Payment risk score: which customer will pay late?

It is the fourth of the month, a Thursday morning. The finance lead opens the AR ageing report and three lines sit there: three invoices past ninety days, all issued in the same quarter, together a meaningful slice of what the business expected to collect that month. All three customers looked fine on the sales side — the quote was approved quickly, the contract signed, the first invoice cleared on time. The real problem is not the day the money stopped arriving. It is that the small signals piling up over the previous three months were never gathered in one place: a part payment here, an unanswered statement there, a promise to pay pushed back twice. Each was seen. None sat next to the others.

This article walks through how to turn those signals into a single number. In order: what a payment risk score actually measures, which data can feed it, why average days late misleads, how to build the score with sensible windows and weights, how to derive thresholds from your own collections history, which decisions the score should influence, how to read it alongside AR ageing, how to wire it into daily collections, how it connects to profitability and cash planning, the mistakes that recur, and how to tell whether the model is earning its keep.

What does a payment risk score actually measure?

A payment risk score is a summary, not a prophecy. What it measures is narrow and useful: how strongly this account's past payment behaviour supports the expectation that the next invoice will clear on the due date. It does not tell you who will default; it tells you who deserves closer attention this week. The distinction matters, because a number claiming the first will freeze commercial decisions unfairly, while a number doing the second points a small collections team at the right accounts.

Be equally clear about what the score is not. It is not a credit rating: it draws on no bureau file, no court record, no industry intelligence — only your own trading history with that customer. It is not an insolvency prediction, and certainly not a decision to end a relationship. The same discipline on the relationship side is what you meet when you build a customer health score; the two answer different questions — one asks whether the customer will stay, the other when their money will arrive.

Which data can feed the score?

Most of the inputs already live in your books. The work is collecting them consistently. Measured per invoice rather than per customer, the useful ones look like this:

• Days late against terms: payment date minus due date for every settled invoice. Keep it at invoice level, not account level.
• The shape of lateness: median and worst case alongside the average. An account that usually pays on time but slips seventy days once a year is a different animal from one that is reliably eight days late.
• Part-payment frequency: settling an invoice in instalments is often the earliest visible sign of a cash squeeze.
• Open balance relative to volume: the same open balance means very different things for a customer billing five figures a month and one billing four.
• Disputes and credit notes: repeatedly queried invoices stretch the collection cycle whatever the customer's intentions.
• Instrument history: a dishonoured cheque or a bounced direct debit is one of the sharpest behavioural signals you will get.
• Responsiveness: how long a statement of account goes unanswered, and whether promises to pay hold.

Most of these only work if the payment lands in the record as it happens. If collection data is reconciled from a bank file once a week, your score is always several days behind reality; once payments and the customer record are wired together, the days-late calculation updates the moment the money clears.

One caution. Not every delay is the customer's doing. A wrong tax rate, a missing purchase order number, or a document raised against the wrong entity will hold an invoice just as effectively as a cash shortage. So track whether the document itself was corrected. A team that understands when to cancel an invoice and when to issue a correction does not charge its own errors to the customer's score.

Why is average days late a misleading summary?

A single average hides direction. Picture two accounts, both averaging fourteen days late over the past year. The first was thirty days late in January and is four days late now. The second paid on time all spring and has drifted to twenty-five days over the last quarter. The average files them in the same drawer, yet one is recovering and the other deteriorating. Where a collections team spends its limited hours depends entirely on that difference.

The fix is to compute behaviour over two windows rather than one — say the last three months against the nine before that. Side by side, the score stops being a snapshot and acquires a direction. The Current Account Credit Risk and Limit screen in Rocketly's Pre-Accounting module is built exactly this way: payment behaviour is calculated in two separate windows and compared, so instead of a flat on-time-rate photograph you also see which way the account is heading. The credit limit suggestion there is not a black box either — the inputs it was derived from are listed in the response, and the output is a status label rather than a silent block on sales.

Which windows and weights should you build it with?

A simple model you can defend beats a sophisticated one nobody can explain. Choose an observation window first: twelve months gives most small and mid-sized businesses enough closed invoices, while seasonal trades are safer with twenty-four. Normalise each input onto a common scale, because days and counts cannot simply be added together. Then set weights according to how well each input separates late payers from punctual ones in your own data — not according to which input sounds most sensible in a meeting.

Three rules save a lot of grief. Never score a new customer low merely for lacking history; keep a separate "insufficient data" state, or the model brands every new account as risky. Decay the weight of older behaviour, so last year's bad quarter does not outvote this year's good one forever. And do not let a single event swing the whole number: a dishonoured payment instrument is better carried as its own flag than buried inside an over-weighted component, because a flag can be read and argued with.

How do you derive thresholds from your own collections history?

The most common mistake here is adopting a threshold read somewhere else. Your sector, customer mix, standard terms and collections capacity are not anyone else's, so your bands cannot be either. The method: take the closed invoices of the past twelve to twenty-four months, write down what "late" means for you (due date plus how many days?), calculate the score retrospectively for each account, then look at what actually happened underneath each score. Cut the bands where observed behaviour separates, not where a round number feels tidy.

The table below is an illustrative skeleton. The labels show structure, not prescribed values, and none of them came from your data. Derive the band boundaries and the actions from your own collections history, and take any step with tax, contractual or security consequences to your accountant before applying it.

Example band	Typical pattern in the data	Question to ask	Step worth considering
Strong	Short, stable delays; no deterioration between the two windows	Could terms be loosened for this account?	Keep current terms; routine reminders are enough
Watch	On-time rate slipping in the recent window	Is the delay about cash or about paperwork?	Share the statement, make contact before the due date, record the reason
Strained	Part payments more frequent, open balance ageing	Should terms on the next order change?	Review payment terms; discuss deposit or security options
Flagged	Dishonoured instrument, or a statement unanswered for a long stretch	Is this a collections problem or a relationship problem?	A human decision outside the score; nothing automatic

No row in that table is a rule. A statement of the form "put any customer below a given score on prepayment" cannot responsibly be made without seeing your data. The defensible version is this: if the gap in realised lateness between adjacent bands is meaningful in your own history, that is where your threshold belongs.

