Why disclosure and privacy are marketing problems
It is tempting to file both under legal and move on. That is a mistake for a practical reason: the decisions that create disclosure and privacy problems are made by marketers, in the moment, dozens of times a week — which tool to open, what to paste into it, whether the face in the hero image needs a label, whether the chat widget should say what it is. Legal can write the policy; only the team can follow it, and only a policy written in the team's own terms will be followed.
There is also a brand reason. People react badly to learning after the fact that they were talking to a system, that a testimonial was synthetic, or that something they told a company ended up in a third party's training data. The reaction is not proportionate to the harm; it is proportionate to the surprise. Disclosure done well removes the surprise. Data handling done well means there is nothing to be surprised about. Both are cheaper than the alternative, and both are easier when the AI uses themselves are known and tiered, which is what the AI governance hub establishes. Accuracy, covered in the accuracy and review workflows guide, is the third leg: a disclosed, privacy-safe output that is wrong is still wrong.
Disclosure to customers and platforms
The question to ask is not "was AI involved" — it almost always was, somewhere — but "would a reasonable person feel misled if they later learned how this was made or who they were talking to". The table gives practical answers for the situations marketing teams meet most.
| Situation | Disclose? | How | Why |
|---|---|---|---|
| A chat assistant or agent converses with customers | Yes | State plainly that it is an automated assistant at the start and when asked; offer a route to a person | People make different decisions when they know; regulatory direction in several regions expects it |
| AI-generated voice on calls or voice assistants | Yes | Say so at the start of the call | As above; a voice is more easily mistaken for a person |
| Synthetic person, face or likeness in advertising | Yes, in most cases | Visible label; never imply a real customer or employee | A synthetic "customer" is a fabricated endorsement |
| Realistic AI-generated or altered imagery on ad platforms | Follow the platform's documented rules | Use the platform's labelling tool where the help centre requires it | Platforms document labelling requirements for realistic synthetic media; unlabelled uploads may be rejected or removed |
| AI-edited product photography | Usually no, unless the edit changes the product | Keep edits to background and lighting; do not alter what the product is or does | Altered product features mislead buyers |
| Reviews, testimonials and ratings | Never generate them | Only real customers' words, with permission | Fabricated reviews are deceptive by any standard |
| Email, ad and page copy drafted with AI | Generally no | Review it as your own words; do not claim a human author who did not write it | Readers expect the brand to own its copy; the risk is false authorship, not AI use |
| Bylined articles and thought leadership | Yes, if the named author did not substantially write it | The author genuinely writes or heavily edits; otherwise no byline, or an honest note | A byline is a claim of authorship |
| Personalised offers based on inferred traits | Often, and avoid sensitive inferences | Explain why the person is seeing this; never infer health, religion, sexuality or similar | Inference from behaviour can expose or discriminate |
| Influencer or creator content made with AI | Follow endorsement rules and platform policy | Disclose the relationship as always; label synthetic elements | Endorsement guidance applies regardless of the tool |
Two rules of thumb cover the rest. If a disclosure would feel embarrassing to make, the problem is the use, not the disclosure. And if a platform documents a rule for generated media, it is the platform's rule that applies on that platform, whatever your policy says; the help centres are the authority and change often, so check them at upload time rather than from memory. If a platform has already rejected or labelled a generated asset, the troubleshooting page for rejected AI-generated ad creative reads the reason as given and traces it to the labelling check, the endorsement, the claim, the rights or the altered product. The AI for creative guide covers the rights and labelling questions in creative work in more depth; the AI for content guide covers authorship and search-engine guidance for written content. For service chat, the AI in customer service guide pairs this disclosure with handover rules and a weekly sampling review of automated answers. When customers ask about AI directly rather than waiting to be told, the guide to answering customer questions about AI prepares replies from the AI use register for staff, social managers and the assistant itself.
Customer data in prompts
The most common privacy failure in marketing AI is mundane: someone pastes a customer list, a transcript or an order history into a tool to get a better output, and the tool's terms allow the vendor to retain and learn from it. No breach notification follows because nobody notices; the data has simply left. A data rule stops this, and it needs to be short enough to remember.
