How can contact centres prove staff are trained to oversee AI agents?

Contact centres prove it by tying training to the queue. Each person who approves agent replies, refunds or account changes holds a certificate checked at source, the right to work a queue depends on that certificate being current, and the filed record shows who approved each held action during their shift. High turnover makes this harder and more important.

What makes contact centres hard to evidence?

Three features of contact centre work strain ordinary training records.

Turnover and seasonality. Teams grow for peaks and shrink after. Training records built for a stable workforce lag behind who is actually on the phones.

Outsourcing. Many centres serve several clients, each with its own agent, policies and limits. A reviewer certified on one client's refund agent is not automatically competent on another's.

Shift work. Approvals happen around the clock. If the only certified approver on the night shift is off sick, held items either wait or are approved by someone who should not.

Clients, auditors and regulators of the client's sector all want the same assurance: the person who approved this action was trained for it, at the time. Contact centres need evidence that answers per action, not per year.

What evidence should a contact centre keep?

• A queue-to-competence map. For each client and queue, which agent capabilities run there, at what depth, and which certificate an approver needs.
• A live certificate register. Every approver, their certificate number, issuer, expiry and the clients they are cleared for.
• Queue rights driven by the register. When a certificate lapses, the person's access to approval queues is withdrawn, not just flagged.
• Shift rosters linked to rights. Each shift has enough certified approvers, and the roster shows it.
• Approval records. Every held action, who approved it and when, kept in an append-only record.
• Client-specific practicum results. Supervised decisions on each client's work before solo approval.

The piece on what contact centres get wrong about AI agent training covers the commonest gaps, especially treating one client's training as covering all.

What should the training actually cover?

Contact centre approvers see the same patterns many times a day. That is the risk: approval becomes reflex. Training and testing should focus on:

• Spotting the exception in a stream of routine drafts, using seeded errors.
• Identity. Callers state who they are; that is declared, not verified. Account changes need verification through the client's approved method.
• Refusals under pressure. When a customer insists, an approver must not process a refused action by hand outside the workspace.
• Disclosure. Under the EU AI Act's Article 50, people must be told when they interact with an AI system. Approvers should know how the client's agent discloses this and notice when a draft does not.
• Vulnerability. Recognising signs that a customer needs a person, not a faster reply.

An example from an outsourced centre

An outsourced centre serves an energy supplier and a retail bank. During a winter peak it brings in 30 temporary staff. All complete the general module and the energy client's refund agent practicum; none complete the bank's.

A team lead short of hands moves five temporary staff onto the bank's queue. Because queue rights are driven by the register, the system blocks them from approving held items on that queue: they can handle unheld contacts but not approve anything the bank's agent held. The bank's audit later samples approvals during the peak and finds every one made by a person cleared for that client. The evidence was produced by the control itself, not assembled afterwards.

What will a client or auditor ask?

• Show me everyone who approved actions on our queue last month, and their certificate status at the time.
• How do you stop someone uncertified on our work from approving?
• How quickly are rights withdrawn when a certificate lapses?
• What did your staff practise on before approving our customers' actions?

Clients in regulated markets may bring their own regulator's expectations; see MAS FEAT principles and training staff and SAMA expectations in Saudi Arabia. Other high-volume settings face similar issues, as hospitals and universities show.

Pryme Intelligence Academy offers a free readiness assessment for teams of up to 25 people, which is a reasonable way to find the gaps on a single client queue before building the rest. The guide to AI literacy and certification explains the broader approach.