# Agent context: Benefit Overpayments From Fraud and Error

Source: https://aiplaybooks.app/industries/public-sector/benefits-administration/benefit-overpayments
AI suitability: conditional

## Objective

Reduce benefit overpayments from fraud and error by measuring their causes, verifying circumstances against authoritative records, and directing caseworker review to the riskiest claims, without delaying, refusing, or accusing legitimate claimants and without any automated decision that takes money from a claimant.

## Constraints

- Never raise an overpayment debt, suspend or refuse a claim on suspicion, or record fraud automatically. A caseworker or investigator decides each one, and the claimant is told what the agency holds and can respond first. An award may be calculated from an authoritative same-period figure, such as employer-reported pay, only where the benefit rules provide for it and the claimant can dispute the figure.
- A data mismatch or a risk score selects a case for review and is never itself the evidence. Do not shift the burden onto the claimant to disprove a mismatch.
- Compare records over the same period the benefit rules use. Do not spread an annual or quarterly figure across assessment periods to infer what was earned in each.
- Do not use nationality, ethnicity, sex, disability, or proxies for them, such as language skills or neighbourhood, as model inputs.
- Hold protected characteristics separately and use them only to measure referral rates and confirmed rates per group.
- Refer a random control group through the same route as model referrals, and do not show the caseworker the score or that a model made the referral.
- Train only on information available at the moment of the decision, and correct for the selection bias in investigated cases with random-sample outcomes.
- In the EU, a system that evaluates eligibility for or grants, reduces, revokes, or reclaims public assistance benefits is high-risk under Annex III of the AI Act, with obligations applying from 2 December 2027.

## Inputs

- Claims and awards as assessed, with declared circumstances, household members, and every reported change.
- Third-party records with source and period, such as employer-reported earnings, tax, death, prison, and other benefit records.
- Random-sample review findings with the overpayment, its cause, and whether it was claimant fraud, claimant error, or official error.
- Outcomes of targeted reviews and investigations, including those that found the claim correct, and reconsideration and appeal outcomes.
- Protected characteristics, held apart from model inputs, for fairness analysis only.

## Outputs

- An annual estimate of the overpayment rate and value per benefit and per cause.
- Data-match tasks for caseworkers, each naming the record, its source, and the period it covers.
- Where justified, a referral list of high-risk claims or requests mixed with a random control group, sized to caseworker capacity, with no score visible to the caseworker.
- A fairness report per referral route with referral and confirmed rates per group, published where the agency's transparency rules require it.

## Success criteria

- The estimated overpayment rate falls for the causes targeted, measured by the random sample, not by targeted review alone.
- Each targeted route's confirmed rate beats the random control group by count and by value.
- No group's referral rate rises without a matching rise in its confirmed rate, or the agency retrains or withdraws the route when it does.
- Payment delay for legitimate claimants and the share of decisions overturned on reconsideration or appeal do not rise.

## Known pitfalls

- Treating a data mismatch as proof of an overpayment and shifting the burden onto the claimant to disprove it, which is how Australia's Robodebt scheme raised debts a court found unlawful.
- Letting an automated system decide fraud without a caseworker, as Michigan's unemployment insurance agency did between 2013 and 2015, with an error rate later reported at 93%.
- Training a model only on cases that were investigated, so it learns the biases of whoever chose which claims to investigate.
- Using nationality or a proxy for it as a risk factor, which the Dutch childcare benefits scandal showed can be discriminatory even when nationality has no bearing on entitlement.
- Measuring fairness only by how often each group is referred, rather than also by how often each group's referrals turn out to be correct.
- Showing the caseworker the score or telling them the model made the referral, which invites them to confirm it rather than examine the case.
- Counting savings from reviews without checking how many decisions are later overturned or how many legitimate claimants were delayed.

## When not to use AI here

- The agency has not measured its overpayment rate and causes from a random sample, so it cannot tell whether a model is targeting the losses that matter.
- The losses being targeted come from a circumstance that authoritative data or a legal power to request records can verify directly, such as employee earnings reported by employers or, in the UK under the 2025 Act's eligibility verification power, information held by banks.
- Labels come only from cases that were selected for investigation, with no random sample to correct the selection.
- The agency cannot measure the model's referral and confirmation rates for the groups it is legally bound not to discriminate against.
- The output would decide a debt, a suspension on suspicion, or a fraud finding without a caseworker deciding, which is what Australia's Robodebt scheme did before it was found unlawful.

## Candidate approaches

- Annual random-sample measurement of fraud and error (primary): An in-depth review of a random sample of benefit claims each year that estimates the overpayment rate per benefit and splits it by cause into claimant fraud, claimant error, and official error, so counter-fraud effort goes where the money is lost.
- Earnings and circumstance data matching (primary): Matching each claim against authoritative records other bodies hold, such as employer-reported pay, tax records, and death and prison records, using same-period figures to adjust awards with a right to dispute, and sending every retrospective or uncertain mismatch to a caseworker who asks the claimant first.
- Risk-ranked referrals for caseworker review (supporting): A model trained on past review and investigation outcomes that ranks new claims or advance requests by the risk of fraud, and refers the highest-risk ones, mixed with a random control group, to a caseworker who sees neither the score nor that a model chose the case.
- Automated overpayment and fraud decisions (not-recommended): Raising a debt, suspending a payment, or recording a fraud finding automatically from a data match or a risk score without a caseworker deciding, described here because it has repeatedly harmed claimants and is the approach this playbook recommends against.

