The EU AI Act for Marketing Agencies: The 2026 Compliance Checklist

The EU AI Act (Regulation (EU) 2024/1689) does not classify marketing agencies as a category of their own in law: it recognizes neither the term "agency" nor "AI agent". What matters is your role: almost every DACH agency is a deployer under Art. 3(4), because it uses AI systems under its own responsibility. It becomes a provider under Art. 3(3) only in narrowly defined cases, for example by offering a system under its own brand or, for high-risk systems, via Art. 25. The obligations follow from this.

• Default role of deployer: Art. 4 AI literacy and Art. 50 transparency. The obligations under Art. 26 (human oversight, monitoring, log retention of at least six months) apply only if the agency deploys a high-risk AI system, for Annex III systems from 2 December 2027 (for AI in products under Annex I from 2 August 2028).
• Already in force: AI literacy (Art. 4) and the prohibitions in Art. 5 have applied since 2 February 2025, regardless of risk and not postponed. Only the new Digital Omnibus ban on AI systems for non-consensual intimate imagery and child sexual abuse material applies later, from 2 December 2026.
• Labeling obligation: Art. 50 has applied since 2 August 2026. Providers must make chatbots recognizable as AI and mark AI outputs in a machine-readable format, while deployers must disclose deepfakes and AI-generated text on matters of public interest published without editorial responsibility. Machine-readable marking by generative systems that were already on the market before 2 August 2026 has a grace period until 2 December 2026.

What role does a marketing agency have: deployer or provider?

The AI Act distinguishes the operator roles in Art. 3(3)–(8). Two are relevant for agencies:

A deployer (Art. 3(4)) is anyone who uses an AI system under their own authority. This is the default role for almost every agency that uses ChatGPT Enterprise, Claude for Work, Microsoft 365 Copilot, Gemini for Workspace or a managed RAG service for copywriting, campaigns, customer service or knowledge management. As long as you use the system for its documented purpose, do not fine-tune or substantially modify the underlying model, do not rebrand it and do not repurpose it towards a new high-risk use, you remain a pure deployer.

A provider (Art. 3(3)) is anyone who develops an AI system, or has one developed, and places it on the market under their own name or trademark. A deployer can also tip into this role, with considerable consequences. Three triggers:

1. Own brand (Art. 3(3)): if you have a white-label chatbot system developed or adapted for you and offer it under your own brand, you are usually its provider, and the disclosure obligation under Art. 50(1) then falls on you. If the system is high-risk, Art. 25(1)(a) governs the change of role, unless a contract allocates the obligations differently.
2. Substantial modification (Art. 3(23), Art. 25(1)(b)): for example, fine-tuning the model in a high-risk system on proprietary data in a way that affects compliance under Art. 8–15.
3. Change of purpose to high-risk (Art. 25(1)(c)): for example, a model fine-tuned on CV data for candidate selection: this is Annex III(4) high-risk, and the agency becomes a provider.

Standard RAG over your own documents, prompt engineering and system-prompt adjustment typically do not trigger Art. 25. Likewise, light LoRA fine-tuning below roughly a third of the original training compute remains under the GPAI threshold set out in the Commission guidelines of 18 July 2025.

The compliance checklist for marketing agencies

The following tick-box checklist bundles what an agency should prepare as a deployer. Rows that rely on Art. 26 are mandatory only for high-risk use cases and otherwise best practice. How labeling works day to day is covered in our article on labeling AI content under Art. 50 in practice.

Two practical points to add: if an AI application monitors or evaluates the performance of employees (for example, a productivity-monitoring copilot), it usually counts as a high-risk system under Annex III point 4. From 2 December 2027, Art. 26(7) then requires informing workers' representatives and the affected workers before the system is put into service. In Austria, the works council's participation rights under §§ 91, 96 ArbVG apply as well. And against shadow AI (private ChatGPT use in the browser that bypasses the governance model), an acceptable-use policy helps, backed by network and identity controls.

