Why marketing needs an AI policy
Most marketing teams did not decide to adopt AI; individuals did, one browser tab at a time. That is fine for a brainstorm and a problem for everything else. Without a policy, nobody knows which tools are approved, what data may go into them, which outputs need a second pair of eyes, or who answers when a generated claim turns out to be false. The failures are rarely dramatic. They are a product description that cites a feature the product lacks, a segment built on a hallucinated field, a support reply that promises a refund policy that does not exist, a customer list pasted into a consumer chat tool whose terms allow training on it.
A policy for a marketing team is short. It names the approved tools and the tiers of use, states what data is never entered into a prompt, says which outputs need review and by whom, requires disclosure in the situations where a customer would reasonably want to know, and names an owner. The NIST AI Risk Management Framework offers a vocabulary that scales down well: govern (who is accountable), map (where AI is used and what could go wrong), measure (how it actually behaves) and manage (what you do about it). The European regulatory approach, which sorts AI uses by risk and attaches transparency duties to systems that interact with people or generate synthetic media, points in the same direction: proportion the controls to the harm. Search engines' quality guidance adds a commercial reason: content that lacks demonstrable experience and expertise (E-E-A-TExperience, Expertise, Authoritativeness, Trust.) is content that struggles to rank, whoever or whatever wrote it. The AI in marketing pillar sets the wider context; the content governance guide covers the editorial standards an AI policy plugs into.
Risk tiers by use case
The single most useful move in AI governance is to stop arguing about "AI" as one thing. A model that drafts an internal meeting summary and a model that talks to customers about a medical device are not the same risk, and treating them the same either strangles the first or under-governs the second. Sort uses into tiers by what happens if the output is wrong, and attach controls to the tier.
| Tier | Typical marketing uses | Why it sits here | Governance that fits |
|---|---|---|---|
| Internal and reversible | Brainstorming, outlining, summarising your own notes, first-pass research, rewriting internal documents | A person reads everything before it goes anywhere; errors cost minutes | Approved tools; the data rule; no formal review |
| Published with an editor | Draft copy for pages, emails and ads; social posts; image concepts; reports used for internal decisions | Reaches customers or shapes spend, but passes through a human step | Source grounding; fact and claim check; named editor sign-off |
| Customer-facing and automated | Chat assistants; agents that reply or act on records; send-time personalisation; recommendation and prediction triggers | Output reaches people without a person reading each one | Guardrails per step; sampling review; logs; disclosure; an owner on call |
| Regulated or sensitive | Claims about health, finance, safety or environmental impact; synthetic likenesses and voices; decisions that affect what an individual is offered | Wrong output can mislead, harm or breach a rule | Specialist review; legal or compliance sign-off; disclosure by default; often "do not automate" |
The tiers are a planning device, not a legal classification. A use can move tiers as it scales: a draft assistant that becomes an auto-sender has changed tier and needs a new review. The applications section shows the tiers in practice across content, creative, analysis and reporting, ad platforms and automation.
Accuracy and review, disclosure and privacy: the guides
The core guides carry most of the weight in this section.
Accuracy and review workflows is about the output. It catalogues the ways AI-assisted work goes wrong — fabricated facts, invented sources, unsupported claims, brief drift, inherited errors — and builds a review workflow around them: grounding the model in sources, a fact and claim check that treats every number, name and comparative as unverified until shown otherwise, review tiers that match the risk table above, roles and sign-off, and a log that turns each review into a better prompt or a tighter tier. It covers the design of the review, not the writing process; the AI-assisted content with review guide owns brief-to-draft-to-edit.
Disclosure and data privacy is about the people on both ends. On one end are customers, who may need to know they are talking to a system, that a face in an advert is synthetic, or that a review was not written by a reviewer, and platforms whose rules on generated media are documented in their help centres. On the other end is the data that goes into prompts: which customer fields are ever appropriate, what vendor terms must say, how long logs and transcripts are kept, and what the regulatory direction in Europe and the United States asks for at the level of principle. It is a practical guide and says so; it is not legal advice.
Read them in that order if the team mostly produces content; in the reverse order if the team mostly runs tools that talk to customers or touch records.
More governance guides
The core guides above carry the review and disclosure work. The pages below cover the rest of a team's governance: the written policy, the rules on claims, the open questions on copyright, data handling, measurement, skills, incidents, tool choice, the agencies and contractors who work on a team's marketing, the sampling plan for an automated use, and the questions customers ask about a business's AI use. Where they discuss rules they are educational, not legal advice. The foundations section explains the mechanics they assume, including how a team prompt library is kept and when retrieval or fine-tuning suits brand knowledge.
- AI Claims, Endorsements and Consumer Protection: claims that a product uses AI, generated reviews and endorsements, synthetic people and substantiation.
- Answering Customer Questions About Your AI Use: how to prepare honest answers when customers ask whether they spoke to a bot, whether an image or review is real, or what happened to their data in AI tools.
- Building AI Skills in a Marketing Team: a baseline, training on real tasks, review practice, shared prompts, champions and protected judgement.
- Copyright and Training Data: Open Questions: the unsettled copyright and training-data questions around AI marketing output, framed as questions for counsel and vendors, with habits that lower exposure.
- Evaluating AI Tools for a Marketing Team: how to evaluate an AI tool before a marketing team adopts it: a trial on real tasks, data and training terms, admin controls, model changes and exit.
- Fact-Checking an AI Answer Step by Step: a routine for checking an AI tool's answer before you use it: split it into claims, find primary sources, check quotes, figures and dates, mark the unknowns.
