PrymeIntelligence

Training teams for AI agents·Playbook

How to run an AI agent incident drill for the contact centre

Run an AI agent incident drill for the contact centre: wrong fee details sent at scale, pause to draft-only, surge staffing, escalate, record and report.

A contact centre drill should rehearse scale. Use a scenario in which the reply agent has given thousands of customers wrong information about a fee change, then practise pausing it to draft-only, rebuilding the rota for the held queue, escalating to the product owner and compliance, recording decisions, and reporting. Allow about two hours, because staffing decisions take time.

What makes a contact centre drill different?

Most incident drills test whether a few people do the right things in the right order. A contact centre drill also tests capacity. When the reply agent drops from depth 3 to depth 2, every reply it used to send alone now waits for an adviser. A queue that was manageable becomes a flood within the hour.

The contact centre manager owns the drill. A senior team leader or the workforce planning lead facilitates. Team leaders, quality analysts, workforce planners, someone from the product team and someone from compliance take part. For the general idea, see what an incident drill is.

The scenario

Friday, 09:30. A product change moved a late payment fee from one amount to another, effective at the start of the month. The agent's source was updated two days early. Since Wednesday, the agent has answered around 4,000 chat and email queries at depth 3, telling customers the new fee already applied. Some customers paid early to avoid it; others complained. Social media posts start appearing.

The drill step by step

  1. Person (facilitator): reads the scenario and starts the clock.
  2. Person (quality analyst): spots the error during sampling and alerts the team leader.
  3. Person (contact centre manager): pauses the agent to draft-only. The centre's own limits are tightened so every reply is held. The agent drops to depth 2.
  4. Person (workforce planner): rebuilds the rota. How many advisers are needed to clear held replies within the service target? Who comes off phones, who extends shifts, who is called in?
  5. Person (contact centre manager): escalates to the product owner and compliance. Product confirms the correct effective date.
  6. Person (team leader): corrects the agent's source with the product owner and checks a sample of new drafts before any depth is restored.
  7. Person (quality analyst): lists every affected reply from the record.
  8. Person (contact centre manager): decides on remedy and customer contact, with product and compliance: a correction message to affected customers, refunds of any fee charged early, and a script for advisers.
  9. Person (compliance): decides on reporting to leadership and, if customer treatment rules require it, the regulator.
  10. Person (facilitator): runs the debrief.
Step The agent does Depth Stays with a person
Alert Has answered at scale 3 Spotting in sampling
Pause Holds every reply Dropped to 2 Tightening limits
Rota Drafts replies into the held queue 2 Staffing decisions
Escalate Nothing new 2 Product and compliance
Fix source Drafts from corrected source 2 Checking new drafts
Scope and remedy Drafts correction messages 2 Approval; refunds
Report Nothing 2 Reporting decisions

What does the drill reveal about the centre's readiness?

The rota step usually exposes the real gap. Many centres find they have too few trained reviewers to clear a full held queue. Advisers who have never approved agent drafts cannot simply be switched over. The debrief should produce a minimum number of trained reviewers per shift, as described in who in a contact centre needs AI agent training first.

It also shows the value of sampling. A quality analyst caught the problem. Without regular sampling of depth 3 replies, the first warning would have been social media.

Finally, the drill should test the adviser script. Customers who call about the fee will hear something from the first adviser they reach, trained or not. A short, approved script gets the same answer to everyone within the first hour.

Which controls are tested, and what evidence do they leave?

The pause. Evidence: the limit change in the record, name and time.

Surge rota. Evidence: the rota change and the trained reviewers assigned.

Source correction. Evidence: the change to the agent's source, who made it, and the sample checked before depth was restored.

Customer remedy. Evidence: correction messages approved by named advisers, and each refund released by a person.

What should the contact centre manager measure?

  • Minutes from the quality alert to the pause.
  • Hours to clear the held queue within the service target.
  • Trained reviewers available per shift against what the surge required.
  • Time until all affected customers received a correction.

Where do contact centre drills go wrong?

Ignoring the queue. Teams focus on the error and forget that every new contact is now held. Customers wait, and complaints pile up on top of the original problem.

Switching in untrained advisers. It clears the queue faster and creates a second incident.

Restoring depth too soon. Check a sample of new drafts first.

Compare drills in teams that face different stakes: the risk team drill focuses on judgement, and the lending team drill on decisions with financial effect. Managers can build drills into planning with the AI agent training plan template for contact centre managers. The full method is in training teams for AI agents.

Where the Academy fits

Pryme Intelligence Academy's Operator track includes modules on running the rhythm and on when something goes wrong, which cover the two halves of a contact centre drill: keeping the queue moving and handling the error. For teams, an admin view shows which advisers hold current certificates, which is what the surge rota needs to know.

Questions people ask

Should the agent send correction messages itself?

It can draft them at depth 2 for advisers to approve. Once a correct template is approved, sending it in bulk may run at depth 3 if the agent is certified for it.

How do we practise the surge without disrupting real customers?

Run the rota step on paper, using real shift data. You do not need to move people to test whether you have enough trained reviewers.

How often should a contact centre drill?

Twice a year at least, and before any seasonal peak, so the rota plan is fresh.

Train the people who run your agents.

Pryme Intelligence Academy: free courses on the real product, audio lessons, labs in a free workspace, and certificates anyone can check on VerifyDoc.ai. Teams get seats, an admin view and a free readiness assessment.