What problem does it solve?
Most collections teams hold more overdue accounts than they have collectors. Without a score, the worklist is ordered by balance, by age of the debt or by the order in which cases arrived, and every debtor in the same bucket receives the same letters and calls. That spends effort on people who would have paid after one reminder and misses people whose payments are slowly falling away. HMRC's Business Propensity to Pay model replaced a simple segmentation by exposure to the pandemic, and before that segmentation most customers received the same debt collection journey. Camden Council says its rent teams struggled to focus on the right areas and missed chances to give preventive support to tenants who could not pay.
A score also separates two groups who need opposite treatment. A debtor who is likely to pay needs a prompt and an easy way to pay. A debtor whose payments are declining may be in financial difficulty and needs an earlier and more careful conversation or a referral for support. Used carelessly, the same score can do harm. Australia's Robodebt scheme shows what happens when an automated process raises debts, puts the burden of disproving them on the person and has no human check.
- Camden Council reports that its total council housing rent arrears were second highest in London and nearing 18 million pounds in December 2023.Camden Council: RentSense (algorithmic transparency record) (2023)
How does it work?
- Build the account history. Link every debt a customer holds, as HMRC does across VAT, PAYE and Corporation Tax, and add payment history, balances, changes of payment method and, where it is lawful, contact history.
- Train and calibrate the model. A supervised model, usually gradient boosted trees (HMRC and DWP both use them), predicts whether a debtor will pay or engage within a set window. HMRC calibrates the raw output with a logistic regression so the score reads as a likelihood.
- Band the scores. Scores become bands such as high, medium and low, because collections systems and caseworkers act on bands.
- Choose the treatment. The band, together with other rules, sets the timing, frequency, channel and wording of the first contact, or places the case on a caseworker's list with the reasons and suggested actions. HMRC says the score is only one factor in the collection path.
- Leave consequences to people. The score orders work and selects communications. People agree payment plans, handle vulnerability and decide on enforcement.
- Retrain and monitor. DWP trains a new model every month and accepts it only above a precision floor. HMRC runs its model monthly, watches input drift on a dashboard and reviews it every year.
- Audience
- Back office
- Autonomy
- Assist
- Adoption
- Early adopters
- Channels
- Internal tools, API and system to system
What is it worth?
Benchmarks are computed from the public deployments below: one data point per organization per KPI, with who made each claim.
| KPI | Median | Reported range | Data points | Claimed by |
|---|---|---|---|---|
| Interactions handled | Not pooled | 100,000 to 4 million | 2 | 2 organization |
| Accuracy | Too few to pool | 86% | 1 | 1 organization |
| Revenue recovered | Not pooled | Not pooled: up to USD 50 million | 0plus 1 up to | 1 organization |
Value drivers: Risk and loss reduction, Employee productivity, Lower cost to serve, Customer experience.
Indicative value
A water utility with 1 million household accounts
USD 240,000 to USD 2.2 million
Additional overdue money recovered per year
How this is calculated
Formula: accounts * arrearsShare * overdueBalance * recoveryUplift. The low scenario uses every low input, the high scenario every high input.
| Input | Low | High | Basis |
|---|---|---|---|
| Household accounts accounts, accounts | 1,000,000 | 1,000,000 | The reference utility. |
| Share of accounts that fall into arrears in a year arrearsShare, fraction of accounts | 0.08 | 0.12 | Editorial assumption, replace with your own arrears data. |
| Average overdue balance per account in arrears overdueBalance, USD per account | 150 | 300 | Editorial assumption, replace with your own arrears data. |
| Relative increase in overdue money recovered after targeting recoveryUplift, fraction of overdue balance | 0.02 | 0.06 | Deliberately below the 8 percent total growth in collections that the New York State tax department reports for its first year, because the authors attribute only 40 to 50 million of the 83 million dollar growth to the system and no other record on this page publishes a recovery result. |
What it leaves out: Gross cash only. It leaves out the cost of building and running the model, money that would have been recovered anyway, the effect of earlier support on bad debt and customer retention, and any cost of treating vulnerable customers differently.
Who already uses it?
4 public deployments, strongest evidence first. Grades: A regulator or audit, B the organization itself, C vendor case study, D anonymous or estimate.
HM Revenue and Customs
United Kingdom · Government and public sector · 2026
HMRC's Debt Management function runs a machine learning model called Business Propensity to Pay that scores about 4 million business debtors each month on how likely they are to engage with debt recovery. A gradient boosted decision tree, built in SAS Viya, learns from debt value and past repayment behaviour across VAT, PAYE and Corporation Tax debts linked at customer level.
The segmented score helps decide which collection journey a debtor receives, including the timing, frequency and content of the first letters. HMRC says the model replaced a simple segmentation based on exposure to the pandemic (before that segmentation, most customers received the same journey) and that the score is only one factor in the journey. The record discloses no recovery results.
- Interactions handled: about 4 million, per month
"The model is run on a monthly basis and creates scores for approximately 4 million customers each month."
