AI for Collections: Working Aged Debt Without Damaging the Relationship

Every day a balance sits unpaid is working capital you are lending to someone else for free. Yet most credit control teams spend most of their week not collecting. They spend it assembling: finding the invoice, digging out the purchase order, hunting for the proof of delivery, checking the contract terms, then reading months of email to work out whether this customer is disputing something or is simply slow. By then half the morning has gone and they have made four calls instead of twenty.

This is the part AI agents are good at. Not the conversation with the customer, which stays human, but everything that has to be true before it is worth having.

Triage and worklist prioritisation

An aged debtor report sorted by days overdue says little about where to spend effort. A ninety day balance from a customer who always pays at ninety days is not a problem. A fifteen day balance from one whose behaviour just changed is.

The agent reads the open receivables ledger alongside payment history, credit limits, order activity, dispute history and external credit signals. It scores each account on two axes that matter more than age: likelihood of payment without intervention, and value at risk if you do nothing. It then builds each collector a worklist ordered by expected recovery per hour, with a plain sentence explaining why each account is on it. Accounts that will self cure get left alone. Accounts drifting out of pattern surface early, while they are still collectable. The collector can reorder the list.

The single case view

Most collections tools promise a complete view of the account and few deliver, because the evidence is scattered. The invoice is in the ERP, the purchase order is in the customer's portal, the proof of delivery is a scanned image in a logistics system, the contract is on a shared drive.

Before an account reaches the worklist, the agent assembles the case: the invoice and its line detail, the matching purchase order, the signed proof of delivery or service acceptance, and the contract, from which it extracts payment terms, late payment interest, dispute notification windows and any agreed retention. It reads the email and ticket history and produces a short chronology.

The collector sees one page: what is owed, what backs it up, what has been said, and what is missing. If the proof of delivery cannot be found, the agent says so rather than leaving a gap to discover mid call.

Telling a dispute apart from a slow payer

These two look identical on an aged debt report and need opposite responses. Chasing a customer who raised a valid short delivery query and never got an answer does real damage. Treating a habitual late payer as a dispute case gives them another month.

The agent reads correspondence and transactional evidence together. A genuine dispute leaves traces: a query against a specific line, a credit request, a returns record, a partial payment matching the undisputed portion exactly. A slow payer leaves a different signature, with no substantive objection, payment eventually arriving in full, promises quietly missed.

The agent classifies the account, states the evidence, and splits the balance so the undisputed portion can still be chased. Anything it is unsure about goes to a collector with both readings laid out.

Drafting the chase in the right tone

The same balance warrants a different message for a strategic account in a temporary cash squeeze, a small business that forgot, and a serial late payer on their fourth broken promise.

The agent drafts the contact from the case file. It cites specific invoice numbers and amounts, states what has already been agreed, offers the supporting documents, and matches tone to the relationship and the stage of the process, following your approved templates and escalation ladder. Where a payment plan exists, it references the plan instead of demanding the full balance, which is the most common error in manual chasing.

Every outbound message is reviewed by a person before it goes anywhere, and not as a temporary measure while the technology matures. Collections contact carries legal and commercial consequences, and a human should own every word of it. The agent removes the twenty minutes of assembly, not the judgement.

Cash application and remittance matching

Unapplied cash is a quiet cause of bad collections. While a payment sits unallocated, the ledger says the customer owes money they have already paid, and someone chases them for it.

The agent matches incoming payments to open invoices, including the cases that break rules engines. Remittance advice arrives as an email body, an attachment, a portal download, or not at all. Customers pay round sums covering several invoices, deduct settlement discounts, net off credit notes, or pay short by exactly the amount of a disputed line. The agent reads the remittance, proposes an allocation and shows its working. Where a short payment matches a specific line, it raises a probable dispute rather than a payment error. Unmatched cash and low confidence allocations queue for a human with the best proposal pre-filled.