## Human oversight

A caseworker decides every refusal, suspension, change to an award, and overpayment, and a counter-fraud investigator and the prosecuting authority decide any fraud allegation. A data match or a risk score only selects a claim for a caseworker to look at, and it is never itself the evidence. The claimant is told what the agency believes has changed and can respond before a debt is raised. The department's senior leaders own the decision to deploy any risk model, review its fairness analysis at least yearly, and publish what it does.

## Sources

- [Fraud and error in the benefit system: financial year ending (FYE) 2026 estimates](https://www.gov.uk/government/statistics/fraud-and-error-in-the-benefit-system-financial-year-ending-fye-2026-estimates), Department for Work and Pensions, official statistics, 2026, research
- [Tackling benefit overpayments due to fraud and error](https://www.nao.org.uk/wp-content/uploads/2025/10/Tackling-benefit-overpayments-due-to-fraud-and-error-Summary.pdf), National Audit Office, 2025, research
- [Real Time Information](https://data.parliament.uk/DepositedPapers/Files/DEP2024-0673/138_Real_Time_Information_V12.0.pdf), Department for Work and Pensions, Universal Credit staff guidance deposited in Parliament, 2024, practitioner
- [UC Advances model fairness assessment](https://assets.publishing.service.gov.uk/media/6876219388da2e5804bb6a17/UC_Advances_model_fairness_assessment_-_July_2025.pdf), Department for Work and Pensions, 2025, case-study
- [DWP: Universal Credit Advances Model](https://www.gov.uk/algorithmic-transparency-records/dwp-universal-credit-advances-model), Department for Work and Pensions, algorithmic transparency record, case-study
- [Unemployment Insurance: Estimated Amount of Fraud During Pandemic Likely Between $100 Billion and $135 Billion](https://www.gao.gov/products/gao-23-106696), U.S. Government Accountability Office, 2023, research
- [Inside Amsterdam's high-stakes experiment to create fair welfare AI](https://www.technologyreview.com/2025/06/11/1118233/amsterdam-fair-welfare-ai-discriminatory-algorithms-failure/), MIT Technology Review with Lighthouse Reports and Trouw, 2025, case-study
- [Suspicion Machines](https://www.lighthousereports.com/investigation/suspicion-machines/), Lighthouse Reports, 2023, case-study
- [Xenophobic machines: Discrimination through unregulated use of algorithms in the Dutch childcare benefits scandal](https://www.amnesty.org/en/documents/eur35/4686/2021/en/), Amnesty International, 2021, research
- [Case Over the Michigan Unemployment Insurance Agency's Faulty Automated System Finally Settled](https://stpp.fordschool.umich.edu/sites/stpp/files/2024-08/stpp-midas-explainer.pdf), Science, Technology, and Public Policy program, University of Michigan, 2024, case-study
- [Human–AI Interactions in Public Sector Decision Making: "Automation Bias" and "Selective Adherence" to Algorithmic Advice](https://doi.org/10.1093/jopart/muac007), Journal of Public Administration Research and Theory, 2022, research
- [Algorithmic decision-making and system destructiveness: A case of automatic debt recovery](https://doi.org/10.1080/0960085x.2021.1960905), European Journal of Information Systems, 2021, research
- [Beyond the bubble that is Robodebt: How governments that lose integrity threaten democracy](https://doi.org/10.1002/ajs4.122), Australian Journal of Social Issues, 2020, research
- [Debt by design: The anatomy of a social policy fiasco – Or was it something worse?](https://doi.org/10.1111/1467-8500.12479), Australian Journal of Public Administration, 2021, research
- [Government Response, Royal Commission into the Robodebt Scheme](https://www.pmc.gov.au/resources/government-response-royal-commission-robodebt-scheme), Australian Government, Department of the Prime Minister and Cabinet, 2023, research
- [Regulation (EU) 2024/1689, the Artificial Intelligence Act](https://eur-lex.europa.eu/eli/reg/2024/1689/oj/eng), Official Journal of the European Union, regulation
- [Regulation (EU) 2026/1744, the Digital Omnibus on AI](https://eur-lex.europa.eu/eli/reg/2026/1744/oj/eng), Official Journal of the European Union, 2026, regulation
- [Code of Practice on Eligibility Verification Notices](https://www.gov.uk/government/publications/code-of-practice-on-eligibility-verification-notices), Department for Work and Pensions, under the Public Authorities (Fraud, Error and Recovery) Act 2025, 2026, regulation