Think Build Implement Repeat
London, UK +44 7367 067226
WhatsApp FOLLOW f in X
AI & Machine Learning

Predicting Late Payments Before Invoices Go Overdue

Last updated:

Chasing after the due date is already too late

Most credit control works on an aged debt report. An invoice crosses its due date, it appears on the list, somebody sends a reminder, then a firmer one, then picks up the phone. By the time a human is involved, the invoice is already late and the customer has already decided which suppliers to pay first this month.

The customers who pay late are rarely a surprise in hindsight. The same names recur. What credit control lacks is a way to see, on the day an invoice is issued, which of this week's 400 invoices are the ones worth a friendly call before the due date.

What the model predicts

There are two useful ways to frame late payment prediction, and the choice depends on how you act.

FramingOutputBest when
ClassificationProbability the invoice is paid more than N days lateYou run a simple 'call the risky ones' process
RegressionExpected days lateYou also want to feed expected payment dates into a cash flow forecast
Survival modelProbability of payment by each future dateTiming matters and many invoices are part-paid or disputed

For most businesses we start with classification, because it maps neatly onto a work queue. If the finance team also wants better cash projections, expected payment dates become an input to cash flow forecasting with machine learning, and that is where the regression or survival version earns its extra complexity.

The signals that predict a late payment

  • The customer's own history. Average days late over the last six and twelve months, and whether that is getting worse. Trend matters more than the long-run average.
  • Invoice size relative to what they normally receive. An unusually large invoice often goes through an extra approval step at the customer's end.
  • Terms and timing. Invoices dated just after a customer's monthly payment run can wait nearly a month.
  • Disputes and credit notes. Any recent query on the account is a strong early warning.
  • Purchase order quality. Invoices missing a PO number or sent to a generic inbox sit in someone's queue.
  • External credit signals. Useful for larger customers, optional for the rest.

Notice how many of these are about process at the customer's side rather than their ability to pay. That is typical. In B2B, a lot of late payment is administrative, which is good news, because administrative lateness responds well to a polite early nudge.

An illustrative example

Picture a facilities maintenance company invoicing 1,200 times a month across 300 commercial clients, on 30-day terms, with about a quarter of invoices paid more than two weeks late. One credit controller cannot call 300 invoices, so she calls none until they are overdue.

A model scoring each invoice at issue might put 150 of them in a high-risk band, and in testing on last year's data that band might contain most of the invoices that ended up badly late. A call or a check that the invoice reached the right person, made in the first week, is a far lighter piece of work than chasing at day 45. The numbers here are an example, but the shape of the result, a small band containing most of the problem, is what we usually see when we test this on real ledgers.

Turning scores into a credit control process

  1. Score every invoice when it is raised, and rescore weekly as payment behaviour and disputes change
  2. Put high-risk invoices into a pre-due-date queue with the top reasons shown
  3. Confirm receipt and approval status with the customer's accounts payable team within a few days of sending
  4. Automate the routine reminders for low-risk invoices so people only handle the exceptions
  5. Record the outcome of every contact so the next model version learns which interventions worked

Step four is where most of the time saving comes from. Our automation work for finance teams often starts with the reminders and remittance matching, and the prediction model slots in on top once the basics run cleanly. If invoicing itself is still manual, sort that first; our guide to automating invoice billing covers it.

When a model is not worth building

If you raise fifty invoices a month to twenty customers, your credit controller already knows who pays late, and a model adds nothing but a monthly fee. The same is true where a few large customers dominate receivables; those relationships need a person, not a score.

It is also wasted effort when the real cause is on your side. If invoices go out late, with wrong PO numbers or to the wrong contact, a prediction model will faithfully learn that your own mistakes cause lateness. Fix the invoice quality first. It is cheaper and nobody has to trust a probability.

Fairness, credit decisions and data

Using a score to decide who gets an early courtesy call is low risk. Using the same score to cut credit limits or change terms automatically is a different decision, with commercial and in some cases legal consequences if the customer is a sole trader, since that is personal data about an individual. Keep automated scores in the prioritisation role and leave credit decisions with a person who can see the reasons.

At SpiderHunts we also check early whether the model has quietly learned something unhelpful, such as penalising customers in one region because a courier problem delayed their invoices last year. That is a data issue, and it only shows up if someone looks.

Frequently asked questions

Can machine learning predict which invoices will be paid late?

Yes, usually with useful accuracy for businesses with a steady volume of invoices and repeat customers. Payment history, invoice size, terms and recent disputes carry most of the signal. It is much weaker for one-off customers with no history, where external credit data does more of the work.

What accounting data do we need for a payment prediction model?

Invoice date, due date, amount, customer, paid date and ideally part-payments, credit notes and disputes, for at least a year. Most of this can be exported from Xero, Sage, QuickBooks or an ERP. Contact logs from credit control improve the model but are not essential to start.

Does our accounting software already do this?

Some platforms now show predicted payment dates or risk flags, and for a small business they may be enough. A custom model makes sense when you have higher volumes, data from outside the ledger such as delivery or dispute systems, or you need the prediction to drive a specific workflow.

Will contacting customers before the due date annoy them?

Not if it is framed as a service check: confirming the invoice arrived, has a PO number and is with the right person. Most accounts payable teams welcome that because it saves them a query later. Pressure tactics before the due date are a different matter and we would not recommend them.

Keep reading

Chasing the same slow payers every month?

Send us an anonymised export of a year of invoices and payment dates. We will tell you how predictable your late payers are and whether a model would beat your credit controller's instinct.

Book a free 30-minute call Get a project estimate WhatsApp us

Related services

What we build for problems like this one

AI AgentsMachine LearningAI Integration