How to Prioritise Collections with AI in 6 Steps

Key Takeaways

• Rank overdue invoices by expected cash impact (amount x late-payment risk x touch effect): EUR 120,000 five days late can outrank EUR 3,200 at 95 days.
• Take disputed amounts out of the score, hold promised invoices until the promised date, and move broken promises up.
• A collector makes 15-20 calls a day, so everything below the human cut-off (about rank 40 for two collectors) goes to an AI agent.
• Among Europe’s 1,000 largest companies, DSO was 48.5 days in 2024, with an 18-day gap between median and top performers (The Hackett Group, 2025).

In This Article

• Key Takeaways
• What is AI collections prioritisation?
• Why does oldest-first collections leave cash on the table?
• What do you need before you prioritise collections with AI?
• How to prioritise collections with AI in 6 steps
• Example: one collections worklist, ranked two ways
• Common mistakes when prioritising collections with AI
• How do you measure whether AI prioritisation works?

What is AI collections prioritisation?

AI collections prioritisation is a scoring method that ranks every overdue invoice by the cash a collection touch is likely to bring in, so the highest-value work happens first.

The output is a daily ranked worklist: every overdue invoice gets a next action and an owner, collector or AI agent. An aging report sorts by time; the score sorts by expected cash.

Prioritisation is the first step of the wider loop covered in how AI automates collections: dunning, calls, promise tracking and dispute handling.

Why does oldest-first collections leave cash on the table?

Oldest-first collections leaves cash on the table because days overdue says nothing about the amount at stake, the customer’s payment habits or whether a call will change anything.

Among Europe’s 1,000 largest companies, DSO rose to 48.5 days in 2024, with an 18-day gap between median and top performers (The Hackett Group, 2025).

More than 12% of European businesses’ revenues are now paid late, above the level they consider sustainable (Intrum, 2026). Intrum surveyed 8,385 businesses in 20 European countries.

Western European companies lose 1.6% of B2B invoiced turnover to credit losses on average, and nearly one in four report losses of up to 5% (Atradius, 2026).

Collections prioritisation by age spends the first calls on small, old balances that are often lost, while a large invoice from a slipping customer waits because it is only five days late.

What do you need before you prioritise collections with AI?

To prioritise collections with AI you need open items, at least 12 months of payment history, promises, disputes, contacts and credit limits in one place, plus a written policy.

• Data: ERP open items per invoice, 12+ months of cleared payments (more if seasonal), each promise to pay with date and amount, open disputes, the contact log, the credit limit and any credit-risk signal you buy.
• Policy: which accounts stay with a named person, what may never be automated, and when a case goes to an agency.
• Baseline: the share of overdue invoices touched within 24 hours, DSO and promise-kept rate, measured over the four weeks before you switch.

How to prioritise collections with AI in 6 steps

Prioritising collections with AI takes six steps: unify the data, predict each payment, score expected cash impact, apply policy rules, split the worklist by capacity, and recalibrate.

‍

‍

1. Put every overdue invoice and its customer history in one record, refreshed daily.
2. Predict each invoice’s payment: the chance it is paid late and the expected date.
3. Score expected cash impact: open amount x late-payment probability x touch effect.
4. Apply policy rules for disputes, promises and named accounts.
5. Split the worklist by capacity: collectors take the top, an AI agent the rest.
6. Re-rank daily and recalibrate monthly.

Step 1: Put every overdue invoice and its history in one record

Scores are per invoice; exposure is per customer. Keep both: each invoice carries its customer’s total open balance, credit limit use, last promise and last dispute.

Refresh it daily after cash application, so already-paid invoices never reach the list. Vero, the agent in Transformance, works on Customer 360, one graph per customer that is its working memory, so the last promise, dispute and contact are known before every touch.

Step 2: Predict each invoice’s payment

Each invoice gets two outputs: the probability that it is paid more than N days late (N from your policy, say 30) and an expected payment date.

The strongest B2B inputs are the customer’s own: days-late history, invoice size against its usual invoices, month or quarter end, past disputes, response to reminders and recent partial payments. Features and models are covered in machine learning payment prediction.

Step 3: Score expected cash impact

The formula: priority score = open amount x probability of late payment x probability that a touch changes the outcome.

Don’t skip the third factor. A reliable payer three days late will pay without a call; a customer who pays faster after contact has a high factor. Estimate it from how much sooner similar invoices were paid after a touch.

Modifiers multiply the score; keep them few and written down: a broken promise (x1.5 in the example below), an invoice about to cross the next aging bucket (x1.25), rising credit limit use, a credit-risk signal. That makes it risk-based collections, not amount-based.

Risk is rising: 53% of European businesses expect late or non-payment to increase within 12 months (Intrum, 2026); EU bankruptcies rose 5.7% quarter on quarter in Q2 2026 (Eurostat, 2026).

