat tool, an automated reply nobody had read, a campaign nobody could trace back to the prompt that made it.
| Decision | The question it answers | Where the detail lives |
|---|---|---|
| Scope | Which tools and features count as AI for this policy | The register of uses |
| Tiers | How much control each kind of use needs | The governance hub's risk tiers |
| Data | What may and may not go into a prompt or upload | Disclosure and data privacy |
| Review | Which outputs are checked, how, and who signs off | Accuracy and review workflows |
| Disclosure | When customers and platforms are told AI was involved | Disclosure and data privacy |
| Tools and owners | Which tools are approved, on which plan, and who owns each use | The register of uses |
| Incidents | What counts as one, who is told, what happens next | The incident list |
| Review date | When the policy is checked against how the team actually works | The policy header |
The policy states the decisions; the guides hold the method. That split keeps the policy short enough to remember and lets the method change without reissuing the policy. The NIST AI Risk Management Framework gives a useful frame for the whole document: govern (who is accountable), map (where AI is used and what could go wrong), measure (how it behaves in practice) and manage (what you do about it). The register maps, the tiers and review rules manage, the incident and review clauses measure, and the owners govern. The NIST AI Risk Management Framework for marketing teams shows how the four functions fit together beyond the policy.
Before you start
- The governance hub read. The risk tiers in the AI governance hub are the backbone of the policy; this tutorial assumes them rather than repeating them.
- An owner for the policy. One named person who drafts it, keeps the register and calls the review. In a small team that is often the head of marketing or marketing operations.
- Access to the stack. The list of tools the team pays for, plus admin access or a contact who can check which AI features are switched on in each.
- A reviewer from legal, compliance or data protection. Someone who will read the draft before it is published, particularly the data and disclosure clauses.
- Existing standards to plug into. The editorial standards in the content governance guide and the sign-off routes in approvals and compliance reviews. The AI policy should reuse those routes rather than invent parallel ones.
- A couple of hours. The register takes most of the time; the clauses go quickly once it exists.
Writing the policy, step by step
List current uses
Start with what the team already does, not with what it should do. Ask each person to list every task where they use an AI tool or an AI feature, however small: brainstorming, rewriting, summarising reports, generating image concepts, drafting replies, asking an ad platform to generate assets, using an email tool's subject-line suggestions. Then check the stack yourself, because built-in features are routinely missed. Record each use as a row: the task, the tool and plan, the person who does it, what data it sees, what it produces and where that output goes. This register is the first thing the policy points to, and the regulatory direction rewards it: both the NIST framework's mapping function and the European Commission's risk-based approach start from knowing where AI is used.
Sort them into tiers
Give each row a tier using the hub's four: internal and reversible, published with an editor, customer-facing and automated, and regulated or sensitive. Tier by what happens if the output is wrong and nobody catches it, not by how impressive the tool is. A summary of your own notes is internal; the same model drafting an email to customers is published with an editor; the same model replying to customers on its own is customer-facing and automated. Mark any use that handles health, finance, safety or environmental claims, synthetic likenesses or decisions about what an individual is offered as regulated or sensitive. Some uses will turn out to need stopping now; stop them and note it.
Write the data rule
Write the rule about data in one short paragraph a person can recall in the moment. The disclosure and data privacy guide sets out three classes: public and non-personal data may go into any approved tool; personal data goes 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 does not go into a prompt for marketing purposes. Your First-Party DataData a business collects directly from its own customers, visitors and users. is mostly personal data, so the rule matters most where the team is tempted to paste exports. Add the minimisation habit in a sentence: share only the fields the task needs, prepared with the data minimisation routine.
Set review by tier
For each tier, state what review is required and who signs off. Internal uses need no formal review; the person using the output reads it. Published uses need a fact and claim check and a named editor's sign-off recorded. Customer-facing automated uses need guardrails, a sampled review on a schedule, logs and an owner. Regulated uses add specialist sign-off. Point to the accuracy and review workflows guide for the method rather than writing it out. Add one sentence on search quality: content is judged on demonstrated experience and expertise (E-E-A-TExperience, Expertise, Authoritativeness, Trust.), whoever drafted it, so review checks for substance as well as errors.
Set disclosure rules
Write when customers and platforms are told. The disclosure guide's position is a good default: tell people when they are interacting with an automated system, label synthetic people, voices and realistic generated media, follow each platform's documented rules for generated media at upload, never generate reviews or testimonials (the AI claims and consumer protection guide covers why), and do not put a person's byline on writing they did not substantially write. The European Commission's description of its approach attaches transparency expectations to systems that interact with people and to synthetic content, which points the same way. Keep the clause to principles and link the guide for the situations table.
Name approved tools and owners
List the approved tools, the plan each is approved on and the tiers it may be used for. The plan matters because consumer and business plans of the same product often carry different data terms. Then give every row in the register a named owner who can answer for it, reviews its logs and decides when it changes tier. Add the route for approving a new tool: who receives the request, what they check, and the principle that nothing above the internal tier runs on an unapproved tool. That route can be the same front door the guide to handling requests for new marketing tools sets up for every tool, with AI requests flagged and triaged against this register.
Define incidents and escalation
Say what counts as an incident: a wrong fact or claim that reached customers, personal data entered into an unapproved tool, an automated system acting outside its limits, generated media published without the required label, or output that breached a platform rule. Say who is told first, who decides on correction and who involves legal or data protection. The guide to handling an AI mistake sets out the response itself. Every incident goes into the incident list with what happened, who noticed, what was corrected and what changed afterwards. The point of the clause is that people report early, so say plainly that reporting a mistake is expected and will not be treated as the mistake.
