Does the EU AI Act apply to AI agents that send email?

Yes. The Commission’s Article 50 guidelines name AI agents that manage correspondence, and give as an example of disclosure an email written by an AI agent with an AI label at the top.

What is due on 2 December 2026?

Machine-readable marking of AI-generated text under Article 50(2), for AI systems placed on the market before 2 August 2026 (Article 111(4)). Disclosure under Article 50(1) has applied to everyone since 2 August 2026.

Who is responsible: the model maker or the company running the agent?

Usually the company running it. The provider is whoever puts the AI system into service under its own name, including a company that builds an agent in-house for its own use.

Does a person reviewing the agent’s drafts remove the duty to disclose?

Not by itself. The guidelines say the mere possibility of human review should not be used to avoid disclosure. A message a person writes and sends with AI as a tool is a different case.

Does Pidgeon add an AI disclosure to agent email today?

Not yet. Signed, machine-readable disclosure is being built into Pidgeon; until it ships, the agent’s own message has to carry the disclosure. Watermarking the text itself is for the provider of the model that writes it.

AI agents that email people: what the EU AI Act requires by 2 December

Q: What is the fine for breaching Article 50?

Up to €15 million or 3% of worldwide annual turnover, whichever is higher. For SMEs and start-ups, whichever is lower.

The short version

Article 50 of the EU AI Act has applied since 2 August 2026. For an AI agent that writes email to people, it asks two things of the agent's provider:

Say it is an AI. A person the agent writes to must be told, clearly and at the latest in the first message, that they are dealing with an AI system. The Commission's guidelines add that an agent should also say on whose behalf it acts.
Mark what it writes. AI-generated text must carry a machine-readable mark that software can detect. For systems already on the market before 2 August 2026, this part is due by 2 December 2026.

Breaching either can cost up to €15 million or 3% of worldwide annual turnover, whichever is higher; for SMEs and start-ups, whichever is lower.

This page explains both duties as they apply to email, from the Act's text and the Commission's guidance as read on 30 October 2026, and ends with a checklist. It is not legal advice.

Who the duties fall on

Article 50(1) and 50(2) are duties of the provider: the company that places an AI system on the market or puts it into service under its own name. The Commission's guidelines give two examples that fit most agent teams. One is a company offering an AI agent to customers under its own brand, free or paid, wherever it is based. The other is a company that builds an agent in-house and runs it for its own use under its own name. Building on another company's model does not move the duty: if you put the agent into service under your name, you are its provider.

Deployers, who use an AI system under their authority, have separate duties. For text, the one that matters is Article 50(4): text published to inform the public on matters of public interest must be disclosed as AI-generated, unless a person reviewed it and someone holds editorial responsibility. A one-to-one email is not usually published to the public, so for agent email the provider duties are the ones to plan for.

Duty one: disclose the AI, from the first message

Article 50(1) requires interactive AI systems to be designed so that the people interacting with them are informed they are dealing with an AI, unless that is obvious to a reasonably well-informed person. Article 50(5) says how: "in a clear and distinguishable manner at the latest at the time of the first interaction or exposure", meeting accessibility requirements.

The guidelines (C(2026) 5054, 20 July 2026) settle three questions for email:

Email counts as interaction. It is named among the channels through which people give input to an AI system, and interaction "does not have to be initiated by a human". An agent that writes first and invites a reply is in scope.
Agents are named. AI agents are covered when they interact with people while carrying out tasks, and "managing correspondence" is one of the examples given. They should disclose "both their artificial nature and the person on whose behalf they are acting". Where a provider cannot know in advance whether a person will read the output, the agent should disclose itself wherever that is reasonably likely, including when the person acts for a company.
The example is an email. Among the ways to disclose, the guidelines give "an email generated by an AI agent sent to a natural person that features an AI label at the top". They also name disclosing "AI identifiers, and credentials", with agents disclosing their AI identity "to the extent feasible in a verifiable manner".

