Most customer service agents sit in the EU AI Act's transparency tier, not the high-risk one. They must tell people they are dealing with AI, must never manipulate or exploit vulnerable customers, and need staff trained to supervise them. A support agent becomes high-risk when it starts assessing creditworthiness or eligibility for listed services.
What does the Act require of a support agent?
Four duties apply to almost every customer service agent:
- Disclosure (Article 50). People must be told when they are interacting with an AI system, unless it is obvious from the context. The information must be given clearly, at the latest at the first interaction.
- Marking of generated content (Article 50). Certain AI-generated content must be marked as such. For a support agent this matters most where it produces synthetic audio, images or text published to inform the public.
- No prohibited practices. Manipulative or deceptive techniques that materially distort behaviour and cause significant harm are banned, as is exploiting vulnerability due to age, disability or economic situation.
- AI literacy (Article 4). Staff who supervise, configure or rely on the agent must have sufficient understanding of it.
If the agent is high-risk, the full deployer set under Article 26 applies on top: oversight, input data control, monitoring, logs and notices. For a primer on the Act itself, see what the EU AI Act is.
Where does the risk tier usually land?
Answering questions, tracking orders, resetting passwords, explaining fees, updating addresses: none of these is in Annex III. A support agent doing this work is limited risk, with transparency duties. The question whether a customer support chatbot is high-risk is usually answered no.
The tier changes when service blends into decisions. Annex III lists creditworthiness assessment and credit scoring of individuals, risk assessment and pricing in life and health insurance, and some decisions on eligibility for essential public services. A bank's support agent that starts telling customers whether they qualify for an overdraft, or an insurer's that quotes individual life premiums, may be taking part in a listed use.
That drift is the main risk. It rarely happens by decision. It happens when a team adds a connection to the lending system "so the agent can answer eligibility questions", and the agent begins doing in conversation what the credit process does formally.
What changes when the support agent can act?
A chatbot answers. An agent can also refund, waive a fee, change a direct debit, close an account or cancel an order. The Act's tiers do not turn on that difference, but the risk does, and so do the other rules you already follow.
Three duties take a specific shape for acting agents:
- Escalation. The Act does not require a human hand-over for limited-risk systems. Consumer law, financial conduct rules and plain good practice do. A customer must be able to reach a person, and the agent must hand over when it meets a complaint, a vulnerable customer, or a request outside its bounds.
- Manipulation in retention flows. An agent rewarded for keeping customers can learn to push. Pressure, false urgency and obstructing cancellation are where service crosses into the prohibited practices.
- Actions with consequences. Closing an account or cancelling a policy is hard to undo. Whatever the tier, those actions should be held for a person or bounded tightly.
The broader argument is in the EU AI Act and agents that act.
How do you meet the Act for a support agent?
| Requirement | Control | Evidence it leaves |
|---|---|---|
| Tell people it is AI | Disclosure in the first message in every channel, including voice | Disclosure text; channel captures; voice script |
| Mark generated content where required | Labels or technical marking on synthetic media | Marking method and samples |
| No manipulation or exploitation | Brief bars pressure tactics; vulnerability signals route to a person | Brief restrictions; escalation records |
| Stay out of high-risk uses | Connections to credit or underwriting systems reviewed before enabling; eligibility questions routed | Change review records; routing rules |
| Easy escalation | Hand-over on request, on complaint and on out-of-bound requests | Escalation counts and waiting times |
| Bound consequential actions | Refunds and waivers within limits; closures and cancellations held | Bound settings; held action records |
| AI literacy | Training for agents' supervisors and the team that receives escalations | Training records by role |
What will a supervisor or auditor ask?
- How does a customer know they are dealing with AI, in each channel?
- Which systems can the agent reach? Can it influence a credit or underwriting decision?
- Show us the escalation rules and how quickly customers reach a person.
- What can the agent do without approval, and what is the limit?
- How do you spot and protect vulnerable customers?
- Show us your retention script and the controls on it.
What does the Act not cover in customer service?
It does not set service standards, complaint handling rules or response times; those come from consumer and sector law. It does not require human review of routine service actions. It does not displace GDPR, which governs what personal data the agent may see and keep. And it is not the only frame: firms that also follow the voluntary US framework will find a comparison in EU AI Act vs NIST AI RMF for AI agents. For a deployer-wide view, use our EU AI Act compliance checklist for AI agents and the guide to AI agent regulation, mapped to controls.
This is an explanation of the rules, not legal advice; confirm how they apply to you with counsel.
Where Pryme fits
In Pryme, a support agent's capabilities each carry their own depth. Answering is depth 1, drafting a refund for a person to send is depth 2, issuing refunds within the customer's limits is depth 3. Actions above the agent's earned depth are held for a named person, and anything outside its role scope is refused, so an eligibility question cannot quietly become a credit decision.
Questions people ask
Does a voice agent have to say it is AI?
Yes, unless it is obvious from context. In a phone call it rarely is, so the disclosure should come at the start of the call, before the customer shares anything.
Can a support agent tell a customer whether they qualify for credit?
It can explain published criteria. Assessing an individual's creditworthiness is a listed high-risk use, so an agent that does it brings the high-risk duties with it. Route individual eligibility to the credit process.
Is a human hand-over legally required?
Not by the AI Act for limited-risk agents. Consumer and conduct rules, and the practical risk of complaints, make it necessary in most regulated businesses.