Agentic Accounts Receivable: How AI Agents Score Credit Risk and Automate Dunning Before Invoices Become Write-Offs
written by Cooter:Labs
published on August 9, 2026
Introduction
Most receivables don't go bad suddenly. An account slides from current to 30 days past due, then to 60, then to 90, and by the time it lands on a collections manager's desk as a write-off candidate, the pattern that predicted it — slower payment on the last three invoices, a dispute that never got resolved, a credit limit that was raised eighteen months ago and never revisited — was visible in the data the whole time. The reason it wasn't acted on earlier usually isn't that nobody could see it. It's that credit risk scoring in most ERPs runs at customer onboarding and maybe an annual review, and collections works off an aging report that's a snapshot of what already happened, not a signal for what's about to.
Collections is a lagging process built on top of a leading-indicator problem
An aging bucket tells you an invoice is 60 days past due. It doesn't tell you that this same customer's average days-to-pay has been drifting upward for two quarters, that a growing share of their invoices are getting short-paid against a deduction code that's never validated, or that their order volume just spiked right as their payment behavior started to slip — the classic pattern of a company over-ordering on credit before a default. None of that requires new data. It requires correlating payment history, open disputes, order patterns, and the credit limit on file continuously instead of at the moment someone happens to pull a report. An agent doing that correlation on every new invoice and every payment posting can flag the drift while there's still a decision to make, instead of after the account has already aged into a collections escalation.
Most ERPs set a credit limit once, at onboarding, based on a credit application, a bureau pull, or a trade reference check, and then leave it untouched unless someone manually flags the account for review. An agent with access to the AR ledger can recompute a risk score on every payment event instead — average days-to-pay over a trailing window, the trend direction of that average, the ratio of disputed to clean invoices, and how close the customer is running to their existing limit. None of this replaces the initial underwriting; it keeps the number current instead of stale. A customer whose payment behavior has quietly degraded for two quarters shows up as elevated risk before they blow through their limit, not after a manual review eighttteen months later happens to catch it.
The default dunning setup in most systems is one cadence for everyone: a reminder at 7 days past due, a second notice at 30, a call at 60, regardless of who the customer is or why they're late. An agent can route by segment and by cause instead — a large strategic account with an otherwise clean payment history that's three days late gets a different treatment than a small account with a six-month pattern of paying only after the second reminder. It can also distinguish 'hasn't paid because they haven't gotten to it' from 'hasn't paid because of an open dispute,' and hold the dunning sequence on disputed invoices instead of sending an increasingly aggressive reminder about money the customer doesn't believe they owe — a mistake that reliably escalates a billing question into a relationship problem.
A large share of late payment isn't nonpayment — it's a partial payment against a deduction: a shortage claim, a pricing discrepancy, a damaged-goods allowance, a promotional credit the customer thinks they're owed and the seller's system doesn't reflect yet. Each of these needs to be validated against source documents — the shipment record, the price list in effect at order time, the signed proof of delivery — before anyone knows whether the deduction is legitimate. An agent can pull the relevant source documents automatically, resolve the deductions that check out cleanly against a rule (a valid promo code, a shipment record that confirms the shortage) without a person touching them, and route only the ones that don't reconcile — the ones where the claim and the paper trail actually disagree — to a person for a judgment call.
When a collections call ends with a customer committing to pay by a specific date, that promise is usually logged as a note in a CRM or collections tool and then nobody systematically checks whether it was kept — the account just ages further and the next call starts from scratch. An agent can track the promise as a structured commitment with a due date, watch for the payment against it, and escalate specifically when a promise is broken rather than on the generic aging schedule. A customer who has broken two promise-to-pay commitments in a row is a materially different risk than one who's simply 45 days out with no broken commitments yet, and the escalation path should reflect that difference instead of treating both as the same aging-bucket problem.
Looking Ahead: Challenges and Innovations
Aggressive automated dunning can damage the accounts that matter most
A strategic account with real negotiating leverage and a long relationship history doesn't respond well to an automated reminder sequence built for a small, low-touch customer — and collections mishandling a large account can cost more in relationship damage than the late payment itself was ever going to cost in carrying cost. The workable design routes large or strategic accounts to a human relationship owner for anything past the first automated reminder, and reserves the fully automated sequence for the long tail of small accounts where the economics of a human touch don't make sense in the first place.
A validated deduction still needs someone to own the write-off decision
An agent that confirms a deduction claim matches the source documents has done real work — it's replaced a manual document pull with an automated one — but deciding to actually write off the disputed amount is a financial decision with accounting and customer-relationship consequences that shouldn't be fully automated even when the underlying facts are clear. The agent's role is narrowing the decision to a clean recommendation with the evidence attached, not making the write-off itself; someone with authority over bad debt reserves still signs off.
Credit holds carry real revenue risk if the underlying score is wrong
Putting an account on credit hold blocks their next shipment or service delivery, which means a false positive — flagging a customer as high risk when they're not — has an immediate, visible business cost, not just an internal process cost. That asymmetry argues for the agent recommending a hold with its reasoning surfaced, rather than triggering one automatically, at least until the scoring model has enough track record on a given customer segment to earn that level of autonomy. A wrongly held shipment to a good customer is a worse outcome than a slightly late escalation on a bad one.
The metaverse
The natural next step is connecting AR risk monitoring to the AP-side agentic workflows already covering vendor payments and three-way matching, since many companies are simultaneously a customer and a supplier to the same counterparties. A unified view of a business partner's payment behavior on both sides of the ledger — how reliably they pay you, alongside how reliably you're paying them — is a more complete risk picture than either side tracked in isolation, and it's a natural extension for companies that have already built the continuous-monitoring infrastructure for one side of the transaction.
Conclusion
Receivables age into write-offs gradually, following patterns that are visible in payment history, dispute logs, and order data well before an account becomes a collections problem. The gap isn't a lack of data — it's that most AR processes only look at that data on a fixed schedule: an annual credit review, a weekly aging report, a dunning cadence that doesn't know the difference between a slow payer and a disputed invoice. An agent that scores risk continuously, segments dunning by actual behavior instead of a single cadence, and triages deductions against source documents doesn't remove the judgment calls that matter — credit holds, write-offs, strategic-account handling — it just gets the right accounts in front of the right person while there's still time to act instead of after the invoice has already aged past the point of an easy recovery.
Share this post:
Curious what this means for your business?
Get a personalized ROI estimate, or book a free discovery workshop with our team.