What is out of scope: traditional out-of-office replies and rule-based auto-responses, because they are not AI systems; and machine-to-machine exchanges whose output is not meant to reach a person. A person reviewing the agent's drafts does not remove the duty on its own: the guidelines say the mere possibility of human intervention should not be used to get around disclosure. When a person genuinely writes and sends the message using AI as a tool, the interaction is mediated, and Article 50(1) does not apply to it.

Duty two: mark the text, by 2 December

Article 50(2) requires providers of systems that generate synthetic text to mark the output "in a machine-readable format and detectable as artificially generated or manipulated". The technical solutions must be "effective, interoperable, robust and reliable" as far as technically feasible.

For agents, the guidelines are specific: Article 50(2) applies when the agent's action produces AI-generated content that people perceive, such as text. An email the agent wrote qualifies. Its reasoning steps and its web requests do not.

The date. Article 113 makes Article 50 apply from 2 August 2026. Article 111(4), added by the Digital Omnibus, gives providers whose systems were on the market before that date until 2 December 2026 to comply with the marking duty. It covers marking only: disclosure has applied to everyone since 2 August. A system launched after 2 August had to mark from the start.

Exceptions. Marking is not needed where the system only assists standard editing or does not substantially alter the input, such as fixing spelling in a person's own text. The guidelines also describe a narrow business-to-business exception, but every one of its conditions must hold: the output is strictly technical, read only by a limited set of professionals inside the provider's and deployer's organisations, and never meant to leave the company. An email to a customer, a supplier or a candidate fails the last condition.

How to mark. The Act's recitals list watermarks, metadata, cryptographic proofs of provenance and authenticity, logging, and fingerprints. The Commission's Code of Practice on marking and labelling AI-generated content, finalised in June 2026, is the voluntary route to showing compliance, and the Commission and the AI Board have confirmed it as adequate. In practice the mark is applied mostly where the text is generated: for long text the Code looks to watermarks embedded by the model, which the provider of the generating system supplies and no mail header can add afterwards. So ask your model provider how its output is marked. A machine-readable disclosure in the email itself, such as a signed header saying an agent wrote it and for whom, complements that mark and carries the Article 50(1) disclosure in a form software can read; it does not replace a watermark.

Keeping records

Article 50 does not itself require a log. But each of the duties above is something a provider may have to show it met, and enforcement sits with national market surveillance authorities. So keep evidence of what the agent sent, to whom, with what disclosure and what mark, and which person or company it acted for. The simplest record is the sent mail itself, kept, with the headers intact.

A checklist for agents that email people

Decide who the provider is. If your company puts the agent into service under its name, it is you.
List where the agent writes to people. Customers, suppliers, candidates, the public, and people who act for companies all count.
Disclose at the top of the first message. A plain line such as "This email was written by an AI agent acting for Acme Ltd." Repeat it when the agent starts a new conversation; the guidelines expect disclosure at every new interaction.
Say whom it acts for. The company, and where relevant the person, on whose behalf the agent writes.
Check how the text is marked. Ask your model provider how its output is watermarked, and add a machine-readable disclosure to each message, ideally signed so it cannot be stripped or forged unnoticed. Marking is due by 2 December 2026 for systems already on the market.
Don't rely on review to escape disclosure. A person able to edit the draft does not make the agent's message human-written.
Keep the evidence. The sent messages, their headers, and which agent and operator sent each one.
Check out-of-scope cases rather than assume them. Rule-based auto-replies are out; an AI that drafts a reply sent automatically is in.

Talk to us

We are building signed, machine-readable disclosure into Pidgeon, so that an agent's mail says what it is and whom it acts for by default. If you run agents that email people in the EU and want to talk through the 2 December deadline, write to us.

Sources

All read on 30 October 2026.

Regulation (EU) 2024/1689 (the AI Act), Article 50, Article 99, Article 111 and Article 113, in the consolidated text published by the Commission's AI Act Service Desk, as amended by the Digital Omnibus on AI.
European Commission, Guidelines on the implementation of the transparency obligations for certain AI systems under Article 50, C(2026) 5054, 20 July 2026: points 11, 30, 31, 36, 63, 73 and 87.
European Commission, Transparency obligations under Article 50 of the AI Act: FAQ.
European Commission, Code of Practice on AI-generated content.