EU AI Act for Marketing and Customer Decisioning: Compliance Checklist, Deadlines and High-Risk Rules

To meet the EU AI Act in marketing, inventory every AI system fed by your customer data platform (CDP), classify each one as prohibited, high-risk, transparency-only or minimal risk, and work to the dates: Article 50 transparency applies from 2 August 2026, and stand-alone high-risk duties (credit scoring, life and health insurance pricing, emotion recognition) from 2 December 2027 15. Ordinary segmentation and personalization is usually minimal risk.

Who this is for: marketing operations leads, DPOs and architects who need a class, owner and evidence list per AI feature. Part of the regulation hub; dates also sit in the deadline tracker.

EU AI Act deadlines after the Digital Omnibus

The "Digital Omnibus on AI" is law, not a proposal. Regulation (EU) 2026/1744 was adopted on 8 July 2026, published in the Official Journal on 24 July and entered into force on 27 July 2026 1. It moved the high-risk dates to fixed calendar days and kept Article 50 on 2 August 2026, with a marking grace period.

Three clarifications:

• Fines attach as soon as each article applies. Prohibited practices: up to EUR 35 million or 7% of worldwide turnover. Deployer duties (Article 26) and transparency (Article 50): EUR 15 million or 3%. Misleading authorities: EUR 7.5 million or 1% (Art. 99) 45.
• The broader "Digital Omnibus" is still a proposal. The Commission's 19 November 2025 package to amend the GDPR, ePrivacy and the Data Act is separate and not law. Plan against the GDPR as written.
• Two softer changes help marketing teams. Article 4 now asks providers and deployers to "take measures to support the development of AI literacy" rather than ensure a level of it, and new Article 4a allows narrowly conditioned use of special-category data to detect bias 1. Gibson Dunn summarizes the package 15.

Who is responsible: provider or deployer?

A provider develops an AI system, or has it developed, and puts it on the market or into service under its own name; a deployer uses a system under its own authority 5. The Act reaches non-EU companies whose output is used in the Union (Art. 2(1)(a) to (c)) 4.

Your role can change. Under Article 25(1) a deployer becomes the provider of a high-risk system if it puts its own name on it, substantially modifies it, or changes the intended purpose of a non-high-risk system so that it becomes high-risk 3. A team that repurposes a propensity model as a credit eligibility filter has become a provider.

Which marketing and customer-decisioning use cases are high-risk?

Classification follows intended use, not technology. A system is high-risk when it falls in an Annex III use case (Art. 6(2)) or is a regulated product component (Art. 6(1)) 4. Profiling matters in one way: an Annex III system that profiles natural persons never qualifies for the Article 6(3) "narrow task" exemption 4. Profiling alone does not make a marketing model high-risk.

The Commission's draft classification guidelines of 19 May 2026 are non-binding and may change, but they are the best signal available 7. The matrix below applies them.

Three details decide edge cases. Credit is "essential" for loans, mortgages and credit-line or card-limit extensions, not premium cards, leisure loans or securities; a system serving both is high-risk. The fraud exception applies only if fraud detection is the main purpose. Insurance covers life, health, long-term care and credit life; claims management and product design models fall outside 7.

What high-risk duties mean in a customer-data stack

Where private deployment and auditability help

Terms first. On-premise means the software runs in your data center; private cloud is a single-tenant environment in your own cloud account; self-hosted means you operate it; customer-owned means you hold the license, keys and logs; sovereign means the jurisdiction and key control are yours.

The AI Act requires none of these, and none makes a model compliant. Software as a service (SaaS) is a reasonable answer for minimal-risk personalization. Location matters when you provide or deploy an Annex III system and must reproduce a decision years later: logs help only if you can export them, keep them immutable and replay the model version that decided. In a vendor tenant the logs sit under the vendor's control and need contractual export rights; in an environment you own, you set retention, access and immutability yourself. PrivyCDP (immutable append-only audit ledger, customer-held keys), Meiro and vendor-operated single-tenant options such as Tealium Private Cloud can support that; see the self-hosted CDP guide and explainable AI for customer decisions.

Article 50: what customer-facing bots and generated content must do

Article 50 applies from 2 August 2026, enforced mainly by national market surveillance authorities; breaches carry up to EUR 15 million or 3% of turnover 5.

The Commission's FAQ and guidelines 56 add the following.

