The Fraud Frame and Its Limits
Benefits fraud is real. Across OECD countries, it is estimated at 2–5% of total social programme spend. In the United States, improper payments exceed $175 billion annually. These are serious numbers, and agencies are right to address them. The problem is not that agencies focus on fraud. The problem is that fraud has become the primary organising frame for AI strategy in benefits administration — and that frame is both incomplete and, in important ways, actively misleading.
It is incomplete because fraud is not the primary driver of improper payments. The majority of overpayments in most benefits programmes result not from deliberate misrepresentation but from a simpler failure: the agency did not know that a recipient’s circumstances had changed. A recipient who returned to work and did not report their earnings may be committing fraud. More often, they did not understand the reporting requirement, reported it through a channel the agency failed to process, or genuinely did not recognise that their part-time income crossed a relevant threshold. The overpayment accumulated — for months, sometimes years — not because the recipient was running a scheme but because the agency had no mechanism to detect the change until a periodic review caught it.
It is misleading because fraud AI does nothing to address the other half of the dual mandate that every benefits agency carries. The obligation to ensure that entitled citizens receive the support they are owed is not served by better fraud detection. It is undermined by it, if fraud-optimised systems generate false positives that create barriers for legitimate claimants — which they do, in ways that most agencies are not measuring.
What Overpayments Actually Are
The distinction between fraud and circumstance-driven overpayments matters enormously for AI strategy because the two problems require different solutions.
Fraud requires detection of intentional misrepresentation — signals of deliberate concealment, organised ring behaviour, identity manipulation. The models that address this problem are adversarial: they are trying to identify actors who are actively working to evade detection.
Circumstance-driven overpayments require something simpler and more tractable: real-time awareness of when a recipient’s income, household composition, employment status, or living situation changes. A recipient who starts a job, gains a partner, moves address, or inherits an asset is not necessarily concealing anything. The agency simply does not know the change has occurred until a re-determination cycle triggers a review — often months or years later.
The financial scale of this problem dwarfs deliberate fraud in most programme budgets. And its solution is not adversarial AI. It is data integration: connecting the benefits agency’s recipient records to the wage data, housing records, and financial data that would make circumstance changes visible in near-real time rather than at the next scheduled review.
The economic stakes of getting the timing right are material. Recovery rates on overpayments drop sharply as debt ages — often to below 20 cents on the dollar after two years. Earlier detection, at the point when circumstances change rather than at the point when a periodic review finally catches up, is the largest single lever on recovery yield. A programme that detects overpayments 60 days earlier recovers more, writes off less, and generates a smaller improper payment rate — not because it detected more fraud, but because it knew faster.
The Other Side of the Mandate
The human cost of the fraud frame’s blind spot is less easily quantified but more consequential.
Every benefits system has a take-up problem. Entitled individuals who do not claim the support they are owed — because they do not know about a programme, cannot navigate the application process, or have been deterred by a previous negative interaction — represent a failure of the agency’s core mandate that is invisible in most performance dashboards. Fraud prevention rates are tracked. Overpayment recovery yields are tracked. The rate at which entitled citizens are not receiving entitled support is, in most agencies, unmeasured.
Fraud-oriented AI does not help here. A model trained to identify suspicious application behaviour has no capability to identify the entitled individual who has not applied. Worse, if that model generates false positives that create friction in the application journey — requests for additional documentation, verification delays, referrals for investigation — it actively suppresses take-up among the legitimate claimants who most need to reach the system.
The agencies that have recognised the take-up problem as a financial and human priority — and have invested in proactive outreach models that identify entitled individuals who are not claiming — have not built a separate AI programme alongside their fraud detection capability. They have recognised that identifying non-claimants and identifying circumstance changes are the same underlying problem. Both require knowing, in as close to real time as possible, who is entitled to what based on their actual current circumstances.
One Problem, Three Symptoms
This is the core insight that should reorganise AI strategy in benefits administration: fraud, overpayments, and benefit take-up gaps are not three separate problems requiring three separate AI investments. They are three symptoms of a single underlying failure — the agency does not have real-time visibility into recipient circumstances.
The recipient who is concealing earnings is a circumstance change the agency has not detected. The recipient whose overpayment has been accumulating for eighteen months is a circumstance change the agency detected too late. The entitled individual who is not claiming is a circumstance change — a job loss, a disability onset, a household change — that the agency has not identified at all. All three situations would be addressed by the same capability: real-time data matching across the sources that carry the relevant circumstance signals.
Most agencies are currently running separate programmes against all three. Fraud detection is owned by a counter-fraud team with its own models and data infrastructure. Overpayment prevention is a compliance function. Outreach to unclaimed entitlements is a communications or digital engagement function. These programmes share a common data need but do not share data, models, or even, in most cases, a common governance structure. The result is three separate investments solving three manifestations of one problem — at significantly greater total cost and significantly lower effectiveness than a unified capability would achieve.
The Recommendation
Replace the fraud frame with a circumstance change detection frame — and build the data and model infrastructure that frame requires.
In practice, this means three things.
First, invest in the cross-agency data matching infrastructure that makes real-time circumstance detection possible. Wage data from employment agencies. Housing and address data from local authorities. Financial data from relevant sources within legal frameworks. This infrastructure is not a fraud investment and it is not a compliance investment. It is the foundational capability that serves eligibility accuracy, fraud prevention, and take-up identification simultaneously. Agencies that frame it as a fraud investment will fund it at fraud-programme scale and deploy it against fraud objectives alone. That is a significant underinvestment in the capability’s actual potential.
Second, measure performance against all three outcomes. Fraud detection rates, overpayment recovery yields, and benefit take-up rates should appear together on the same dashboard — because they are all measures of the same underlying capability’s effectiveness. An agency that improves its fraud detection rate while its take-up rate falls has not improved its performance. It has traded one failure mode for another.
Third, evaluate re-determination processes as the highest-priority AI application in the near term. Event-driven re-determination — triggered by a detected circumstance change rather than a calendar date — is the single intervention with the widest financial impact across all three problem dimensions. It reduces overpayment accumulation. It identifies the earnings concealment that would eventually become a fraud case. And it creates the data touchpoints that enable proactive outreach to recipients whose circumstances have changed in ways that open new entitlements.
The agencies that build this capability will reduce their improper payment rates. They will improve their fraud detection yields. And they will reach entitled citizens who their current systems are missing entirely. Not because they invested more — but because they stopped treating one problem as three.