Three classes of data. Public and non-personal data — product information, your own published content, aggregated statistics, anonymised figures — can 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. — names, contact details, order histories, behavioural records, transcripts, anything that identifies a person or could with a little effort — goes only into tools with contractual terms that forbid training on inputs, limit retention and keep the data in an agreed region, and only when the task needs it. Special-category data — health, finances in detail, religion, politics, sexuality, ethnicity, children's data — does not go into a prompt at all for marketing purposes, with the rare exception handled by legal rather than by the team.
Minimisation is the working principle. A personalisation prompt needs the first name and the category of the last purchase, not the full record. A win-back draft needs the tenure band and the cancellation reason, not the email address. Where the task needs individual-level data at scale — send-time personalisation across a segment, say — the right source is the customer data platform (CDPUnified first-party customer profiles.) or the automation platform's own fields, pseudonymised at the boundary, rather than an export. The same discipline already governs server-side event sharing with ad platforms: a Conversions APIServer-side event sharing with Meta. integration hashes identifiers and respects consent before anything leaves, as the Conversions API guide describes, and AI tooling deserves no less. Purpose matters as much as content: data collected to fulfil orders was not collected so that a model could write marketing, and the privacy and consent guide covers how consent and lawful basisOne of the six grounds the GDPR accepts for processing personal data, such as consent, contract or legitimate interests, chosen for each purpose. follow data through the stack. The AI for marketing automation guide shows where these prompts sit inside live workflows. For the hands-on preparation of a real export, transcript or ticket set, the routine for minimising and redacting data before prompting walks through which fields to drop, aggregate or pseudonymise and how to mask free text.
A fashion retailer rolls out a shopping assistant
A fashion retailer with 1.2 million email subscribers adds a chat assistant to its site. The governance review sorts the data it may see: the product catalogue, stock levels, delivery and returns policies and the current promotions, all non-personal; for signed-in customers, order status and the last 3 orders by product category, pseudonymised with a session token rather than the customer ID; nothing else. The vendor contract specifies no training on inputs, 30-day deletion of transcripts and EU hosting. The assistant opens every conversation by saying it is an automated assistant and that typing "agent" at any time reaches a person, and 6% of conversations do. A weekly sample of 100 transcripts is read by the service lead. In the first month the sample finds 4 conversations where a customer typed a full address or card detail unprompted; the team adds a filter that masks those patterns before the message reaches the model, and the retention clock for masked transcripts stays at 30 days.
Vendor terms and data use
The terms that matter are rarely in the marketing material and often differ between a consumer plan and a business plan of the same product. Read for six things. Training use: are inputs and outputs used to train or improve the vendor's models, and can that be turned off contractually rather than by a settings toggle a new user might miss. Retention: how long prompts, uploads and outputs are kept, and whether deletion on request is honoured and confirmed. Sub-processors: which other companies see the data, and whether you are told when the list changes. Region: where data is stored and processed, which matters to transfer rules. Access: who at the vendor can read your data and under what conditions. And ownership of outputs: whether you may use what the tool produces commercially, and whether the vendor warrants anything about it. When a vendor describes its product as "trained on your brandTraining an existing language model further on a set of examples so that a style, format or task pattern is built into the model itself.", these questions apply with extra force, and the retrieval vs fine-tuning comparison explains what the phrase can mean and why the mechanism changes what happens to the material supplied.
Business or enterprise tiers usually answer these acceptably; consumer tiers usually do not, which is why the data rule above says "approved tools" and the register says which plan. AI features inside existing platforms — the email tool's subject-line generator, the ad platform's asset generation, the analytics suite's natural-language querying — are covered by that platform's existing agreement, but the agreement may have changed when the feature arrived; check the data-processing terms for the AI feature specifically. The tool evaluation guide folds these questions into procurement, and the data flow and integrations guide shows where in the stack AI tools receive data and what they pass on. The tag governance guide is the on-site equivalent: AI chat widgets and personalisation scripts are tags, and the same approval and audit process applies.
