Skip to main content
AI and law

Do You Have to Tell Customers the Answer Comes from an AI? Transparency Requirements in Practice

Distinguish between direct AI interaction, machine-readable marking and the rules for published content. Practical examples for chat, email and documents.

Author: Agent360 editorial team
Published: 29 July 2026
Last updated: 23 October 2026
Reading time: 14 min read

When interacting directly with an AI system, the person must as a starting point be told so, unless it's obvious from the situation. An AI-assisted email draft or document doesn't automatically trigger the same duty. Assess interaction, generated content and the parties' roles separately; below we distinguish between legal requirements and recommended openness.

What the AI Act says about the duty to inform

The EU's Artificial Intelligence Act (the AI Act) contains a number of transparency obligations for certain types of AI system, regardless of whether the system is otherwise classified as high-risk. The purpose is to prevent people from being misled or subjected to imitation without knowing it.

The core of the rule for customer-facing AI systems is simple: if an AI system is intended to interact directly with a natural person, it must be designed so that the person is told they're talking to an AI. That applies unless it's obvious from the point of view of a reasonably well-informed, observant and circumspect person, taking the situation and context into account.

Article 50(1) has a narrow exception for AI systems authorised by law to detect, prevent, investigate or prosecute criminal offences, with appropriate safeguards for the rights of third parties. The exception doesn't apply to systems that are available to the public for reporting a criminal offence. It isn't a basis for omitting disclosure in ordinary commercial customer flows.

When is it "obvious" that it's AI?

"Obvious from context" isn't the same as "we think it's clear ourselves". The assessment must be made from the point of view of a reasonably well-informed, observant and circumspect person, in other words an objective yardstick, not the company's own perception. In practice, that means a discreet icon or a line in the terms is typically not enough to carry the exception on its own. What genuinely makes something "obvious" is usually clear marking in the interaction itself: a name like "AI assistant", a visible label in the chat window, or an opening sentence that makes it clear there's no human at the other end.

If you're unsure whether your flow is clear enough to fall under the exception, the robust route is to disclose it directly rather than lean on an interpretation of "obvious". It's cheaper to write one sentence than to have to defend an interpretation afterwards.

The law enforcement exception

The law enforcement exception in Article 50(1) relates specifically to uses authorised by law and doesn't cover publicly available systems for reporting criminal offences. Ordinary customer communication, customer service, sales and administration must be assessed under the relevant rules for the specific flow.

Does it apply to your customer chat, email and inbox?

It depends on the specific function. Direct AI interaction must be assessed under Article 50(1); content marking follows separate rules. Human approval is relevant, but it isn't a general exemption from all requirements.

A customer chat where the AI answers independently: Here there's direct interaction between the AI system and the customer. The duty under Article 50(1) rests on the provider. If you put a system into service under your own name, you can yourselves be the provider under Article 3(3). Check the information given to the customer and the exceptions described above.

An email where an employee uses AI to write a draft, but reads, edits and sends it under their own name: An internal writing tool isn't, for that reason alone, direct AI interaction with the customer under paragraph 1. Assess separately any rules on content, data protection and misleading practices. Voluntary openness can still be relevant.

Automatic inbox replies and receipts that are pure templates with no AI model behind them: Fall outside the scope of the AI Act, because there is no AI system generating the content at all. Template text isn't the same as AI-generated text.

AI-generated documents or reports sent to the customer as a standalone product: The document format doesn't determine the duty to inform. Examine whether there is direct AI interaction, or whether a separate content rule applies. A clear sender and a statement about AI use can be chosen as good practice.

The assessment has to be made flow by flow. Distinguish between the customer's direct interaction with an AI system, the provider's marking of generated output, and your responsibility for the content you publish or send.

