AI for Collections in Utilities: Domestic Arrears, Affordability and Arrangements That Hold

A domestic arrears book behaves nothing like a commercial receivables ledger. The customer cannot be cut off without a regulated process and often has not refused to pay at all. Some are in genuine financial difficulty. Some dispute a balance built from estimated reads. Some moved out two years ago and the final bill went to an address they left.

They all sit on the same aged debt report in descending balance order, and one dunning path applied evenly across them harms the households that needed help, misses the ones who could have paid, and chases balances that should never have been pursued. The work that matters is separating those populations before anything reaches a customer.

Ability to pay and willingness to pay drive everything downstream

Two accounts with identical balances and ages can need opposite treatment. A willingness case is contactable, has the means and has not paid. An ability case has income below essential expenditure and will break any plan that ignores it. The distinction sets the channel, the tone, whether an arrangement is offered before a demand, and whether the account belongs in a support queue rather than a collections one.

An agent can assemble the evidence that separates them: the shape of payment history rather than its age, prior arrangements and how they ended, whether payments arrive as a benefit deduction, and any contact where the customer described their circumstances. It presents that reading with the evidence attached. It does not act on it.

Affordability assessment and the income and expenditure statement

An arrangement is only as good as the affordability work behind it: income including benefits, essential household expenditure, other priority debts such as rent, council tax and the other utility, then non-priority credit, ending in a disposable figure shared across priority creditors rather than claimed by whoever asked first.

The mechanical parts suit an agent. It reads uploaded bank statements, award letters and wage slips, categorises transactions into your headings, and flags a figure far outside typical guidance. The arithmetic belongs to ordinary code, never the model, because a plan built on a wrong subtotal fails by month two. And the customer's own account of their circumstances stays the record: an agent may prompt a question, never replace a declared figure with an inference.

Setting an arrangement, and re-planning a broken one

The most common design error is a plan that repays the arrears but does not cover ongoing consumption, so the customer pays every week and the balance still climbs. Both must be affordable together, and a plan set against summer usage on a heating account is unaffordable by January unless it was built for that. When a plan breaks, the useful question is why, not how many instalments were missed. An agent can detect the break early, gather what changed since the plan was set, and propose a revision. Where the original was unaffordable, the answer is a lower plan, not a stage advance.

Prepayment meters and credit meters take different paths

On a prepayment meter, arrears are recovered as a deduction from each top-up, so recovery competes directly with supply. Set the rate too high and the customer stops topping up, and that failure is silent: it shows as an absence of vends, not a complaint. Treat falling top-up frequency as a welfare signal and a reason to review the deduction rate, not grounds to escalate. Credit meters take the longer route through billing, arrangement, formal notice and steps needing external authority. Moving a customer between meter types is itself a controlled action, which an agent can prepare and document but must not decide.

Vulnerability is a stop, not an input

A vulnerability signal or a priority services register flag must act as a gate that runs before any scoring, not as a feature inside the score. A feature is a weight, and a large enough balance will eventually outweigh it.

As a stop, the flag takes the account out of automated contact and every stage advance path and routes it to a trained human. The valuable job for an agent runs the other way: finding vulnerability nobody recorded. Signals sit in call notes, inbound correspondence, letters from advice agencies, evidence of medical dependency on supply and repeated self-disconnection. The agent proposes the flag and the evidence. It does not choose the treatment.

Estimated reads, disputed balances and back-billing limits

A meaningful share of any domestic arrears book is not arrears. A long run of estimated reads produces a balance that dissolves the moment an actual read lands, and a catch-up bill after a missed meter exchange looks like wilful non-payment when it is a billing failure. Where the supplier failed to bill accurately for a period, back-billing limits mean the older portion is not collectable at all.

Run that verification before the contact, not after the complaint. Is the balance built on actual reads or estimates. Is there an unbilled gap. Does any part fall outside the back-billing window. Was there a meter exchange, a crossed meter or a wrong serial. Expect the counter-intuitive result: the book gets smaller, because accounts failing the check leave recovery and go to billing for correction.

