AI Governance for Marketing Teams: A Policy Template Aligned With the EU AI Act

AI governance for marketing is a written policy that fixes which AI uses are approved, what data they may touch, who reviews outputs, what customers are told and how each AI-assisted action is logged. Marketing needs its own version because it processes personal data at scale in tools such as the customer data platform (CDP) and sends content straight to customers. This article gives a ten-section template with the EU AI Act and GDPR mapping.

Who this is for: heads of marketing operations, DPOs, CISOs and counsel who must put a usable policy in front of a marketing team. You will get the legal timeline, the template, a RACI table and a checklist. This page belongs to the AI hub.

Why marketing needs its own AI policy

A company-wide AI policy usually covers model risk and acceptable use. It tends to miss what is particular to marketing:

• Personal data at volume. Segmentation, scoring and personalisation run on profiles, events and consent records.
• External output. Content and chat replies reach customers directly, so errors become public statements.
• Consent and suppression. A wrong send to an opted-out person is a legal problem, not only a brand problem. The UK regulator fined HelloFresh £140,000 in January 2024 over about 80.9 million marketing messages, citing consent that was not specific or informed and contact after people asked to stop 15.
• Claims. The US Federal Trade Commission's Operation AI Comply (25 September 2024) stated there is no AI exemption from existing law and pursued deceptive AI claims and a tool that generated fake reviews 12.
• Embedded AI. AI features arrive inside tools you already license.

The legal timeline

The EU AI Act's dates moved in 2026. Regulation (EU) 2026/1744 was published in the Official Journal on 24 July 2026 and entered into force on 27 July 2026 2.

What Article 50 asks of a marketing team

GDPR provisions the template relies on

Article 5 sets minimisation, purpose limitation, accuracy, storage limitation, integrity and accountability 4. Article 7 requires the controller to demonstrate consent and lets people withdraw it at any time 5. Article 9 prohibits processing special categories unless an exception applies 6. Article 22 gives a right not to be subject to solely automated decisions with legal or similarly significant effects, with human intervention and a right to contest 7. Article 25 requires data protection by design and by default 8. Article 35 requires an impact assessment for, among other cases, systematic and extensive profiling with significant effects 9. The European Data Protection Board's Opinion 28/2024 (18 December 2024) addresses AI models and personal data 10.

Frameworks: ISO/IEC 42001 and NIST AI RMF

The policy template

Copy the sections below and replace the bracketed items.

1. Purpose and scope

This policy governs the use of AI systems by [Company]'s marketing, lifecycle, CRM, analytics and web teams, and by contractors acting for them. AI systems include generative models, agents, predictive models and AI features inside third-party tools. It covers content, segmentation, personalisation, scoring, chat and reporting. Policy owner: [role]. Effective: [date].

2. Approved use cases and prohibited uses

Every use case is recorded in a register with an owner, purpose, data classes, vendor, model and tier.

• Tier 1: no personal data; internal drafting and summarising of public material. Approved by Marketing Operations.
• Tier 2: personal data or customer-facing output. Requires DPO approval and human review.
• Tier 3: decisions with legal or similarly significant effects, special-category data, or any use that may fall under AI Act Annex III. Prohibited unless Legal and the DPO approve after an impact assessment.

Prohibited in all tiers: manipulative or deceptive techniques; inferring or targeting on special-category data without DPO approval; fabricated reviews or testimonials; realistic synthetic depictions of real people without consent and a label; autonomous sending, publishing or audience activation; any change to consent or suppression state; entering credentials, health data or cardholder data into unapproved tools; unsupported claims about what our AI does; emotion recognition or biometric categorisation pending legal review. The 7 things agents should never do expands the agent list.

3. Data rules

• Non-personal data: any approved tool.
• Personal data: approved tools only, under a data processing agreement, minimised and pseudonymised where possible, in approved regions.
• Health data: only in environments covered by a business associate agreement, limited to the minimum necessary 13; marketing uses need patient authorisation unless an exception applies 14. See HIPAA Marketing Rules Explained.
• Cardholder data: never enters an AI tool. See PCI DSS v4 and Marketing Data.
• Special-category data: not used as input or for inference without DPO approval.
• Prompts and outputs containing personal data are personal data and follow the retention schedule. Vendors may not train on our data unless the contract allows it.

4. Human review

A named person reviews all customer-facing AI output before publication. AI-assisted segments, rules and journeys are approved before activation. For Tier 2 and above, the reviewer is not the requester. Agents have propose-only rights: any send, publish, delete or consent change needs approval, consistent with OWASP's guidance on limiting autonomy 16. Reviewers check facts, claims, consent, fairness and brand. Approvals are recorded.

5. Transparency to customers

AI chat and voice agents state that they are AI at first interaction unless obvious (AI Act Article 50, applying from 2 August 2026) 1. Realistic AI-generated or altered images, audio or video of real people are labelled. We do not strip vendors' machine-readable markings. Our privacy notice describes AI-assisted profiling. Where a decision is solely automated with significant effects, we provide human intervention and a route to contest 7. Wording is owned by Legal.

6. Vendor and model approval

Each vendor passes a security and privacy review before use: data processing agreement, sub-processors, data location, retention of prompts and outputs, training on our data, business associate agreement where needed, log access and exit terms. We record the vendor's role under the AI Act. Model changes that alter behaviour trigger re-approval. AI features inside existing tools stay off until reviewed.

7. Logging and audit

Every AI-assisted action that touches customer data or reaches a customer is logged with actor, model and version, input reference, tools called, output reference, reviewer, decision, purpose and time. Logs are append-only and out of reach of the agent. PrivyCDP, for example, records AI-assisted actions in an append-only audit ledger. Retention: [period]. Sample audits run quarterly. See Explainable AI for Customer Decisions.

8. Incident handling

AI incidents include leaked data, harmful or false published content, an unauthorised send or activation, a successful prompt injection and a consent breach. Steps: contain (disable the tool, revoke tokens); tell the DPO immediately so any notification deadline can be met; preserve logs; correct and notify; review within [number] working days; update the register.

9. Roles and responsibilities

See the RACI table below. The policy owner maintains the register; the DPO owns data protection decisions; the CISO owns security and logging; Legal owns disclosure wording and Tier 3 approvals.

10. Review cadence

Review the policy every [six] months and on triggers: new AI Act guidance or dates, a new vendor or model, an incident, or a change in law. Review the register and sample logs quarterly. Train staff on joining and annually, as AI literacy measures.

RACI for the policy

R is responsible, A accountable, C consulted, I informed.

Adoption checklist

• Name the policy owner and approve the RACI.
• Build the use-case register and classify each current use into a tier.
• Inventory AI features already enabled in martech tools and switch off unreviewed ones.
• Classify data and record which classes each tool may receive.
• Confirm data processing and business associate agreements for each vendor.
• Set human-review rules for each channel and who may approve.
• Draft chatbot, deepfake and synthetic-content disclosure wording.
• Turn on action-level logging that agents cannot edit.
• Write the incident playbook and run one tabletop exercise.
• Schedule the six-monthly policy review and the quarterly register review.

For a walkthrough of how a customer-owned deployment supports these controls, with approval gates and an audit ledger, see the compliance guide.

Related reading: the AI hub, Running AI Agents on Customer Data Without Leaving Your Boundary and the AI marketing operations roadmap.