/

Updated 18 November 2026

What Agencies Must Do Under the EU AI Act (2026 Guide)

AI Disclosure: This content includes AI-generated text.

Short answer: if your agency uses AI to create content for clients, or builds AI features into client websites, the EU AI Act gives you duties of your own, not just advice to pass on. Depending on the project you can be a deployer (you use AI tools to produce work) or a provider (you build an AI system and put it into service). Since 2 August 2026, Article 50’s transparency rules apply to both roles, and clients increasingly expect their agency to handle them.

This guide covers how to work out your role per project, what that means for deliverables, and how to put it into contracts and workflows without slowing projects down.

Step 1: work out your role, per project

The AI Act assigns duties by role, and the same agency can hold different roles on different projects.

What the agency doesLikely roleMain duties
Writes blog posts or copy for a client with ChatGPT or ClaudeDeployer of the AI toolPublic-interest text and deepfake rules, AI literacy
Creates AI images or video for a campaignDeployerDeepfake disclosure where content looks real
Installs and configures a chatbot widget (Tidio, Intercom and similar) on a client siteUsually neither provider nor deployer: the vendor is provider, the client is deployer. You configured it.Make sure the setup discloses the AI, because you wrote the greeting
Builds a custom AI assistant on a model API for a clientThe client is often the provider if it runs under the client’s name; the agency can be provider if it offers the system as its own productAI interaction notice built in, documentation
Sells its own AI tool or plugin to many clientsProviderAll provider duties for that system

Two points that trip agencies up. First, being “just the developer” doesn’t remove responsibility for the design: if the system you deliver can’t tell users it’s an AI, it was built wrong. Second, the client being the legal provider or deployer doesn’t stop them from asking you why you didn’t flag it. Write the role down in the project brief.

Step 2: what changes in your deliverables

Content (text)

AI-generated text published to inform the public on matters of public interest needs a disclosure, unless a person with real expertise reviewed it and someone holds editorial responsibility. For agencies this matters most with clients in law, finance, health, energy and public affairs. You have two clean options per piece: deliver it labelled, or deliver it with a documented substantive review and let the client take editorial responsibility. A spelling check doesn’t count as review. Our Article 50 content test has the details.

Images and video

Realistic AI images of people, places, products or events that would falsely appear authentic are deepfakes and need a visible disclosure. Typical agency risk areas: fake “team” or “customer” photos, staged product shots showing features that don’t exist, and AI-edited photos of real locations. Stylised illustrations are fine without a label. Put a visible label in the deliverable or the publishing spec, because the invisible watermark from the image tool doesn’t satisfy the deployer duty.

Chatbots and AI features

Every chatbot or voice assistant you deliver should identify itself as an AI at the first interaction. Build it into your setup checklist: greeting text, header label, human hand-off messages. See our chatbot compliance guide for wording.

Personalisation and recommendations

AI-driven product recommendations and personalised prices usually aren’t Article 50 issues, but consumer and platform law adds its own transparency duties. Our ecommerce guide covers them.

Step 3: put it in the contract

Most disputes come from nobody deciding who labels what. Add a short AI clause to your statements of work:

  • Which AI tools the agency uses, and for which deliverables.
  • Who reviews AI-generated text on public-interest topics, and who holds editorial responsibility for publication (usually the client).
  • That deepfake-type visuals will be labelled, and who places the label on the live site.
  • For AI systems you build: who is the provider, and that AI disclosure is part of the delivered system.
  • That the client informs the agency if it changes a delivered disclosure.

Also check your client contracts and NDAs for restrictions on feeding client data into AI tools, and your AI vendors’ terms on training with your inputs. That’s a GDPR and confidentiality issue rather than an AI Act one, but clients ask about it in the same breath.

Step 4: build it into your workflow

  1. Project intake: ask whether the client is in a public-interest sector and whether the site has or will have a chatbot.
  2. Production: tag every AI-assisted asset in your project tool. It takes seconds and makes the later audit trivial.
  3. Review: for public-interest text, a subject-matter reviewer signs off with name and date. Re-review if AI rewrites the draft afterwards.
  4. Launch checklist: add three lines: chatbot discloses AI, deepfake visuals labelled, public-interest AI text reviewed or labelled.
  5. Handover: give the client a one-page record of AI use and disclosure decisions. Our website AI audit template works for this.

Step 5: train the team (Article 4)

Since 2 February 2025, providers and deployers must take measures to support the AI literacy of people who operate AI on their behalf, and that includes freelancers working for you. For an agency, a practical minimum is a short internal guide on approved tools, what not to paste into them, how to fact-check AI output, and when a label is required. Keep a record of who received it.

Turning compliance into a service

Many clients don’t know any of this yet. An “AI transparency check” is a natural add-on to maintenance retainers: audit the site, fix chatbot greetings, label the right images, and set up a disclosure process. It’s quick to deliver and easy to explain, because the deadline has already passed.

For WordPress client sites, the free EU AI Act Ready plugin gives you a repeatable setup: per-post disclosure levels your editors can set, AI image labels, notices for common chat widgets, and a compliance report you can hand to the client. It’s GPL, has no account or licence keys to manage per site, and runs on the client’s own server.

Frequently asked questions

Can the agency be fined, or only the client?

Fines attach to whoever breaches an obligation in their role. If the agency is the provider of a system or the deployer that generated a deepfake, the duty and the risk are the agency’s. Up to 15 million euros or 3% of turnover applies to Article 50 breaches, with the lower cap for SMEs. See fines and penalties explained.

Do we need to tell clients we use AI?

The AI Act doesn’t require you to tell clients. Your contracts, confidentiality duties and honesty with the client may. In practice, being open about it avoids disputes, and many clients now ask in procurement.

Is an agency building a chatbot for a client a “provider”?

It depends on whose name the system is put into service under. If the client commissions it and runs it as its own, the client is typically the provider. If you offer the chatbot as your own product to several clients, you are. Either way, build the AI disclosure in.

Does this apply to agencies outside the EU?

Yes, when the output is used in the EU, for example on a client site aimed at EU visitors.

Practical guide: EU AI Act checklist for agencies.