A borrower misses a payment on day one. Your servicing team has to decide: send a text, call, hold, or refer to collections. In a manual operation that decision depends on who is working that day. In a rules-engine operation it depends on which segment the account falls into. Neither gives you a record that shows a regulator, or a borrower who complains, exactly why that account was treated that way and not another.
AI-driven dunning is not about removing people from collections. It is about making every contact decision deliberate, consistent and auditable, from the first missed-payment nudge through to debt sale or write-off. The mechanism that makes that possible is an agent that reads account state, applies a defined procedure and writes a record of what it did and why, every time.
What an AI collections agent actually consists of
The phrase covers a range of implementations. In a lending context, the components that matter are these.
- Account-state reader. The agent reads the current ledger position: days past due, arrears balance, prior contact history, payment plan status and any vulnerability flags recorded on the account. It does not guess; it reads a structured event log.
- Contact-strategy selector. Given account state, the agent selects the next action from a defined matrix: channel (SMS, email, in-app, call), message type, and whether a human agent should be offered immediately. The matrix is configured by the lender and constrained by jurisdiction rules, including the contact-frequency limits that CONC 7 imposes and the equivalent requirements under each market's consumer credit framework.
- Forbearance assessor. Where an account shows signs of financial difficulty, the agent identifies which forbearance options are available under the lender's policy, presents them consistently and records what was offered. Consumer Duty requires lenders to deliver good outcomes for customers in financial difficulty; consistent offer-and-record is the mechanism that demonstrates that.
- Audit writer. Every decision, every contact, every borrower response is written back to the ledger as a timestamped, immutable event. That record is what a complaint handler or an FCA reviewer reads. If it does not exist at the time the decision is made, it cannot be reconstructed later.
- Escalation router. The agent knows its own limits. Accounts that cross a threshold, a borrower indicating vulnerability mid-conversation, a disputed balance, a legal notice are handed to a human with the full event context already loaded.
How a collections procedure runs in practice
Take a borrower who was due to pay on the 14th. It is now the 16th, no payment has arrived, and there is no prior arrears history on the account. The agent runs the following sequence.
- Read account state. The agent pulls the current ledger position: two days past due, no active payment plan, no vulnerability marker, no prior collections contact in this cycle. It confirms the contact channel the borrower has consented to.
- Select first contact. The contact matrix maps this state to an SMS reminder, not a call. The message text is drawn from an approved library. The agent records the outbound contact as a ledger event, including the template version used, so the exact wording can be retrieved if the borrower later disputes what they were told.
- Wait and re-read. The agent is scheduled to re-read the account after a defined interval. If payment arrives, it closes the dunning sequence and records resolution. If not, it advances to the next contact step, which may be email, then a call queue referral, each governed by the same matrix and the same record requirement. CONC 7.9 limits how often and in what manner a lender may contact a borrower in arrears; the matrix encodes those limits so they cannot be overridden by individual agent discretion.
- Assess for forbearance. By day ten, if the account remains unpaid, the agent checks whether the borrower has engaged with any contact. No engagement triggers a forbearance-offer step: the agent presents available options (a payment arrangement, a short deferral) through the agreed channel and records what was offered and whether it was accepted.
- Escalate or refer. If the borrower accepts a plan, the agent sets up the arrangement on the ledger and schedules a monitoring event. If the account remains unresolved beyond the lender's defined threshold, it is referred to the collections queue with full event history attached, so the human collector does not start from scratch.
The outcome is a ledger that shows, for every account, exactly which contact was made, when, on what basis and with what result. That is the record a complaint upheld by the Financial Ombudsman Service requires you to produce, and it is the record that agent operating procedures are designed to generate automatically.
Why a contact matrix is not the same as a prompt
Some teams build collections automation by feeding account data into a large language model and asking it to decide what to do next. The problem is not that the output is wrong. The problem is that it is not reproducible. Two identical accounts processed a day apart may receive different treatment because the model's response shifted. You cannot show a regulator, or a borrower, that the decision was made on a defined basis, because there was no defined basis: there was a probability distribution sampled at a point in time. CONC, Consumer Duty and equivalent frameworks in other jurisdictions require consistent, fair treatment. Consistency requires that the decision procedure is fixed, not inferred.
A properly structured collections agent uses a language model, where it uses one at all, for drafting communications within an approved template set or for summarising borrower responses, not for making contact-strategy decisions. The strategy decisions come from the matrix. The matrix is the thing the lender owns, configures, tests and can explain. That separation is what makes the system auditable, and auditability is what makes it defensible when a complaint arrives. For more on how the ledger handles the underlying arrears position, see past-due credit: how a ledger should handle it.
AI-driven dunning earns its place in a collections operation not because it is faster or cheaper, though it can be both, but because it makes every contact decision a ledger event: defined, recorded and retrievable at the moment a borrower or a regulator asks why their account was handled the way it was.