Article 50 of the EU AI Act requires that people are told when they are interacting with an AI system, and that certain AI-generated content is marked. For organisations running agents, staff must be trained to keep that promise at every handoff: what the person was told, how to answer "am I talking to a machine?", and what to record.
What does Article 50 ask of the people around an agent?
Article 50 is a transparency rule. Its core duty is that AI systems meant to interact directly with people are designed so those people are informed they are dealing with an AI system, unless that is obvious to a reasonably well-informed person in context. It also requires certain synthetic content to be marked as AI-generated, and puts disclosure duties on deployers who publish some kinds of generated or manipulated content.
The design duty sits mainly with providers. The practice sits with the deployer's staff. A notice in the chat window means little if an adviser who takes over the conversation lets the customer believe they spoke to a person all along, or if an outbound voice agent is introduced by a team lead as "our new colleague". Article 50 is kept or broken in those moments. The fuller treatment in Article 50 transparency for AI agents covers the provider side.
Who does it apply to?
Article 50 applies to providers and deployers of AI systems that interact with people or generate content, regardless of risk level. A low-risk chat agent that answers delivery questions is in scope. In a typical organisation, the people whose behaviour decides whether the duty is met are:
- Frontline staff who take over conversations from an agent in chat, email or voice.
- Team leads who write handoff scripts and greeting templates.
- Marketing and communications staff who publish agent-drafted content.
- Product owners who configure where and when the notice appears.
- Outsourced service staff working your channels on your behalf.
What must staff be trained to do?
Transparency sounds simple and fails in small ways. Training should cover six practical skills.
Know what the person was told. Before replying in a conversation an agent started, staff must check what notice was shown and when. If it was missing, they must say so plainly at the handoff.
Answer the direct question truthfully. "Was that a bot?" deserves a clear yes. Staff should never blur the answer to protect a satisfaction score.
Keep the line between drafted and sent. When an agent drafts at depth 2 and a person sends, the person has adopted the message. When an agent sends routine messages at depth 3 under its own name, the notice must travel with them. Staff should know which of their agents do which, capability by capability.
Recognise people who may not understand the notice. Older customers, people in distress and people using assistive technology may miss a small label. Staff should restate it in plain words when in doubt.
Handle requests for a person. The notice often prompts the request. Staff need to know how the agent hands over, how long a held conversation waits, and who picks it up.
Treat published content with care. Staff who publish agent-generated images, audio or text must know which items need marking, and should ask before publishing anything that could pass as real footage or a real person's words.
These skills overlap with the wider literacy covered in what AI literacy means for people who supervise AI agents, but they are distinct. A supervisor can be excellent at reviewing held actions and still mishandle a disclosure.
How do you meet Article 50 and show it?
| Requirement | Control | Evidence it leaves |
|---|---|---|
| People are told they are dealing with an AI system | Notice at the start of every agent-led interaction, in each channel | Notice text versions and where each appears |
| Notice survives handoff | Handoff script requires staff to confirm the AI status of earlier messages | Transcripts showing the confirmation |
| Truthful answers to direct questions | Training and quality criteria on disclosure | QA scores on sampled conversations |
| Accessible notice | Plain-language notice, restated in voice and assistive channels | Channel-by-channel notice review |
| Marked synthetic content | Publishing checklist for generated media and text | Checklist completions; marked assets |
| Competent staff | Disclosure module in induction for frontline and publishing roles | Training records per person |
| Outsourced channels | Contract terms and sampling of partner transcripts | Partner attestations; sample results |
Write the notice once, carefully. Our explainer on what a transparency notice is sets out what it should contain. The rest of the table is about people, which is why training belongs next to design.
What will a supervisor or auditor ask?
- In which channels do people meet an agent first, and what are they told?
- Show us a conversation where an adviser took over. Was the AI status clear?
- What do staff say when asked if they are a bot?
- How do outbound voice agents identify themselves?
- Who checks generated content before it is published?
- How do you sample outsourced channels for disclosure?
What does Article 50 not cover?
It does not require explanation of how the agent reached a decision; that is a separate matter, discussed in what AI transparency is. It does not cover human oversight, which for high-risk systems sits in Article 14 and is a different skill; see Article 14 training for people who oversee agents in fraud alerts. It does not replace data protection notices under GDPR, which still apply when an agent processes personal data. And it does not set out the content of staff training; it sets an outcome you must be able to show. The guide to AI literacy and certification for people who work with agents places this skill among the others.
This is an explanation of the rules, not legal advice; confirm how they apply to you with counsel.
Where the Academy fits
Pryme Intelligence Academy covers disclosure as part of running an agent, not as a separate course. Foundations explains ownership, data and trust for anyone who meets agents. The Operator track, for people who run a live workspace every day, teaches human review and what to do when something goes wrong, with module tests set in the learner's own industry. Lessons and labs are free, and labs run in a free Build workspace.
Questions people ask
Do we need to disclose if a person reviews every message before it is sent?
If a person reviews and sends the message, they have adopted it, and the customer is dealing with that person. Where the agent itself converses with the customer, the customer must be told.
Is a small label in the chat window enough?
It can be, if a reasonably well-informed person would see and understand it. Staff should still restate it when a customer seems unsure or is using voice or assistive tools.
Does Article 50 apply to internal agents used only by staff?
Staff using an internal agent usually know it is an AI system, so the interaction duty is often met by context. Content marking duties can still apply if the output is published.