Claimed by: organization
Department for Work and Pensions
United Kingdom · Government and public sector · 2025
The Child Maintenance Service (CMS) in the Department for Work and Pensions (DWP) uses a machine learning model to predict which paying parents on compliant Collect and Pay cases are likely to miss a payment. A model trained each month on case, payment, liability and contact data (an XGBoost classifier with SHAP explanations) scores around 100,000 cases and shows caseworkers a high, medium or low risk with the reasons and suggested actions, so they can step in before a payment breaks down. Caseworkers decide what to do; the score is one input. The tool went to a wider group of caseworkers in March 2025 and DWP reports validation accuracy of 0.86 and precision of 0.79.
- Interactions handled: about 100,000, cases scored each month
"The tool is used at the start of each month to predict the risk of non-compliance, around 100,000 cases are processed by the model and presented to the CMS2012 System."
Claimed by: organization - Accuracy: 86%, validation data from May 2025
"Accuracy: 0.86 Precision: 0.79"
Claimed by: organization
London Borough of Camden
United Kingdom · Government and public sector · 2025
Camden Council's housing income team piloted Mobysoft's RentSense for six months to decide which council tenants in rent arrears housing officers should contact first. The supplier's supervised machine learning reads rent transaction data from the council's NEC Housing system, looks at payment patterns and payment types such as direct debit and housing benefit, and predicts who will or will not pay. Each week officers get a prioritised list with plain reasons, such as an increasing balance or a broken arrangement.
Mobysoft's own statement, quoted inside the record, is that accuracy is typically 95%+ across social landlords; Camden does not report it for the pilot. Camden says there is no automated decision making, that the supplier would not disclose how the algorithms weigh features, and that it would test the outputs for bias against equalities data held by the council. Total council rent arrears were nearing 18 million pounds in December 2023. No pilot result is published.
No outcome disclosed.
New York State Department of Taxation and Finance
United States · Government and public sector · 2010
New York State's tax department built a collections system with IBM Research and IBM Global Business Services that decides, for each delinquent taxpayer, which collection action to take next. It models every taxpayer from about 200 features of their history and uses constrained reinforcement learning to choose actions that maximise long term returns within staff capacity and legal limits, instead of applying one rule to every case. The system became operational in December 2009.
According to the department's own paper, total collections from delinquent revenue rose 83 million dollars (8 percent) from 2009 to 2010 with the same resources. Because collections had grown 2 to 4 percent a year before, the authors place the increase due to the system in its first year at 40 to 50 million dollars. The system chooses collection actions and is not described as a payment likelihood score.
It is an early and well documented example, published in the Franz Edelman Award section of the INFORMS journal Interfaces; its current status is unknown.
- Revenue recovered: up to USD 50 million, first year after launch (2010), the authors' estimate of the increase due to the system, given as a range of 40 to 50 million dollars
"increases ranging between 2–4 percent during the past several years; this places the revenue increase from CISS in its first year in the range of $40–$50 million already realized, with a promise of $120–$150 million over the three-year period originally projected."
Claimed by: organization
How do you implement it?
A model agnostic playbook: what to prepare, the order to build in, and what goes wrong.
Data you need
- Account level payment history with outcomes (paid, partly paid, broken arrangement) for at least a year
- Debts linked across products or tax types at customer level
- Contact and arrangement history from the collections system
- A lawful basis and a record of which fields are excluded, such as protected characteristics
Systems to integrate
- Collections or debt management system that can store a score and a band
- Case management or caseworker desktop to show the score with reasons
- Letter, SMS and email engine that can vary timing and wording by band
- Data warehouse or feature store with monthly refresh and drift monitoring
Complexity: Medium
The build needs an extract of account level payment history with known outcomes, usually a year or more, joined across every product a debtor holds, and a collections or case system that accepts a score and a band. HMRC tuned its model on 9.5 million rows and DWP trained on about 200,000 labelled cases. Wiring a score into letter timing or worklist order is a small change to a collections system. A score that sits near enforcement, disconnection or benefit decisions needs model risk review, bias testing and a recorded human decision.
- 1
Name the outcome the model predicts
Pick one measurable event, such as a payment within 30 days of the first reminder, engagement with the recovery process (HMRC) or a missed payment on a case that is currently compliant (DWP). A vague target produces a score nobody can check.
- 2
Decide what changes by band before training
Write down the letter timing, channel, caseworker order or support offer that differs for each band. A score that changes no action cannot show any value.
- 3
Choose an explainable model and test on later months
Compare a few simple model families on the same data, as HMRC did with a single tree, a random forest and gradient boosting, and keep the most recent months aside as a hold out. Set the minimum precision a retrained model must reach, as DWP does.
- 4
Test for bias before anyone uses the output
Keep protected characteristics out of the inputs, then compare who lands in each band against the equalities data you hold. Camden planned this check for its pilot and kept that data away from the supplier.
- 5
Measure uplift against a holdout
Leave a random share of debtors on the old treatment for a few cycles and compare the money recovered, the cure rate and the cost to collect in both groups. The difference is the uplift the model can claim.