Retention and deletion
AI tools create new stores of personal data that did not exist before: prompt histories, uploaded files, conversation transcripts, generated outputs that embed customer details, and review logs. Each needs a retention decision, and the decisions need to connect to the retention schedule the business already has, so that a customer's deletion request reaches the assistant's transcript store as reliably as it reaches the CRM.
A workable set of defaults. Prompt and conversation histories in assistants: short, measured in weeks, unless a specific reason requires longer. Uploaded files: deleted when the task ends, with the source of truth staying in its own system. Generated outputs that contain personal data (a personalised email, a drafted reply): kept only as long as the message itself is kept in the sending system. Review and sampling logs: kept for as long as governance needs them, but with personal data stripped or masked before logging, so the log can be long-lived without becoming a liability. Vendor-side retention is a contract term, not a hope; if the vendor's documented deletion period is longer than your policy, your policy is in practice the vendor's. Where a tool offers no deletion controls, it is not approved for personal data.
Regulatory direction
Three bodies of guidance shape how most marketing teams approach these questions, and all three can be read at the level of direction without a lawyer. What follows is that direction, with the sources linked from this page's source list; it is not legal advice, and the specifics depend on where you operate, what you sell and to whom.
The European approach to AI regulation is risk-based: obligations scale with the potential for harm, a small set of uses is prohibited outright, high-risk systems carry heavy duties, and the rest are mostly subject to transparency expectations. For marketing the transparency strand is the one that bites: people should be told when they are interacting with an AI system, and synthetic or manipulated content that could be mistaken for real should be identifiable as such. Most marketing uses sit well below the high-risk category, but a chat assistant and a synthetic face in an advert are squarely in the transparency conversation.
European data-protection law, as it applies to AI tools, asks the questions it has always asked: is there a lawful basis for the processing; was the data collected for this purpose or something compatible with it; is the processing limited to what is necessary; can people exercise their rights, including deletion; and are decisions with significant effects on individuals subject to safeguards and human involvement. Putting customer data into a model does not change the answers; it adds a processor to the chain and a new place where the data lives.
In the United States, the consumer-protection direction is that the standard for deception does not change with the technology. Claims must be truthful and substantiated whether a person or a model wrote them; endorsements and reviews must be genuine; "AI-powered" is itself a claim that must be true; and using AI to deceive — fake reviews, impersonation, misleading personalisation — is treated as deception, not as a technology question. The business guidance pages linked in this page's sources set the direction in plain language. The advertising-standards side, from claims that a product uses AI to generated reviews and synthetic spokespeople, is covered in the guide to AI claims, endorsements and consumer protection. Across all three, the same practical conclusions hold: tell people when they are talking to a system, label synthetic media, do not fabricate endorsements, put only necessary data into tools whose terms you have read, and keep the records that show you did.
A checklist
Use this before a new AI use goes live and at each review of an existing one. Items that cannot be ticked are reasons to stop, not notes for later.
Treat the checklist as a KPIA metric tied directly to an objective. source as well as a gate: the share of live uses that pass it at review is a cleaner measure of governance than the length of the policy.
Common mistakes
- Asking "was AI used" instead of "would anyone feel misled". The first is nearly always yes; the second is the question.
- Letting the chat widget pretend. A friendly name and a human avatar on an automated assistant is exactly the surprise disclosure exists to prevent.
- Generating testimonials "based on real feedback". They are not real feedback.
- Trusting a settings toggle instead of a contract term. Training opt-outs that live in a user's preferences disappear with the user.
- Pasting the full record. Minimisation is the rule; the prompt needs a few fields, not the export.
- Forgetting that AI features inside existing tools are new processing. The agreement you signed may predate the feature.
- Logging transcripts in full. Review logs that contain personal data become a second, unmanaged database.
- Reading the regulation as a checklist of clauses. Read it for direction, then ask counsel what applies to you.