• Chatbots. The duty covers genuine two-way exchanges where AI talks to natural persons directly; background and machine-to-machine systems are out. The "obvious" exception is read narrowly, and disclosure must come from the start of the first interaction, so a clause in the terms and conditions will not do. A brand that builds a branded assistant on a model API is the provider of that assistant, so Article 50(1) lands on you, not just the model vendor.
• Generated content. A generator's machine-readable mark does not discharge your deepfake-labeling duty. The deepfake test is objective: content resembling real people, objects, places or events that would falsely appear authentic. The text-labeling duty is tied to public-interest topics such as politics, public health and consumer safety; the FAQ does not single out marketing copy, and a law-firm reading says product descriptions fall outside 16. Treat health and safety claims with care.
• Grace period. Only marking by generators already on the market before 2 August 2026 is deferred, to 2 December 2026.
• Code of Practice. The final Code on marking and labeling AI-generated content (10 June 2026) is voluntary; non-adherents must show compliance by other equally adequate means 614.

How the AI Act interacts with GDPR Article 22

The AI Act adds product rules; it does not replace the GDPR. Article 22 gives people the right not to be subject to a decision "based solely on automated processing" with legal or similarly significant effects, subject to exceptions (contract necessity, authorizing law, explicit consent) and safeguards: human intervention, a chance to express a view and to contest 9. Article 21(2) gives an unconditional right to object to profiling for direct marketing; Article 35(3)(a) requires a data protection impact assessment for extensive profiling-based evaluation with significant effects 9.

Three recent enforcement and court signals:

• Uber, EUR 824,990,000. The Dutch data protection authority fined Uber B.V. in August 2026 for unlawful automated decision-making (Article 22) in deactivating driver accounts, and for inadequate information (Article 13). The case began with 171 French drivers, covers 2018 to 2022 and is under appeal 12.
• SCHUFA (C-634/21, 7 December 2023). A credit score is itself an Article 22 decision when the third party receiving it relies heavily on it 10.
• Dun & Bradstreet Austria (C-203/22, 27 February 2025). Controllers must explain the procedure and principles applied so the person can understand and challenge the result 11.

For credit, the Consumer Credit Directive (EU) 2023/2225 applies from 20 November 2026: where the creditworthiness assessment uses automated processing, the consumer may obtain human intervention, a clear explanation, the chance to state a view and a review (Art. 18(8)) 13. AI Act Article 86 adds a right to an explanation of the AI system's role for Annex III decisions, but only where other EU law does not already provide it 4. Design one explanation and review path that satisfies all three. For the UK, which diverges, see the Data (Use and Access) Act guide.

EU AI Act compliance checklist (20 points)

Use the AI governance policy template to anchor ownership.

Inventory

• 1. List every model, rule engine and large language model feature that reads from or writes to your CDP: vendor, version, purpose, inputs, outputs.
• 2. Record whether you are provider, deployer or both for each; check the Article 25(1) triggers.
• 3. Map each output to the decision it drives (message, offer, price, limit, eligibility, account action) and whether a human reviews it.
• 4. Flag anything touching credit, insurance, employment, public benefits, or voice and face analysis: your Annex III candidates.
• 5. Inventory chatbots, voice agents and generative tools, including those in vendor suites and agency workflows.

Classify

• 6. Screen every feature against Article 5: hidden emotion inference, exploiting financial vulnerability, manipulative interfaces. Retire or redesign matches.
• 7. Write a dated Article 6(3) analysis for each Annex III candidate; profiling removes the exemption.
• 8. Separate marketing models from creditworthiness models and block data flows between them.
• 9. Assign tier, owner and review date; re-run on any change of purpose.

Configure

• 10. Add an AI disclosure at the first moment of every bot and voice interaction.
• 11. Set a labeling rule for realistic synthetic images, video and voice in campaigns.
• 12. Build oversight into high-risk workflows: review queue, override, stop switch, trained reviewers.
• 13. Log inputs, model version, output, reviewer and outcome; retain at least six months.
• 14. Implement the Article 22 path (explanation, contest, human review) and map Article 21(2) objections to suppression.

Vendors

• 15. Obtain each vendor's written role statement, intended purpose and risk classification.
• 16. For high-risk tools, request instructions for use, declared accuracy metrics, documentation summary, conformity route, EU database registration and monitoring plan.
• 17. Contract for log export, serious-incident notice, model-change notice, audit rights and cooperation with authorities. Ask which harmonised standards the vendor applies.

Evidence

• 18. Keep AI literacy records: who was trained, on what, when.
• 19. Keep the fundamental rights impact assessment (Art. 27) or data protection impact assessment, cross-referenced.
• 20. Calendar 20 November 2026, 2 December 2026, 2 December 2027 and 2 August 2028; name an owner to track the Commission's final classification guidelines.

If you are deciding where credit, insurance or other Annex III decisioning should run before 2 December 2027, a deployment qualification review maps your systems to on-premise, private cloud or hybrid options. Related playbooks are in the regulation hub; financial firms should also read DORA and the marketing stack.