- 6
Monitor monthly and review every year
Watch input drift, band mix and complaint rates each month, and retrain when behaviour changes. HMRC's current model superseded an earlier segmentation built around exposure to the pandemic.
Guardrails
- The score orders work and selects communications; no enforcement, write off or disconnection follows from a score alone
- Protected characteristics and special category data stay out of the model inputs and are used only to test outputs
- A hardship or vulnerability flag overrides the score and routes the case to a person
- Caseworkers see the reasons for a score and the suggested actions beside the number
- A retrained model goes live only above a precision floor and after a drift check
KPIs to instrument
- Payment or cure rate by band, against a holdout group on the old treatment
- Precision of the high risk band, meaning the share of high risk cases that did miss a payment
- Contacts and complaints per 1,000 debtors by band
- Share of debtors with a vulnerability flag in each band
- Cost to collect per pound recovered
Human in the loop
Collectors and caseworkers read the score, check the case file and decide on contact, payment plans and referrals for support. Managers sample low band cases every month to catch debtors the model ranked too low, and review the model owner's bias analysis before each retraining.
Common failure modes
- The score is treated as fact
- Officers read a probability as a certainty. Camden listed this risk in its record. Train staff on the limits, show reasons beside the score and track how often officers override it.
- Vulnerable debtors fall into the low band
- A debtor whose payments are steady until a crisis hits looks safe. Camden also lists the risk of high priority cases treated as low priority. Add vulnerability signals and sample low band cases.
- The model finds people who would have paid anyway
- High scores can simply mark customers who pay after one reminder. Measure uplift against a holdout group and report it next to accuracy.
- Behaviour changes and the model does not
- A model trained before a price rise, benefit change or pandemic stops describing today's debtors. Monthly drift checks and scheduled retraining catch it.
What are the risks and rules?
EU AI Act
Depends on design
A score that only orders a collector's worklist or chooses the timing of a reminder is not listed in Annex III. The same score becomes high risk under Annex III point 5(b) if it is used to evaluate the creditworthiness of natural persons, and may fall under point 5(a) if a public authority uses it to grant, reduce, revoke or reclaim essential public benefits or services. Keep the scope to prioritisation and document where the line sits.
Guidance
- Guidance on AI and data protection (Information Commissioner's Office, Europe). Covers fairness, accuracy and explanation duties when personal data feeds an AI system, including statistical accuracy of the scores.
- Guidance for firms on the fair treatment of vulnerable customers (FG21/1) (Financial Conduct Authority, Europe). Sets out how firms should recognise and treat customers in vulnerable circumstances, which applies to FCA regulated firms that use a score to select collections treatment.
- Algorithmic Transparency Recording Standard hub (UK Government, Europe). The public records from HMRC, DWP and Camden Council on this page follow this standard and show what a public body discloses about a debt model.
Controls to put in place
- Model card with the outcome predicted, training window, inputs, exclusions and the validation result
- Bias and fairness analysis before launch and after each retraining, with a named owner
- Monthly dashboard of drift, band mix, overrides and complaints
- Logged score, band and treatment for every case, so a complaint can be traced
- Public transparency record where the deploying body is a public authority
When it went wrong elsewhere
- Royal Commission into the Robodebt Scheme: report volume 1 (archived). Robodebt raised welfare debts by averaging income and placed the burden on recipients to disprove them. It used no propensity model. It is the best known example of an automated debt process without a human check, and the lesson applies to any score that triggers recovery action.
Frequently asked questions
- How accurate are propensity to pay models?
The published figures come from validation data. DWP reports accuracy of 0.86 and precision of 0.79 for its child maintenance model, which predicts missed payments on compliant cases. Mobysoft, the supplier of RentSense, claims typical accuracy of 95% or more, and Camden Council quotes that claim in its transparency record without verifying it. Accuracy depends on how many debtors miss payments, so compare precision and uplift against a holdout.
- Does propensity to pay scoring increase recovery?
No record on this page measures the recovery effect of a propensity to pay score. The closest result comes from New York State's tax department, whose system chooses collection actions for each delinquent taxpayer and is a related optimization method, not a payment likelihood score. Its collections rose 83 million dollars (8 percent) from 2009 to 2010, and the authors estimate that 40 to 50 million dollars of that came from the system, because collections normally grow 2 to 4 percent a year. HMRC, DWP and Camden Council disclose how their models work and how many cases they score, but no recovery result.
- Is propensity to pay scoring high risk under the EU AI Act?
It depends on use. Ordering a worklist or timing a reminder is not an Annex III use. Scoring that evaluates the creditworthiness of natural persons is high risk under point 5(b), and a public authority using a score to reclaim essential benefits or services may fall under point 5(a).
How to cite this page
Blits.ai AI Use Case Library, "AI propensity to pay scoring for arrears and debt collection", last verified 13 October 2026, https://www.blits.ai/ai-use-cases/propensity-to-pay-scoring. Licensed under CC BY 4.0. Method: how we verify use cases.
Changelog
- 13 October 2026: First published