Quality analysts and team leaders first, then the advisors who approve held actions, then the whole floor. A contact centre should train the people who will judge the agent's handling before it handles anyone, run it in shadow on real contacts, and widen what it may do only when quality scores, reviewer decisions and the agent's certification agree.
Why start with quality and team leaders?
In a contact centre, the people who define "good" are the quality analysts and team leaders. They score calls, coach advisors and handle escalations. When an agent arrives, they are the ones who must decide whether its replies, refunds and handoffs meet the standard. If they are trained last, they will mark the agent against a scorecard written for humans and miss what matters: whether the agent acted within its bounds and whether its sources were sound.
So the first cohort is small and senior. Once they can read the agent's decisions and calibrate one another, they become the coaches for everyone else.
What order should the floor follow?
- Quality analysts and team leaders. The full reviewer path, including a practicum on the agent's shadow drafts. They then build the centre's calibration set from real contacts.
- Advisors on approval queues. The people who will approve held refunds, goodwill credits, account changes and complaint replies. Certified before the agent goes live at depth 2 or above on their queue.
- All advisors. The shared vocabulary: what the agent does, what a held or refused action means, how a handoff arrives. A few hours, delivered in short blocks that workforce management can schedule.
- Workforce planners and process designers. How the agent changes volumes, handling time and skills mix, and who builds and certifies the agents.
What do advisors need to learn about handoffs?
The handoff is where most contact centre agent deployments succeed or fail. When the agent passes a customer to a person, the advisor receives its summary and the record of what it tried. A trained advisor reads that record first: which facts were declared by the customer and unverified, which came from the account system, and why the agent stopped. An untrained advisor asks the customer to start again, and the customer leaves angry.
Advisors also need to know what the agent must tell people. In the EU, Article 50 of the AI Act requires that people are told when they are interacting with an AI system. Advisors picking up a contact should never contradict that or pretend the earlier conversation was with a colleague.
What may the agent do, and what stays with a person?
| The agent may | Stays with a person |
|---|---|
| Answer account and order questions from sealed records (depth 1 to 2) | Complaints that allege harm or discrimination |
| Draft replies for an advisor to send (depth 2) | Any contact where the customer shows signs of vulnerability |
| Issue goodwill credits within the centre's limits, once certified (depth 3) | Credits above the limit, and repeat credits to the same account |
| Change contact details after verified authentication (depth 3) | Changes requested on a declared identity alone |
What should be proven before widening?
Before a capability moves from drafting to acting, three records should agree. Quality scores on the agent's shadow drafts should be at least as good as the team's. Reviewers' practicum decisions should show no missed red line, such as an account change on an unverified caller. And the agent itself should be certified at that depth on the centre's own contacts, with depth earned from the evidence rather than set by a planner. If any of the three lags, the capability stays at depth 2.
An example
Take a hypothetical retail contact centre with 300 advisors. Twelve quality analysts and 25 team leaders take the reviewer path while the agent drafts replies to order queries for three weeks. Forty advisors on the refunds queue follow and complete their practicum on held refund drafts. The whole floor takes the vocabulary course in 30-minute blocks over a month. The agent then issues refunds under a small value at depth 3; everything above stays held for a certified advisor.
Methods for that reviewer practice are set out in how to train the contact centre to review AI agent decisions, and the training plan template for contact centre managers turns the order into weeks. For the budget, see what AI agent training costs contact centres.
Questions people ask
Can the agent take contacts before advisors are trained?
At depth 1 or 2, with every action passing through a person, yes. The limits are in can contact centres use AI agents before staff are trained.
Should quality scorecards change?
Yes. Add checks for the agent's sources, its bounds and the quality of its handoff, not just tone and resolution. Pryme Intelligence Academy's Operator modules on reading the workspace cover what to look for.
What is the commonest mistake?
Training the floor and forgetting quality. More in what contact centres get wrong about AI agent training and the guide to training teams for AI agents.