Italian telecoms operators that put AI agents into customer care must train support agents in AI literacy under Article 4 of the EU AI Act, tell callers when they are speaking to an AI system, and keep disputes, number portability and unrequested-service claims with people. The core skill is reviewing held actions and correcting drafts against contract records.
Where does an Italian telecoms support agent's time go?
Disputed charges dominate. Customers find a premium content subscription they say they never activated, a price change after a promotion ended, or a fee for a modem they believe they returned. Unrequested activations are a long-running sore point in the Italian market, so many calls start from suspicion rather than a question.
Switching is the second large block. Customers move between operators often, and number portability, fibre migrations and the return of rented equipment all generate contacts. Behind them sit the reminders, recontacts and conciliation cases that follow when a customer feels a switch was handled badly. Smaller but sensitive: SIM replacement requests, line transfers between family members, and requests from customers in difficulty who cannot pay.
What changes when AI agents handle part of customer care?
The AI agent reads, explains and prepares; the support agent decides where money, identity or a dispute is at stake.
- Charge explanations (depth 2). The agent pulls the contract, the activation record and the consent trail behind a disputed service, and drafts an explanation. If no consent record exists, it says so plainly in the draft.
- Service deactivation on request (depth 3). Removing a value-added service the customer asks to stop is routine and reversible, so it can run within bounds and be logged.
- Refunds for unrequested services (depth 4, held). The agent prepares the refund with the evidence; a named person releases it.
- Portability status (depth 1). The agent reads the migration status and tells the customer where it stands, without changing anything.
- Equipment return tracking (depth 3). The agent checks return records and sends the standard reminder or confirmation.
What changes for the person is attention. A support agent spends less time searching systems and more time judging whether the evidence the AI agent assembled really supports the answer it drafted.
What do support agents need to learn, and what stays theirs?
Start from the role map in our guide to AI agent skills by role. For Italian telecoms care, four areas matter most.
Provenance. Every fact the AI agent uses carries its source. A consent record from the activation system is sealed; a customer's statement that they never clicked anything is declared. Reviewers must read the difference, because a refund decision depends on it.
Verdicts and depth. Support agents should know why a refund over the limit arrives held, why a line transfer is refused for the agent, and that a withheld action is fixed at the platform floor rather than set by the operator.
Corrections that stick. Rewriting a poor draft fixes one reply. Recording why it was wrong fixes the agent. Our guide on teaching support agents to bank corrections shows how.
Limits. Read what support agents should never let an AI agent do before the first live shift.
What stays with the person: refunds above the limit, any decision on a disputed activation, line transfers and SIM replacements, payment arrangements for customers in difficulty, and conciliation cases. People also keep every call where the customer asks for a human. Agents do not remove the need for care staff; they change what care staff do, a point developed in will AI agents replace customer support agents.
Which rules in Italy shape oversight and literacy?
The EU AI Act applies directly. Article 4 requires providers and deployers to take measures so that staff dealing with AI systems have sufficient AI literacy, which makes training a legal expectation rather than a nice extra. Article 50 requires that people are told when they are interacting with an AI system, so a customer chatting with the AI agent must know it. For high-risk systems, Article 26 requires deployers to assign human oversight to people with the competence, training and authority to exercise it; a care agent is unlikely to be high-risk, but that standard is a sound benchmark.
GDPR governs the customer data the agent reads, with the Garante as the Italian supervisory authority. Article 22 limits decisions based solely on automated processing with significant effects, which matters if an agent were ever allowed to decide, alone, on disconnection or a disputed debt. The national communications authority sets rules on contract transparency and runs an online conciliation procedure for disputes between users and operators. Telecoms operators also fall within NIS2, which brings cybersecurity risk management and incident reporting into the picture when an agent connects to network and customer systems.
This is an explanation of the rules, not legal advice; confirm how they apply to you with counsel.
How should the first 90 days of training run?
Weeks 1 to 4. Give every support agent and supervisor a grounding in governed agents and in the Article 50 disclosure script. Run the AI agent in shadow on charge explanations, comparing its drafts with the replies staff actually sent.
Weeks 5 to 8. Let reviewers approve drafts on live contacts. Use the AI agent competency matrix for the support team to record who can review what, and run a weekly session on disputed activations, where the provenance distinction does most of the work.
Weeks 9 to 13. Certify the agent for service deactivation at depth 3 and keep refunds held. Put supervisors through operator certification, and allow only certified people to release held refunds. Keep evidence of the training itself: under Article 4, being able to show how literacy was built is part of the control.
Where the Academy fits
Pryme Intelligence Academy gives Italian care teams a free Foundations track, about four hours, that covers why agents are governed and how the decision chain works. Supervisors take the Operator track, with module tests set in telecoms support and a two-week practicum on real decisions. The admin view records progress and certificates, which supports an Article 4 file.
Questions people ask
Does the AI agent have to announce itself on every contact?
Article 50 requires that people are told they are interacting with an AI system, unless it is obvious from the context. A short, clear statement at the start of each chat or call is the safe practice, and support agents should know the wording.
Can the AI agent decide that an activation was genuine?
It can show the consent evidence and its provenance. The decision on a disputed activation stays with a person, because it turns on judgement about evidence and the customer's account of events.
Is AI literacy training a one-off?
No. The AI agent's certificate lapses after 90 days and its permissions can change at each renewal, so literacy needs short refreshers tied to those changes.