Which decisions should the score actually change?

A score is worth only the decisions it feeds. In practice it earns its place in four: the payment terms on a new order, the order in which the collections team works its list, the tone and cadence of reminders, and whether a deposit or security is worth discussing. None of these should fire automatically. The score brings the proposal; a person makes the call.

On the terms side, the most useful lever is often timing rather than price. Structuring an early payment discount against late payment interest lets you pull cash forward without souring a relationship — provided you have worked out what the discount really costs you in annualised terms first.

There is a record-keeping dimension too: which document the agreed terms actually live in. If the quote says net 30 and the signed contract says net 45, the cause of the delay is your process, not your customer. Being able to see which version of the quote the customer approved settles half of all collections disputes before they start.

How should the score be read alongside AR ageing?

The ageing report looks backwards: which bucket the money is sitting in right now. The score looks forwards: how this account is likely to behave next. When the two agree, the decision is easy. When they disagree, the reading is most instructive. An account that looks clean in ageing but whose score is sliding has not caused trouble yet and is heading that way — which is the cheapest possible moment to intervene. The reverse case, a long-aged item on an account whose score is solid, usually points at a stuck document rather than a stretched customer.

Reading them together also shows where the exposure clusters. If most of the receivable sits with two or three accounts, an average score is falsely reassuring; how dependent your revenue is on a single customer is a separate thing to track, because in that shape one delay carries the whole month on its own.

How does the score connect to day-to-day collections?

A score touches the real world only when it makes someone do something. Rocketly's Payment Risk Score in Pre-Accounting closes that loop: overdue accounts are scored automatically, accounts crossing the threshold you set get a follow-up task, and the task closes by itself once the debt is settled. Two details matter — you set the threshold, not the product, and task closure is not left to somebody remembering. A task list that never clears becomes a pile nobody opens.

The routine layer is carried by Automatic Collection Reminders: notifications for invoices approaching and past their due date, with the reminder email sent from an English, Turkish or Russian template. You are not spending human attention on low-risk customers, and the team's remaining hours go to the flagged accounts.

For reminders to work, the friction in front of the payment has to go too. An email leaving the customer to copy an invoice number and bank details correctly meets willingness with admin; putting a one-click payment link inside the reminder makes visible how much of your lateness was only ever forgetfulness.

How does the score fit into profitability and cash planning?

A late-paying customer is not just a cash problem; they are expensive. The hours spent chasing, the supplier payment pushed back, the overdraft drawn for no other reason — none of it shows up in gross margin, so a large but slow customer keeps looking good on paper. Once you work out what each customer really contributes, the ranking usually rearranges itself.

Looking forwards, the score is an input to the cash plan. Rocketly's Cash Flow Projection spreads open invoices, expenses, recurring items and cheques or notes in the portfolio across the weeks by due date; it marks overdue and undated open items separately, and says on screen when foreign-currency records without a rate have been excluded. Read the score beside that projection and you answer a second question on top of how much is due in week three — namely, how much of it will actually arrive.

The habit travels to the sales side too. The reasoning behind a deal health score and a payment risk score is the same discipline: count observable behaviour, write down the weights, leave the decision to a person.

What goes wrong most often with payment scoring?

• Using the score as a hidden block. If sales cannot see why an order was held, they will find a way around the score. The output should be a status label, not a silent obstruction.
• Charging your own mistakes to the customer. Delay caused by an invoice raised wrongly or sent late has no business in the score.
• Over-weighting one component. A score built on days late alone is blind to part payments and disputes.
• Setting thresholds once and forgetting them. When your standard terms or customer mix change, the bands go stale.
• Keeping the data in separate ledgers. If transactions, receipts and documents live in different systems, the score always summarises an incomplete picture. Being clear about which work is yours and which belongs to your accountant puts both the responsibility and the data ownership back in place.

How do you know the score is working?

A score proves itself only once you measure it after launch. Every quarter, ask four questions. First: did the accounts you placed in the weak bands actually pay later than the rest? If the bands do not separate, the model is not discriminating. Second: has your average collection period shortened under the same terms policy? Third: how many days pass between an account crossing the threshold and the first contact being made? The real gain of scoring often hides in that interval. Fourth: how often did you raise a false alarm and lean on a customer who never needed it?

Asking those four regularly is what turns a score from dashboard decoration into a working tool. The model will stay imperfect; the goal was never perfection, only sending attention to the right accounts. If you want to build that loop on thresholds derived from your own history, you can start a free Rocketly account and begin by bringing your customer transactions into one place.

FAQ

How much history does a payment risk score need?

Usually at least a year of settled invoices before the bands separate meaningfully; two years is a safer base in seasonal trades. With less than that, showing an "insufficient data" state beats publishing a score that invites false confidence.

What score should a new customer get?

None. Scoring an account low because it has no history marks it risky on no evidence. Hold a separate "no history" state and switch the score on once the first few invoices have shown their behaviour.

Is a risk score the same as a credit limit?

No. The score summarises behaviour; a limit proposes a ceiling based on that behaviour and on trading volume. They work together but should not be conflated: the limit is an amount, the score is a way of ranking attention.

Can I borrow thresholds from an external benchmark?

As a starting point for discussion, yes; as the basis for a decision, no. Bands should be derived from your own collections data, and any step with tax, contractual or security implications is worth checking with your accountant first.