EU AI Act Makes AI Agents Say Who Sent Them

The European Commission’s final guidance on Article 50 of the AI Act gives AI agents a disclosure problem that is more precise than a generic “powered by AI” badge. When an agent contacts a person to book a service, manage correspondence, negotiate, conclude a contract or make a purchase, the Commission says it should identify itself as artificial and disclose the person on whose behalf it is acting.

That turns transparency into a runtime identity and delegation issue. The question is whether an agent crossed from a machine workflow into a human interaction, whose authority it carried and whether disclosure survived the handoff.

The obligations in Article 50 apply from August 2, 2026. The Commission published the final guidelines on July 20, less than two weeks before that date. The guidance is non-binding. It states the Commission’s first interpretation of the law, while authoritative interpretation remains with the Court of Justice of the European Union. That boundary matters: companies should treat the document as the enforcement roadmap now available, not as a final judicial ruling on every agent design.

The authority: Article 50 reaches interactive AI

Article 50 of the AI Act creates separate duties. Providers of directly interactive systems must inform people they are dealing with AI unless that is obvious in context. Generative-system providers face machine-readable marking duties, while deployers face distinct rules for emotion recognition, biometrics, deepfakes and some public-interest text.

The agent question sits mainly under Article 50(1), the direct-interaction provision. In the final 51-page guidance, the Commission describes four elements: there must be an AI system, an intended interaction, a direct exchange and a natural person. A single prompt and reply can qualify. So can an AI-initiated message that seeks a human response. Interaction can occur through text, voice, email, a physical action or another channel that permits a genuine responsive exchange.

The guidance draws a useful line around “direct.” A back-end tool that sends suggestions to a customer-service employee is not necessarily communicating directly with the customer. A machine-to-machine call between agents, with no output intended to reach a person, also falls outside this particular disclosure duty. But the analysis changes when the AI itself sends the message, makes the call or acts in an environment perceived by a person.

For agent developers, that is the operative boundary. The same orchestration can move in and out of Article 50(1) as a task progresses. An agent may retrieve data from another agent without a human-facing disclosure event, then trigger one when it emails a supplier or negotiates with a customer.

The mechanism: disclose the agent and its principal

The Commission goes beyond a bare AI-origin notice for agents. It says agents capable of making bookings, managing correspondence, negotiating or concluding contracts, or executing purchases should disclose both their artificial nature and the person on whose behalf they act. The stated reason is to make the origin of the interaction, delegated authority and accountability for resulting actions visible.

That is a material product requirement. “I am an AI assistant” answers only half of the Commission’s formulation. A recipient may also need to know whether the agent represents a consumer, an employer, a merchant or another organisation before deciding how much authority to attribute to its request.

The guidance also treats disclosure as more than a one-time onboarding screen. An agent should identify itself to the person instructing it at key points such as authorisation, validation and reporting. It should disclose itself at every new interaction. If it relies on another AI system’s output, that fact belongs in the control logic around the instructing user even when the downstream model never contacts a person directly.

Multi-agent architecture does not remove the issue. Where a provider cannot reliably know in advance whether an agent will reach a person, the Commission says the agent should be designed and instructed to disclose itself whenever human interaction is reasonably likely. In a chain of specialised agents, the disclosure state therefore has to survive routing decisions rather than depend on the first interface alone.

This is the same practical divide between capability and authority that TECHi examined in AWS OpenSearch agent production controls. Authentication can let an agent reach a system. It does not by itself explain to an outside person whose instructions the agent is carrying or what the agent has authority to commit.

A human somewhere in the loop is not enough

The guidelines close another easy workaround. A provider cannot avoid Article 50(1) merely because a human could intervene or review an output. In a service that blends AI-generated responses with human-curated material, the specific AI outputs remain within scope unless they are properly reviewed and sent by a human acting as the main interlocutor.

That distinction places weight on the actual communication path. A human approval button may matter if a qualified person examines the response and sends it under their own responsibility. A nominal escalation option, with the AI still addressing the recipient, does not convert the interaction into a human conversation.

The format of the notice must also work in context. The Commission says disclosure hidden in terms and conditions, a URL, documentation or machine-readable metadata is insufficient by itself. Generic references to an “assistant,” a broad notice that a website uses AI, or a technical statement that a service uses an LLM may also fail to explain the artificial origin of the specific interaction.

In many ordinary cases, one prominent notice before the first interaction may be enough. The guidance nevertheless points to periodic or context-aware reminders for sustained or riskier exchanges, including financial, insurance, legal, health and complaints settings, and for systems involving vulnerable users or possible emotional dependence. Voice systems may need spoken disclosure; children require age-appropriate information; accessibility rules still apply.

The result is not a universal banner specification. It is a context test. Product teams need to know when a session begins, when a new person enters, when the agent’s role changes and when a recipient could reasonably mistake the agent for a human.