The state of the deadlines in 2026, and what already applies

The Digital Omnibus has been adopted: Regulation (EU) 2026/1744 was published in the Official Journal on 24 July 2026 and has been in force since 27 July 2026. It postpones the bulk of the high-risk obligations: Annex III high-risk now applies from 2 December 2027, and Annex I product safety from 2 August 2028. The other key dates are in our AI Act timeline.

Not postponed are Art. 4 (literacy) and Art. 5 (prohibitions), both since 2 February 2025, the GPAI rules (Art. 51–55) since 2 August 2025, and the fine framework. One addition there is the new Art. 99(6a): for fines under Art. 99(4) and (5), small mid-caps, like SMEs, now face the lower of the two amounts. The Omnibus did, however, soften Art. 4: providers and deployers must take measures to support the AI literacy of their staff, but they do not have to guarantee a specific level. The transparency obligations under Art. 50 have applied since 2 August 2026. A grace period until 2 December 2026 exists only for machine-readable marking under Art. 50(2) by generative AI systems placed on the market before 2 August 2026.

Prohibited practices in marketing (Art. 5)

Agencies should take Art. 5 seriously, since this is where the highest sanction tier lies. Three prohibitions are particularly relevant for marketing:

• Art. 5(1)(a): subliminal, manipulative or deceptive techniques causing significant harm, for example persuasion-optimized marketing agents that exploit cognitive biases below the threshold of conscious awareness.
• Art. 5(1)(b): exploiting vulnerability on the basis of age, disability or socio-economic situation, e.g. agents that target older customers with predatory financial products.
• Art. 5(1)(g): biometric categorization based on sensitive attributes, relevant for targeting and segmentation agents.

In addition, the Digital Omnibus bans AI systems intended to generate non-consensual intimate imagery or child sexual abuse material, or lacking reasonable safeguards against this. The ban applies from 2 December 2026 and also covers deployers that use an AI system for the purpose of generating such content.

Example scenario with figures: a 60-person agency in Vienna

An owner-managed performance marketing agency with 60 employees uses four AI applications: a copywriting agent (internal), a customer service chatbot on its own website, a code-generation assistant and, occasionally, AI-generated product videos for campaigns.

Classification:

• Copywriting agent: minimal risk. Machine-readable marking under Art. 50(2) is the duty of the tool's provider: the vendor for a bought-in tool, the agency itself if it builds the agent on a model API and puts it into service for its own use (Art. 3(3) and 3(11)). As a deployer, the agency must disclose AI use only for texts on matters of public interest published without editorial review (Art. 50(4)).
• Customer service chatbot: limited risk, Art. 50(1): users must be able to tell that they are chatting with an AI assistant.
• Code generation: minimal (internal use).
• Product videos: limited risk + Art. 50(4) deepfake disclosure in full; the artistic-satirical exemption is interpreted narrowly.

Effort: for the 60-person agency, the project stays manageable because none of the four applications is classified as a high-risk system. The focus is on the AI literacy program (3–5 modules are typically sufficient for an operation of this size), plus the use-case register, the design of the Art. 50 disclosures and the vendor due diligence.

Risk of doing nothing: if, for example, the deepfake labeling were omitted, this would fall into the second fine tier under Art. 99, up to EUR 15 million or 3% of worldwide annual turnover. As this is an SME, Art. 99(6) applies: the lower of the two amounts governs. With an annual turnover of EUR 6 million, 3% equals EUR 180,000, so the cap here would be EUR 180,000 rather than EUR 15 million. The AI Act itself sets no fine for a lack of AI literacy (Art. 4); whether and how a breach is penalized is up to the member states under Art. 99(1).

For agencies and B2B decision-makers

Anyone using AI productively in marketing is already under obligation: AI literacy and the prohibition of manipulative practices have applied since February 2025, and the labeling of AI-generated content since August 2026. The fastest lever is a documented AI inventory with clean risk classification. It shows your training needs, your labeling obligations and whether you slip into the provider role anywhere. Blck Alpaca supports agencies and B2B teams across the DACH region in building this inventory, implementing Art. 50 labeling and establishing an audit-proof AI literacy program.

Legal notice: This article serves as professional orientation and does not constitute legal advice. For a binding assessment of your specific individual case, please seek specialized legal advice.