PrymeIntelligence

Training teams for AI agents·Playbook

How to run an AI agent incident drill for the collections team

Run an AI agent incident drill for the collections team: reminders sent to customers on payment plans, pause to draft-only, escalate, record and report.

The best collections scenario is one where the agent keeps chasing people it should have left alone: customers on payment plans or flagged as vulnerable. Run it as a 90-minute tabletop. Collectors and team leaders practise pausing the agent to draft-only, escalating to the head of collections and compliance, recording decisions, and putting affected customers right.

Why does collections need its own drill?

Collections is where a lender, utility or telecoms operator meets customers in difficulty. Agents typically send routine arrears reminders at depth 3 within bounds, draft payment arrangement offers at depth 2, and hold any message that mentions hardship, illness or bereavement for a person. When the agent's bounds fail, the harm lands on people who are already struggling, and complaints and regulatory scrutiny follow quickly.

The head of collections owns the drill. A team leader or the quality lead facilitates. Collectors, a complaints handler, someone from compliance and someone who knows the customer data feeds take part. The general idea is set out in what an incident drill is.

The scenario

Tuesday, 11:00. A customer rings in distress. She agreed a payment plan three weeks ago, has kept to it, and has just received her third reminder threatening further action. A collector checks: the flag that marks accounts on a payment plan stopped flowing from the core system to the agent's sources after a weekend change. Since then the agent has sent reminders to about 900 customers on plans, including around 70 flagged as needing extra care.

The drill step by step

  1. Person (facilitator): reads the scenario and starts the clock. No scripts.
  2. Person (collector): calms the customer, notes the details and escalates to the team leader while still on the call. The test: do they recognise that one complaint may reveal hundreds of cases?
  3. Person (team leader): pauses the agent to draft-only. They tighten the business's own limits so every reminder is held. The agent drops from depth 3 to depth 2. The verdicts now name the business's own limits as the deciding bound.
  4. Person (team leader): escalates to the head of collections, compliance and the data owner. The data owner confirms when the flag feed stopped.
  5. Person (team leader with the data owner): lists every reminder sent to a flagged account since the change, using the append-only record. Agents can read and sort the list at depth 1; a person checks it.
  6. Person (head of collections): decides on remedy. Call every customer flagged as needing extra care, by a trained collector, within a set time. Write to everyone else on a plan with an apology. Reverse any fees or interest added by the reminders.
  7. Person (team leader): records each decision with owner and time.
  8. Person (compliance): decides on reporting to the board and, where customer treatment rules require it, the regulator.
  9. Person (facilitator): runs the debrief. The pause is lifted only once the flag feed is fixed and the agent's next reminders have been checked.
Step The agent does Depth Stays with a person
Alert Has sent reminders 3 Recognising a pattern on a call
Pause Holds every reminder Dropped to 2 Tightening limits
Escalate Nothing new 2 Data owner, compliance
Scope Sorts affected accounts 1 Checking the list
Remedy Drafts apology letters 2 Calls to vulnerable customers; fee reversals
Record Files verdicts 2 People's decisions
Report Nothing 2 Board and regulator

What does this drill teach about provenance?

The payment plan flag came from the core system and was sealed. Sealed facts are the strongest kind, so the agent trusted them. The problem was absence: a missing flag looks the same as a customer with no plan.

The debrief should ask how the team would have caught the gap earlier. A common answer is a simple check: if the number of accounts on plans drops sharply overnight, the agent holds all reminders until a person confirms. That is a change to the agent's bounds, not a retraining of a model.

Which controls does the drill test, and what evidence remains?

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

Scope from the record. Every reminder is a filed verdict, so the 900 can be listed exactly. Evidence: the list and the person who checked it.

Vulnerable customers first. Evidence: call notes showing a trained collector reached each flagged customer.

Fee reversal. Evidence: each reversal released by a named person.

What should the head of collections measure?

  • Minutes from the customer's call to the pause.
  • Hours until every customer flagged as needing extra care was called.
  • Completeness of the affected list against a manual sample.
  • Whether collectors could explain to customers what happened, in plain words.

Where do collections drills go wrong?

The collector handles the call and moves on. The single most important moment is the first collector deciding to escalate. Drill that moment until it is a habit.

Remedy by letter only. A letter to a customer in crisis is not enough. The people flagged as needing extra care should hear from a person.

Lifting the pause when the feed is fixed. Check the next batch of reminders before restoring depth 3.

Teams with overlapping risks can compare the support team drill and the compliance team drill. The AI agent incident drill script template gives a ready structure, and how to train a team to handle an AI agent incident covers the training around it. The overall approach is in training teams for AI agents.

Where the Academy fits

Pryme Intelligence Academy's Operator track includes a module on when something goes wrong and one on reading the workspace, which is how a collector learns to spot that a held queue has gone quiet when it should not have. Module tests are set in the learner's own industry and role.

Questions people ask

Can the collections agent ever contact a vulnerable customer?

It can draft a message for a trained collector to review and send. Messages to customers flagged as needing extra care should not go out at depth 3.

Who decides whether to reverse fees?

The head of collections, with compliance. Decide the principle in the drill so the real event moves faster.

How often should collections teams drill?

At least twice a year, and after any change to data feeds, the arrears process or the agent's depth.

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.