Promises to pay and the follow up

A promise to pay is the most valuable output of a collections call and the most commonly dropped. It gets written in a note field, or nowhere. The agent captures it from the call notes or email thread, records it as a structured commitment with a date and an amount, then watches for the payment. If the money arrives, the case closes quietly. If the date passes, the agent reopens the account at the top of the worklist with the broken promise stated plainly and drafts the follow up. Three broken promises is itself evidence. Payment plans work the same way, each instalment tracked and the first missed one flagged rather than the third.

Disputes, escalation and handover

A dispute sitting in the collections team is a dispute nobody can resolve. Only sales, operations or service can fix the underlying problem. When the agent classifies a genuine dispute, it routes the case file to the owning team with the claim, the evidence either way, and the value at stake, then chases that internal owner exactly as it would chase a customer. Internal delay is a large share of the ageing on disputed balances.

For escalation to legal or an external agency, the agent prepares the pack: transaction history, every contact made and when, the documents supporting the debt, the contract terms relied on, and the promises broken. The decision to escalate stays human. The agent removes the two days of file preparation that usually delays it.

The regulatory constraint you cannot design around

Consumer collections is a regulated activity in many markets, and business to business collections carries its own rules on late payment. Specifics vary by jurisdiction and change over time, so treat this as the shape of the problem rather than legal advice, and involve compliance before anything is built.

What is typically constrained is the frequency and timing of contact, the channels used, what a message may imply about consequences, and how customers in financial difficulty or identifying as vulnerable must be treated.

That changes the design in three ways. Frequency rules must be enforced by the system rather than remembered by a person, so the agent knows it cannot queue another contact today and says why. Any signal of vulnerability or financial difficulty must stop the standard chase path and route to a trained human, one of the few places an agent needs a hard stop rather than a confidence threshold. And there is no automated outbound contact to a consumer without human approval. Not a lower approval rate. None. Build the audit trail assuming a regulator will read it.

How you evaluate before you trust it

Nobody should take an agent's collections judgement on faith. Assess it the way you would assess a new collector.

Build a labelled set from your own history. Several hundred closed accounts where you know what happened, including the awkward ones: valid disputes, partial payments, insolvencies, the ones who paid the day after escalation. That is your fixed benchmark, and it should contain failures, not only clean cases.

Measure the classifications that carry consequence. For dispute against slow payer, what matters is how often the agent called something a slow payer when it was a genuine dispute, because that is the error that reaches the customer. Track it separately from overall accuracy, which will only flatter.

Shadow run before anything goes live. Let the agent build worklists, case files and drafts for several weeks while the team works as usual, then read the disagreements and see who was right. This is also where you discover which data sources are worse than anyone admitted.

Score drafts on tone and factual accuracy separately, because a message can be accurate and still wrong to send. Treat any factual error as a blocking defect. Then watch operational numbers rather than the demo: days sales outstanding, unapplied cash, promise to pay kept rate, disputes resolved inside target. Re-run the benchmark on a schedule, because behaviour drifts.

Where AI fails here

Bad ledger data breaks everything downstream. Duplicated invoices, unapplied credit notes, or the same customer existing three times under slightly different names will produce confident nonsense. This is the most common reason collections automation stalls, and the fix is data work.

Documents that were never filed. If proof of delivery lives in a driver's paper book, no agent can retrieve it. A case file that flags the gap helps, but it is not the same as finding it.

Context that exists only in someone's head. The collector who knows a renewal is being negotiated and now is not the moment to escalate holds information no system has.

Plausible and wrong. Models can produce a fluent draft containing an invoice number that does not exist or a balance that is out of date. Every fact in a draft should be bound to a value read from the ledger at draft time and validated before a human sees it.

Insolvency and hardship need judgement and legal input. The agent should recognise the signals and stop.

Where to start

Pick one segment and one hand-off. Case file assembly into a review queue is the usual first step, because a human still makes every decision, so risk is close to zero while the time saving lands immediately. Measure the baseline first and widen authority only on evidence.

If you want a grounded view of what an agent could take off your collections team, describe your current process to Adopt AI Partners and we will tell you what is automatable now and where we would begin.