Who carries the obligation inside a company

Role assignment is equally consequential. A provider is the entity that develops an AI system, or has it developed, and places it on the EU market or puts it into service under its own name or trademark. A company that takes an existing system, modifies it with new training data and releases the resulting system under its own name can become a provider of the new system. An organisation that builds an internal chatbot and puts it into service under its own name can also be a provider.

A deployer is the person or organisation that decides to use the system under its authority and determines its purpose and manner of use. The Commission says authority does not require low-level technical control. It can exist in decentralised workflows and corporate groups.

Employees are not treated as separate deployers when they use a system under their employer’s control. Nor does an organisation automatically transfer its deployer role to contractors or freelancers operating the system on its behalf and under its responsibility. By contrast, a company that merely commissions an agency to produce an advertisement, without deciding whether or how the agency uses AI, is not a deployer on that fact alone.

For agent businesses, that makes procurement language operational. A vendor may provide the model, an integrator may assemble the workflow, a customer may define the purpose, and an outside agent may contact a fourth party. Contracts can allocate implementation work, but the Commission’s role analysis follows who places the system under a name and who controls its use.

The geographic boundary is broader than EU incorporation. A non-EU provider or deployer can fall within scope when its output is used in the Union, although unforeseeable channels outside a third-country deployer’s control do not automatically produce the same result.

What changes for agent product teams

The final guidance does not prescribe a database schema or an identity protocol. Its logic nevertheless points to five pieces of runtime state that mature agent systems will need to preserve: the agent’s artificial identity, the principal it represents, the authority granted for the current action, the human interaction boundary, and evidence that the required disclosure occurred.

Those elements become harder to maintain as agents gain commercial power. TECHi’s report on Solana-based agent payments examined agents buying API access without a normal account or subscription. Article 50 adds a separate question when a purchasing agent leaves that machine channel and deals with a person: who authorised the spend, and what identity must the recipient see?

The same issue appears at the operating-system layer. The EU’s binding measures opening Android features to rival AI assistants concern agents that can hear context and act for users. As access expands, the difference between an assistant offering advice and an agent exercising delegated authority becomes more important, not less.

For companies, the near-term work is therefore a workflow audit rather than a logo exercise. Teams need to map every channel in which an agent can reach a person, identify the legal entity that provides and deploys the system, define the represented principal, test disclosure across handoffs, and decide when a human becomes the true interlocutor. Those are engineering, governance and legal questions at once.

The Commission notes one possible technical direction: electronic attestations available through EU Digital Identity Wallets, and proposed European Business Wallets, could help verify agent identity, attributes and authorisations. It presents that route as an example, not a mandate. No single credential standard is imposed by the guidance.

The August deadline is sharper for interaction than marking

The timing can be misunderstood because Article 50 contains different duties. The Commission’s official FAQ says the rules apply from August 2. A limited transition until December 2 is available for the machine-readable marking obligation under Article 50(2) for qualifying generative systems already on the market before August 2.

That transition does not postpone the direct-interaction disclosure under Article 50(1). A system that is both interactive and generative may receive extra time for qualifying marking work, while still needing to disclose AI interaction from August 2.

Non-compliance with Article 50 can draw fines up to €15 million or 3% of worldwide annual turnover for an undertaking, whichever is higher. For small and medium-sized enterprises, including startups, the applicable ceiling is the lower of the fixed amount and percentage. National market-surveillance authorities, the AI Office within its remit, and the European Data Protection Supervisor for EU institutions have enforcement roles.

The penalty figure is significant, but it should not obscure the narrower immediate point. An agent company waiting for the December marking transition may still miss an August disclosure obligation that applies to its chatbot, voice agent or autonomous correspondent.

What remains unresolved

The Commission calls these guidelines a first interpretation and says it may revise or withdraw them as implementation experience, enforcement and CJEU decisions develop. Several operational questions therefore remain open.

The phrase “person on whose behalf” is clear at the edges but harder in layered services. An agent may be offered by a model company, configured by an enterprise, instructed by an employee and sent to negotiate with a supplier. The guidance establishes the disclosure objective, but it does not supply a universal method for resolving every principal in that chain.

Interaction boundaries can also be disputed. A new recipient is plainly important, while a reopened session, a transferred thread or an agent that changes roles during one workflow may require contextual judgment. The Commission recommends reminders in higher-risk settings but does not set a fixed interval.

Finally, the voluntary Code of Practice mainly helps providers and deployers demonstrate compliance with the marking and labelling duties under Article 50(2), (4) and (5). The official FAQ says providers can determine adequate measures for direct-interaction disclosure under Article 50(1), taking the guidelines into account. Signing the Code therefore does not replace the design decision for agent identity.

The measured conclusion is firm. The final guidance does not make an AI agent a legal person, and it does not settle every question of liability. It does make the Commission’s expected disclosure architecture visible before August 2: when software speaks or acts directly to a person, it should not present only a capability. It should identify the artificial actor and the authority behind it.