What should contact centre managers never let an AI agent do?

The things that would mislead customers, trap them, or act on many of them at once without a decision. Contact centre managers should never let an agent hide that it is an AI system, block the route to a person, promise outside policy, settle complaints, handle vulnerable customers alone, or send bulk outbound messages that nobody approved.

What must customers always be told and always be able to do?

Know they are dealing with an AI system. Under the EU AI Act, people must be told when they interact with an AI system. Outside the EU, hiding it damages trust the moment a customer finds out. An agent should never present itself as a named human adviser.

Reach a person. The route to a human must be easy to find and work every time. An agent should never argue a customer out of transferring, loop them through repeated questions, or tell them nobody is available when someone is. A customer who cannot reach a person when it matters becomes a complaint and, often, a regulator’s case study.

What must an agent never offer or decide?

Promise outside policy. Compensation, fee waivers, delivery dates or outcomes the business cannot guarantee. These sit in the business’s own limits, one of five bounds every action meets alongside role scope, certificate, sources and the platform guardrails. Outside the limit, the agent refuses or holds.

Settle a complaint. The agent can acknowledge, gather facts and draft. Deciding whether the complaint is upheld, and what remedy to offer, needs a person who can weigh fairness and be accountable.

Handle vulnerable customers alone. Bereavement, serious illness, financial difficulty, signs of abuse. The agent should flag, respond with approved wording, and hand over. A person decides what happens next.

What must an agent never do at scale?

Send unapproved bulk outbound messages. A single wrong reply affects one customer. A wrong outbound message to every customer with an overdue bill, or every policyholder in a region, affects thousands at once and cannot be recalled. Bulk sends should be held for a named manager, with the message, the list and the trigger checked.

Change scripts, disclosures or the never-promise list. Those are the rules agents work within. Changes go through governance, never through the agent.

What does a crossed line look like?

Take a contact centre manager at an insurer. On Monday, a weather warning is issued. An agent, set up to send claims guidance to affected customers, prepares a message to all home insurance customers in three regions, telling them to “claim online for any storm damage; claims will be settled within 48 hours”. The agent held the send because the list exceeded its bulk limit.

The manager reads it. The settlement promise is not policy and cannot be guaranteed during a surge. The manager edits the message to set realistic expectations, narrows the list to customers in the warning area, and approves. Had the bulk limit been set higher, the promise would have gone to tens of thousands of people.

The agent was following its brief, which included an old campaign template. The manager asks for the template to be removed from its sources.

How do managers keep these lines in place?

Write each into the centre’s own limits so the agent refuses or holds it. Set a bulk-send limit low and keep it there. Test disclosure and transfer routes every week by calling in as a customer. Sample the record for promises and complaint handling. Keep track of certificate dates: certificates lapse after 90 days, and a run that clears no depth drops the agent to depth 1 the same day, which changes staffing as well as risk.

A competency matrix for the contact centre helps confirm reviewers can spot these lines. Related lists: what risk managers should never let an agent do and what operations managers should never let an agent do. Market detail differs: compare contact centre training in Nigerian insurance with Indian telecoms. The overview is in AI agent skills by role.