Duty to disclose AI use in typical written customer flows

Situation Duty to inform Why
AI chatbot that answers the customer independently Yes Direct AI-to-person interaction, not obvious without marking
Chat clearly labelled "AI assistant" from the start Met through the labelling The context makes it obvious to an observant person
Employee uses AI for email drafts, reads and sends themselves Not automatically covered by paragraph 1 Internal writing tool; other rules assessed separately
AI-generated report/document sent as a standalone product Depends on function and content The document format doesn't in itself trigger direct interaction
Automatic receipt/auto-reply built on a pure template No No AI system generates the content itself
Law enforcement authorised by law, covered by the paragraph 1 exception Exception with special conditions Doesn't apply to publicly available systems for reporting criminal offences

AI-generated content: marking text and documents

Article 50(2) is aimed at providers of AI systems that generate synthetic audio, images, video or text. Output must be marked in a machine-readable format and be identifiable as artificially generated or manipulated, subject to the provision's technical limits and exceptions. The provider of the AI system isn't automatically the underlying model supplier; the roles follow Article 3. For systems that were on the market before 2 August 2026, the deadline for paragraph 2 is 2 December 2026, cf. Article 111(4). See the Commission's current FAQ on Article 50.

Deepfakes and synthetic content

The deepfake rule covers certain artificially generated or manipulated images, audio and video that may appear genuine. The deployer must disclose the manipulation. For evidently artistic, creative, satirical or fictional works, the duty is limited to appropriate disclosure of the content's existence, without hampering the display or enjoyment of the work; it isn't a full exception. Ordinary AI text isn't a deepfake. For text published with the purpose of informing the public on matters of public interest, paragraph 4 has a separate disclosure rule and, among other things, an exception for human review or editorial control together with editorial responsibility. See Article 50.

Consumer law and the Marketing Practices Act: misleading practices outside the AI Act

The AI Act isn't the only relevant framework. General consumer and marketing law is built on a basic principle that customers must not be misled, whether the deception happens through an AI or in some other way. If a customer would reasonably act differently if they knew the answer came from an AI rather than an adviser, that points towards disclosure being the right course, irrespective of the precise legal classification. The Consumer Ombudsman supervises marketing legislation in Denmark, and companies in doubt about a specific customer-facing AI practice can seek guidance there.

The GDPR perspective: why transparency is linked to data protection

Transparency about the use of AI is often linked to another question: what happens to the customer's data when the AI processes an enquiry? The data protection rules set their own transparency requirements towards the data subject: including the right to be informed about how their information is processed and whether automated processing takes place. It's a separate set of rules with its own conditions, and it doesn't mean that every use of AI in customer contact automatically triggers the same obligations as the AI Act's transparency article. But in practice it's wise to think of the two together: when you have to describe your AI use to the customer anyway, it makes sense to be clear about how data is processed and where it's stored as well. The Danish Data Protection Agency is the relevant authority for that part in Denmark and has rules and guidance available for companies that want to familiarise themselves with the requirements.

To be able to explain the data processing, you need to know the location, access and sub-suppliers in the specific set-up. Your own instance doesn't on its own settle these matters and isn't a legal guarantee of compliance.

How to build the duty to inform into your written AI flows

The practical task isn't just to write a sentence about AI in the terms. It's to build the disclosure in where the customer actually meets the AI.

Chat and inbox

In a customer chat, the simplest solution is a clear label at the start of the conversation ("You're chatting with our AI assistant" or similar) together with the option to be passed on to a human. The same applies if the AI answers enquiries in a shared inbox: the recipient should be able to see that the reply is AI-generated, either in the text itself or in a visible field around it.

Emails and documents

When an employee uses AI as an internal writing tool, direct AI interaction with the customer doesn't automatically arise. For documents and reports, you need to examine relevant content rules separately. You can also choose a visible explanation of AI use as good practice; the recommendation isn't a statement of a general marking duty for all documents.

Your own instance and data ownership

When you choose a platform for this kind of written AI flow, it's worth putting weight on how easy it is to document and explain your set-up if a customer asks. Your set-up should make it possible to explain where data is processed and who has access. Your own systems and your own instance can also use sub-suppliers, whose terms and data processing need to be part of the explanation. It doesn't change the fact that you still have to decide for yourselves when the duty to inform the customer is triggered, but it does make the basis for answering precisely stronger.

What the AI should NOT do

  • The AI shouldn't decide for itself whether a new type of customer flow falls under the exception "obvious from context": that's a legal assessment that people in the company must make, preferably with an adviser, when in doubt.
  • The AI shouldn't omit marking in order to "seem more human" or to improve an experience metric. That is precisely the type of behaviour the rules were created to prevent.
  • The AI shouldn't be used to get round consumer protection or marketing rules by phrasing itself so that the customer wrongly believes they're dealing with human advice or judgement.
  • As internal practice, the AI shouldn't send content to customers until you've clarified human oversight and the relevant disclosure. The specific legal obligation depends on function and content.
  • The AI shouldn't be used to give legal advice on whether your specific set-up complies with the rules. That requires a specific assessment, possibly with an external adviser, not a general text like this one.