Set the review date and publish
Put a review date in the header and a short list of triggers that bring it forward: a new tool, a use moving tier, an incident, a change in a platform's documented rules or a change in the regulatory picture your advisers flag. Send the draft to the legal or compliance reviewer, fix what they raise, then publish it where the team actually works, next to the register. Walk the team through it in a short session using their own register rows as examples, and record that each person has read it.
A one-page policy (illustrative)
The page below shows the shape a finished policy can take. The business is invented; the structure is the point. Note how little of it is method: almost every clause points to a register, a guide or a list.
Cobblemere Print Studio's AI use policy
Cobblemere Print Studio, an invented online stationery printer with a marketing team of 5, writes its policy after a register exercise finds 11 AI uses across 4 tools, 3 of them features built into tools the team already paid for. Its one-page policy reads, in summary. Scope: every row in the register; nothing outside it until added. Tiers: 6 uses internal, 3 published with an editor (product page drafts, email subject lines, social captions), 1 customer-facing (the site chat assistant), 1 stopped (a prompt that pasted the full order export into a free chat tool to write win-back emails). Data: no personal data in any tool except the chat assistant, which sees order status by order number only; never health or payment details. Review: published drafts get a fact and claim check and the content lead's sign-off; the chat assistant's owner reads a sample of 40 transcripts every fortnight. Disclosure: the assistant says it is automated in its first message; no generated reviews; no bylines on drafts the named author did not write. Tools: 2 approved assistants on business plans, plus the email and ad platforms' built-in features. Incidents: report to the marketing lead the same day; logged in the incident list. Review: every 6 months, or sooner on a new tool or an incident. At the first review the register has 13 rows, every row has an owner, and 1 incident (a wrong paper weight in a product description) is logged with its fix.
Rolling it out and keeping it current
A policy fails quietly. Nobody argues with it; people simply stop consulting it as the tools change around them. Three habits keep it alive. First, keep the register as a working document rather than an appendix: when someone starts using a new feature, the register row comes first. Second, review the policy on its date even if nothing seems to have changed, by walking the register row by row and asking whether each tier, owner and data rule still holds. Third, use the incident list as the main input: each incident should change a rule, a tier, a check or a tool, and the policy should show that it did. If the register or an incident shows customer records already going into consumer or personal accounts, the troubleshooting page for customer data found in AI tools starts with the data protection owner and reads the cases for the gaps the policy has to close. The AI governance operations playbook turns these habits into a standing programme, with an owner for the register, a starting rhythm for each review and a decision at each gate. A staff social media policy follows the same pattern of short decisions and linked guides, as the tutorial on writing a social media policy for staff shows.
Measure the policy by how it is followed, not by its length. A few register-based measures work as a KPIA metric tied directly to an objective. for governance health: the share of uses with a current owner, the share reviewed in the last cycle, the number of open incidents and how long they stayed open. The automation governance guide applies the same ownership discipline to workflows, and AI steps inside those workflows should appear in both registers under the same owner.
When the regulatory picture shifts, the policy changes through its review triggers, not through a rewrite. The European approach is still being implemented and platform rules on generated media change often, so the policy should state principles and point to guidance that is checked at the time of use, rather than copying rules that will date.
Verify it worked
Check the policy against the register and the team's practice.
- The register lists every AI use, including features built into the email platform, ad platform or CMS. Each row has a tier, a tool and plan, and a named owner.
- Each tier states its review and who signs off. Pick one published draft and find its recorded fact and claim check and editor sign-off.
- Ask any team member to state the data rule without looking. They name the three data classes correctly.
- The approved list names a plan for each tool and the tiers it may be used for.
- The header carries a review date and its triggers, and each person has recorded that they have read the policy.
If it goes wrong
- The register is shorter than the team expected. Built-in features were missed. Ask a tool admin which AI features are switched on and add rows.
- People cannot recall the rules. The policy holds method that belongs in the guides. Move checklists and situation tables out and keep the rules short.
- A customer export turns up in a consumer chat tool. The data rule was unclear or the wrong plan was approved. Stop the use, log an incident and tell the data protection owner.
- The incident list stays empty. People expect blame. Say plainly that reporting a mistake is expected and will not be treated as the mistake.
Next steps
- Read the content governance guide to line up the published tier's editor sign-off with your editorial standards.
- Read approvals and compliance reviews to route regulated or sensitive uses through the sign-off the organisation already has.
Common mistakes
- Writing the policy before the register. Rules written without knowing the uses either miss the real risks or ban things nobody does.
- Missing the built-in features. Subject-line generators, asset generation in ad platforms and plain-language querying in analytics tools are AI uses with data implications.
- Putting the method in the policy. Checklists and situation tables belong in the guides; the policy that contains them becomes too long to remember.
- One rule for every use. Treating a brainstorm and a customer-facing assistant the same either strangles the first or under-governs the second.
- Approving a product rather than a plan. The consumer plan and the business plan can have different terms on training and retention.
- No owner per use. A use everyone is responsible for is a use nobody reviews.
- Making incident reporting feel like confession. People hide mistakes they expect to be blamed for, and the incident list stays empty while the problems grow.
- No review date. A policy without one describes the tools the team used the year it was written.