Score in group currency at today’s rate, so FX risk doesn’t distort cross-border comparisons.

Step 4: Apply policy rules before anyone is contacted

• Disputes: the disputed amount leaves the score and goes to dispute management; the undisputed balance stays on the list.
• Promises: an invoice with an open promise is held until the promised date plus one business day; if the money doesn’t arrive, the broken-promise modifier moves it up.
• Named accounts: strategic customers, payment plans and legal holds go to their owner with the score attached.
• Agency cases: accounts past your policy threshold leave the worklist; see when to use collections software or an agency.

Step 5: Split the collections worklist by capacity

Two collectors at 15-20 calls a day cover 30-40 calls, so the human cut-off on the collections worklist sits around rank 40.

That cut-off answers who to call first: collectors take the top 40, where the cash is largest and judgement matters most. Everything below gets a touch today from the agent: an email for the reliable payer, a call for one who ignores email.

A human collector makes 15-20 calls a day; the CollectPulse calling agent makes 15-20 calls an hour in 70+ languages. CollectPulse actions 100% of overdue invoices within 24 hours, against 30-40% for manual collections teams.

Step 6: Re-rank daily and recalibrate monthly

Re-score daily, after cash application and every reply, promise or dispute. Monthly, compare predicted with actual payment dates, check that the top 40 brought in the most cash per touch, and adjust weights where the score was wrong. Date each change.

An AI agent in accounts receivable starts by recommending and moves to executing routine touches once its choices match what collectors would do. Transformance uses a 4-level approval model (read, recommend, execute, post), with a human in the loop for ERP postings.

Example: one collections worklist, ranked two ways

In this illustrative six-invoice example, oldest-first order puts a EUR 3,200 invoice at the top, while AI scoring puts a EUR 120,000 invoice with a broken promise first.

Figures are invented, in group currency; the team has one collector call free for these six. Score = undisputed open amount x late-payment probability x touch effect x modifier: this article’s method, not a product’s scoring model.

For customer A: 120,000 x 0.35 x 0.6 = 25,200, and the x1.5 broken-promise modifier takes it to 37,800. For D, 9,500 x 0.80 x 0.4 = 3,040, and x1.25 gives 3,800.

Oldest-first spends the one call on EUR 3,200 that is probably gone, chases F before its promised date and leaves A fifth. The AI order puts the call on A; the agent covers B, C, D and E the same day.

‍

‍

Common mistakes when prioritising collections with AI

The costly mistakes in AI prioritisation sit in the rules around the score, not the model: disputed money stays in, promises get chased early, and nobody works the list.

• Ranking by days overdue with an AI label on top: the score never sees amount.
• Scoring disputed amounts or calling before the promised date: both chase money the customer has already addressed.
• A 400-line list for a team that can make 40 calls: the bottom 360 are never touched.
• Never checking predictions against actual payments, so weights drift every quarter-end.

How do you measure whether AI prioritisation works?

Measure AI prioritisation by coverage, cash and accuracy: the share of overdue invoices actioned within 24 hours, collections effectiveness, promise-kept rate and predicted versus actual payment dates.

• Coverage = overdue invoices touched within 24 hours / all overdue invoices.
• Collections effectiveness index (CEI) = the share of collectible AR collected in the period.
• Promise-kept rate = promises paid in full by the promised date / promises made.
• Prediction error = median gap in days between predicted and actual payment.
• Top-of-list yield = cash collected from the human top 40 / all cash collected.

The guide to collections KPIs covers the wider set. CollectPulse customers see DSO fall by 8-12 days on average (Transformance customer data, 2026).

Conclusion: rank by expected cash, cover the rest the same day

Prioritising collections with AI means ranking by expected cash, not by age, and giving every invoice below the human cut-off to an AI agent within 24 hours. Write the policy down first, start in recommend mode, and judge coverage, promise-kept rate and DSO against your baseline.

Frequently Asked Questions

How do you prioritise collections with AI?

Score every overdue invoice on amount, late-payment probability and touch effect, adjust for promises and disputes, and rank by expected cash; collectors work the top and an AI agent the rest.

Who should collectors call first?

Call first the invoices with the highest expected cash impact where a person makes a difference: large balances, broken promises, accounts near their credit limit. Reliable payers get an automated reminder.

What is risk-based collections?

Risk-based collections orders collection work by the likelihood and cost of non-payment, not by invoice age, so a slipping customer comes before a reliable one with an older invoice.

Can generative AI prioritise collections?

Not on its own: the ranking comes from predictive models trained on your payment history, while generative AI writes the messages and reads the replies. Generative AI in finance explains where each fits.

How long does it take to set up AI collections prioritisation?

With Transformance, a proof of value takes 4-5 weeks on one entity, and a full rollout takes up to 4-8 weeks. Both assume the step 1 data is accessible.