- Handling an AI Mistake That Reached Customers: what to do when AI-assisted output reaches customers with an error: contain it, correct it where people saw it, tell those affected, trace the cause and log it.
- Measuring AI's Impact on Team Output: baselines, time per accepted output, defect and rework rates, outcome quality, and comparisons that hold up.
- Minimising and Redacting Data Before Prompting: a routine for stripping personal data before material enters an AI tool: choose the fields, aggregate, pseudonymise, mask free text and check what is left.
- Setting AI Rules for Agencies and Contractors: how to agree AI use with the agencies and contractors working on your marketing: permitted uses, your data in their tools, review, disclosure, records and exit.
- Writing a Marketing AI Use Policy: register the uses, set tiers, the data rule, review and disclosure rules, owners, and the review date.
- Writing a Sampling Plan for an Automated AI Use: how to write the plan that makes a person read what an automated AI use produced: what counts as a failure, who reads, how the sample is drawn, what is logged and when the use stops.
Vendor terms and data
Most marketing AI arrives inside tools the team already uses — the email platform, the ad platform, the analytics suite, the CMS — plus a few standalone assistants. Governance of vendors is therefore mostly governance of the stack. Three questions go to every vendor before a tool is approved for anything above the internal tier. What happens to inputs: are prompts and uploaded data used to train models, retained, or shared with sub-processors, and can that be switched off contractually. Where the data lives and for how long. And what the tool does on its own: an "AI feature" that rewrites subject lines or expands audiences is an automated decision with a blast radius, and someone should know it is on.
The answers belong in the same place as the rest of the stack's documentation. The tool evaluation guide covers the procurement questions; the tag governance guide covers the parallel problem of scripts on the site that send data to third parties; and the privacy and consent guide covers the consent basis that any customer data entering a prompt must respect. When the tools belong to an agency or contractor rather than the team, the guide to setting AI rules for agencies and contractors covers what happens to the business's data in their tools, who reviews and discloses, which AI settings they may switch on in the business's ad accounts, and what comes back at the end.
Records and accountability
Governance that lives in people's heads lasts until they leave. Four records make it durable. A register of AI uses: each use, its tier, its owner, the tool, what data it sees and what it can do. A version history for prompts, models and settings, so that a change in behaviour can be traced to a change in configuration. Review logs: what was checked, what was found, who signed off. And an incident list: what went wrong, who noticed, what changed afterwards. The AI use register template sets out the register as one row per use, with columns that point to the other three records.
Accountability means one named owner per use who can answer for it and who reviews its logs on a schedule. Treat the register like any other operational KPIA metric tied directly to an objective. source: the share of uses with a current owner, the share reviewed in the last cycle and the number of open incidents say more about governance health than the length of the policy. The automation governance guide applies the same discipline to workflows generally, and its naming and ownership conventions extend naturally to AI steps inside them.
A retailer's first AI use register
A homeware retailer with a marketing team of nine audits its AI use and finds 14 distinct uses across 6 tools, only 3 of which anyone had thought of as "AI". The register sorts them: 7 internal (summaries, outlines, research), 4 published-with-editor (product copy drafts, email subject lines, blog drafts, image concepts), 2 customer-facing (a site chat assistant and the email platform's send-time optimisation) and 1 sensitive (a prompt that was pulling full order histories, including names and addresses, into a consumer chat tool to write better win-back emails). The sensitive use is stopped the same day. Each remaining use gets an owner and a tier; the chat assistant gets a weekly transcript sample of 50 conversations and a disclosure line. Three months later the register shows 16 uses, 16 owners, 2 logged incidents (both caught in review) and a review cycle that takes the owners about 2 hours a week between them.
Common governance problems
Most governance failures announce themselves through a familiar symptom. The table maps each to the guide that handles it.
| Problem | Usual cause | Guide |
|---|---|---|
| A published page cites a study or statistic nobody can find | No source grounding; fact check skipped or treated as proofreading | Accuracy and review workflows |
| Ad copy makes a comparative or "best" claim the product cannot support | Claim check absent; reviewer not briefed on substantiation | Accuracy and review workflows |
| Customers complain they did not know they were talking to a bot | No disclosure rule for conversational tools | Disclosure and data privacy; for the answers staff and assistants give, Answering customer questions about AI |
| A platform rejects or labels creative with a synthetic person | Platform's generated-media rules not checked before upload | Disclosure and data privacy, AI for creative |
| Customer records pasted into a consumer chat tool | No data rule; no approved enterprise tool for the task | Disclosure and data privacy |
| An automated flow sent off-brand or wrong messages to a segment | Offer or legal blocks left to the model; no sampling review; no rate limit | AI for marketing automation |
| An automated AI use is live and nobody reads what it sends after launch week | No sampling plan: no named reader, no schedule, no log and no rule for when to pause | Writing a sampling plan for one automated use |
| A report built with AI assistance contains a figure that does not match the source | Analysis output not verified against the query or dataset | AI for analysis and reporting |
| Nobody can say which campaigns a given prompt version touched | No register, no version history | The records section above; automation governance |
Start with the register. It takes an afternoon, and every other control in this section attaches to it. The applications hub is the companion section: it describes what AI is used for; this one describes how to keep that use safe. Where one of these symptoms keeps recurring, the AI in marketing troubleshooting section separates its causes in a diagnosis table and pairs each with a fix, a check and the control that prevents it.