Final bills, closed accounts and gone away tracing

Closed account debt is its own discipline. The final bill is often estimated, the tenancy change date disputed, and liability between outgoing tenant, incoming tenant and landlord unclear from the records the supplier holds. Chasing the wrong person for someone else's consumption is both a complaint and a data protection incident. So before any tracing effort, an agent can assemble the liability case: the tenancy change record, the last actual read and its date, the incoming occupier's account and any evidence of occupancy. Tracing is then a matching exercise it can rank by likelihood, though confirming a traced record is the same person stays a human decision.

Recovery stages, protections and the audit trail

Treat every stage advance as a gated decision rather than a timer. Typical preconditions: contact attempted across the required channels, an arrangement genuinely offered, affordability attempted, the vulnerability gate clear, the balance verified against actual reads, the back-billing test passed. The agent evidences that those conditions hold; the advance itself should be a rule your compliance team can read, not a model output. Disconnection and enforced meter installation sit behind protections: seasonal restrictions, categories of household exempt from them, and external authority with its own evidence standard. Specifics vary by market, so involve compliance early. For any account you should be able to reconstruct which checks ran, on what data, under which rule version, and who approved.

Placement with and recall from external agencies

Placement with an external agency deserves a preflight list: balance verified, no open dispute, no vulnerability flag, no live arrangement, address confirmed, correct customer identified. Recall is where an agent adds most, because after placement nobody is watching. A customer contacting the supplier directly, a new register flag, a dispute raised, a back-billing finding or a meter type change should each trigger recall, and the supplier stays accountable for the agency's conduct throughout. That is a continuous reconciliation between the placed population and internal changes, monitoring that never gets done by people with a worklist of their own.

Hardship funds, grants and entitlement screening

Most customers who qualify for help never apply. Hardship funds, arrears write-off schemes with matched-payment conditions, efficiency grants and unclaimed benefit entitlement all have criteria that are rule-based and evidence-hungry, which is why applications stall. An agent can screen the arrears book against each scheme and produce a shortlist where the evidence is already gathered from the affordability work, so an adviser offers the scheme rather than waiting for the customer to discover it. Where a scheme depends on the customer maintaining payments for a matched credit, the same agent tracks both sides and flags a break while it is still recoverable. Award decisions stay with the scheme owner.

Where AI fails here

Bad metering and account data defeats it: crossed meters, wrong serials, unresolved tenancy change records and duplicate customers produce confident nonsense at scale.

Circumstances nobody recorded are invisible. If a health condition was described on a call that was never noted, no model recovers it.

Fluent and wrong is the dangerous failure. An arrangement letter citing a balance that has since changed reads perfectly well and is a complaint waiting to happen. Bind every figure in a customer-facing draft to a value read at draft time.

Vulnerability detection errs both ways. Missing a signal harms someone; flagging a household that is coping is intrusive. Only a proposal reviewed by a trained person is safe.

No automated contact should reach a customer in arrears without human approval. Not a reduced rate. None. The same holds for any step affecting supply.

Where to start

Begin with balance verification. It is documents and arithmetic rather than judgement, it has a checkable right answer, it touches no customer, and it removes uncollectable and disputed debt before anyone spends effort on it. Affordability statement assembly into an adviser review queue is a strong second, since the adviser still owns every figure and every word spoken. Placement preflight and recall monitoring follow, because both are reconciliations. Shadow run each against current practice and read the disagreements before granting authority.

Why a recovery-maximising model fails here

Set the objective as maximum recovery and an optimiser finds the accounts where pressure works, many of them households paying the energy bill instead of something else they cannot afford either. It learns to push past a broken arrangement rather than reduce it, to treat a vulnerability flag as a cost rather than a boundary, and to chase balances a back-billing check would have written off. Each breaches the rules this sector runs under, and each turns up later in a complaint file or an audit.

In domestic arrears the right outcome is frequently a smaller, collectable, sustainable arrangement rather than the full balance. So the objective is arrangements that hold and customers still supplied, with recovery as a consequence rather than the target, behind gates no score can override.

If your arrears book holds debt you suspect should never have been chased, or arrangements that break faster than you can reset them, describe the process to Adopt AI Partners and we will tell you plainly which parts an agent should touch and which it should not.