Practical checklist before you go live

  1. Map where AI interacts directly with people, where it's an internal writing tool, and where you publish generated content.
  2. For each flow, assess whether it's "obvious" to an ordinary, observant customer that it's AI, or whether it requires explicit disclosure.
  3. Build the marking into the interaction itself, not just the terms. A visible label in the chat or a line in the email is cheaper than a dispute afterwards.
  4. Agree internally who is responsible for watching whether new AI flows are rolled out without going through this assessment first.
  5. Document your assessment and choices, so you can show that you've actively taken a position, even if you conclude that a specific situation falls under the exception.
  6. Clarify roles, required data processing agreements and data storage in parallel with the transparency assessment.
  7. Seek guidance from the Danish Data Protection Agency or the Consumer Ombudsman in specific cases of doubt. This article is not legal advice.

In short: the rule isn't hard to understand, but it requires you to actually go through your written AI flows one by one, rather than handling it with one sentence in the general terms.

Frequently asked questions

Q: Is there a fine if we don't tell the customer about AI use? The Regulation contains a system of sanctions for infringements, but the specific amounts and enforcement practice depend on the individual case and the authority's assessment. If you're unsure about the consequences in a specific situation, you should seek legal advice or contact the relevant authority.

Q: Do we have to use particular wording to comply with the requirement? The Regulation doesn't prescribe a fixed wording. The requirement is that the person is told they're interacting with an AI system, in a way that genuinely makes it clear. A clear, easily understood sentence or label is better than technical wording hidden in the terms.

Q: Does the duty to inform also apply to B2B customers, not just consumers? The wording of the Regulation refers to natural persons interacting with the AI system; it isn't limited to consumer relationships. A contact person at a business customer who chats with your AI system is also a natural person in this context.

Q: What if our AI only produces drafts that an employee always approves before sending? When AI is used as an internal writing tool, it isn't, for that reason alone, direct AI interaction with the customer under Article 50(1). Human approval does not, however, automatically exempt you from other rules on content, data protection or misleading practices.

Q: Do we also have to mark AI-generated content internally, not just for customers? The Regulation's focus in this context is interaction with natural persons and the marking of generated content generally. Good practice, though, is to keep track internally of which documents and texts are AI-generated, whether or not they go to an external customer, so you can always answer precisely if anyone asks.

Q: How do we document that we've taken a position on the rules? There's no fixed template prescribed in this article, but a written overview of your customer-facing AI flows, your assessment of each one (disclosed/not disclosed/why), and who is responsible is a sensible practice to be able to produce if the question arises.

How this fits with the rest of your AI work

Transparency towards the customer is naturally linked to how you build AI into the company at all. If you're in the process of clarifying what an AI agent for business actually covers, the duty to inform is one of the practical questions that comes with the AI starting to talk directly to customers. The same applies if you work with CRM automation, where AI-generated notes and follow-ups often end up in customer-facing flows without anyone having thought about the marking. And if your AI use is tied to data hygiene in the CRM, it makes sense to take the transparency requirement into the same clean-up, rather than treating it as a separate project.

The rule is easy to summarise, but it requires someone in the company to actually go through the flows: where does the AI talk directly to a customer, is it obvious from the context, and if not, is the marking built in where the customer actually sees it? It's cheaper to do now than to clean up afterwards.

This article is general information and does not constitute legal advice. Contact a lawyer or the relevant authority for an assessment of your specific situation.

Tags: AI and law, the AI Act, transparency, customer service, compliance

Unsure about the rules for your own set-up?

We look at what you use today and what it requires: documentation, consent and what the EU AI Act means for you. 20 minutes, no commitment.

How we built it · Email & prioritisation

The inbox, sorted

An inbox where AI sorts the emails, writes the reason next to each email and can draft a reply based on the customer’s case.

  1. Email in
  2. AI sorts
  3. You reply
See it from inbox to follow-up →