Show your collections staff exactly what the AI agent may now say, offer and agree, and what it hands to them. Every 90 days the agent is recertified on your own accounts, and its rights to send reminders, propose payment plans or flag vulnerability can change. Staff need the delta, practice on realistic accounts and a clear route for anything held.
What changes for collections staff at each cycle?
Two kinds of agent share this desk, so be precise. The people are collections agents. The software is a governed AI agent working alongside them.
The AI agent usually holds four capabilities. It reads the account, payment history and contact log at depth 1. It drafts letters, messages and payment plan proposals at depth 2. It sends routine reminders and agrees standard payment plans at depth 3, inside the lender's own limits on contact frequency, plan length and minimum instalment. Settlements and write-offs are depth 4 actions, hard to undo, and held for a named person.
Recertification re-runs the agent's cases on recent accounts. The escalation gate matters most in collections, because the cases that test it are customers who mention illness, bereavement, job loss or confusion. A capability that now escalates more readily changes the human team's workload, and staff need to know why those cases are arriving. In the UK the FCA's Consumer Duty expects good outcomes for retail customers, which makes the handling of these cases a conduct question as well as a service one.
If a run falls short, the agent keeps only the depth it cleared, and below depth 3 reminders stop going out and every plan proposal becomes a draft. The general idea is in what recertification is.
Who owns the cycle?
The collections team leader owns staff readiness and the rota. The builder who certifies the agent owns recertification. The head of collections owns the lender's own limits. A conduct or compliance reviewer should read the change note when it touches vulnerability.
Steps for each 90-day cycle
- Person (team leader): fix the date away from month-start payment peaks.
- Agent, depth 1: report the cycle. Reminders sent, plans agreed, drafts returned, escalations and refusals, with the deciding bound for each.
- Person (quality analyst): listen to or read a sample of agent-led contacts, with extra weight on vulnerability cases.
- Person (builder): recertify on recent accounts, including any the sample challenged.
- Person (team leader and builder): write the change note: contact rules, plan limits, new escalation phrases, and any change in the agent's depth.
- Person (conduct reviewer): sign the note where it touches vulnerability or contact rules.
- Agent, depth 2: draft practice accounts that show each change, checked by the quality analyst.
- Person (team leader): brief staff and sample the first fortnight.
| Step | The agent does | Depth | Stays with a person |
|---|---|---|---|
| Account review | Reads history and contact log | 1 | Judging context |
| Letters and messages | Drafts for review | 2 | Sending non-standard wording |
| Reminders | Sends within contact limits | 3 | Setting the limits |
| Standard payment plan | Agrees within bounds | 3 | Any plan outside bounds |
| Settlement or write-off | Prepares and holds | 4, held | Approving |
| Vulnerability signal | Escalates to a named person | escalation | All support decisions |
What to retrain, case by case
Vulnerability escalations. Give staff three or four examples of new escalation triggers. They should practise the first call after a handover, because the customer has already spoken to the agent and should not repeat their story.
Held settlements. Read the deciding bound first. "Held: above limit" is a value judgement. "Held: source declared" means the customer's statement of income is unverified, and staff decide whether to ask for evidence.
Refusals. If the agent refused to contact a customer because the contact limit was reached, the right response is to wait or escalate, not to call manually to get round it.
Controls and the evidence they leave
- Contact sample each cycle. Evidence: the sample, scores and corrective actions.
- Change note with conduct sign-off. Evidence: the signed note.
- Staff briefing. Evidence: named attendance and practice answers.
- Named approval on held settlements. Evidence: each approval in the append-only record.
- Practicum for new approvers. Evidence: the practicum log template for collections agents.
What should you measure?
Track escalations raised and the time to a human response, the number of plans agreed by the agent that later broke within one month, and complaints that mention contact frequency. A rise in broken plans after a cycle can mean the plan bounds widened and staff were not told what to watch.
Where does readiness slip?
It slips when staff see the AI agent as competition and stop reading its notes. Handovers then repeat questions, and customers in difficulty feel unheard.
It slips when settlement authority changes quietly. A change in who may approve a settlement belongs in the note, not in a rota email.
It slips when certification is treated as optional. The argument for requiring it before go-live is in the case for certifying collections agents before AI agents go live. Underwriting teams face similar handover issues; compare keeping underwriters ready for recertification.
Where the Academy fits
Pryme Intelligence Academy's Operator track covers human review, teaching the agent and when something goes wrong, which are the skills collections staff use at every handover. Lessons, labs and module tests are free, and tests are set in the learner's own industry and role. Certified Pryme Operator adds a written exam and a two-week practicum, and is valid 12 months. More in the guide to AI literacy and certification for people who work with agents.
Questions people ask
Should customers know they are talking to an AI agent?
Yes. Tell them plainly, and make it easy to reach a person, especially for anyone who says they are struggling.
Can staff change a plan the agent agreed?
Yes, through their own authority. The change and the reason go into the record alongside the agent's original verdict.
What if the agent escalates too often?
Raise it with the builder before the next cycle. Over-escalation is safer than the opposite, but it hides real cases in volume.