AI for Fraud Review in Government

An officer opens the next item in the queue. The payment has already gone out. The applicant's address is shared with several others in the same round, the bank details resemble an account used on a different scheme two rounds ago, and the supporting evidence is a photograph of a letter from a body nobody on the team has heard of. Nothing here says fraud. It also does not say error, and the difference decides whether the next contact with this person is a request for information or the opening of an investigation.

This is fraud and error work inside a public body: grants, benefits, subsidies, licences and reliefs, and it is not commercial fraud review. The counterparty holds a legal entitlement, the decision is appealable, and wrongly withholding money from somebody owed it is a serious failure in itself, landing on the person least able to absorb it. Scheme rules differ by jurisdiction; what follows is the shape of the work, not the law.

Fraud and error are not the same finding

A commercial team can treat loss as loss. A public body cannot: the two findings carry different powers, different recovery rules and different consequences for the person. Error splits again, into claimant error and official error, and only one is usually recoverable in full.

What separates them is intent, which is almost never in the data. A discrepancy does not show whether the applicant knew, or whether the office applied an update it received. An agent can characterise the discrepancy, set out the explanations consistent with it, and say which evidence would separate them. It should never call a case fraud itself.

The line the agent does not cross

The agent gathers, matches, evidences, drafts and flags. It does not decide eligibility, it does not suspend or reduce a payment, and it does not score a person.

An entitlement decision is a legal act by an officer who has to explain it in their own words to a tribunal. A number produced by a model is not a reason. Suspension is the most harmful thing the department can do on a suspicion, so it needs somebody accountable for having formed it. Build the pipeline so the agent cannot trigger any of those. Not policy, plumbing.

Verification at the point of application

Evidence arrives as scans, phone photographs, and self declarations that are assertions until something corroborates them. The question is not completeness but whether the claim is consistent with everything else in view. An agent does the mechanical part: extract the declared facts, check them against each other and against what the body already holds, and list what is missing or unreadable.

Absence of a document is not evidence of ineligibility, and the applicants least able to produce clean paperwork are often the most clearly entitled. An agent that treats an unreadable scan as an absent one has changed the applicant's position without anyone deciding to.

Matching registers that were never built to agree

All of this rests on joining records held by systems designed separately, for different purposes, and never reconciled. A match is therefore a hypothesis. Two records sharing a name and an address may be one person, two people with a common name in a converted building, or the same person before and after a move one register missed.

So the agent's job is to make match strength visible: which fields matched, which did not, how stale each source is, and what an innocent explanation would look like. A pipeline that emits matches without that context manufactures suspicion at volume.

Duplicates, related applicants and the entity that exists only on paper

Schemes repeat, so the same applicant appears across rounds, across related schemes, and occasionally twice in a round under variations of a name. Clustering near duplicate applications is worthwhile because coordinated submissions share text, but identical wording is also what you get when an adviser or a support service helps people apply. Textual similarity is a prompt to look, never a finding.

Organisations bring their own version: an entity registered shortly before the scheme opened, no trading history, an address shared with other applicants, an account opened for the purpose. What an agent assembles here is a description, and whether it describes a shell or a genuinely new business is judgement.

The risk of opening a review on someone in need

A review opened on a household with no margin does damage even when it ends with no finding. The correspondence frightens people, the evidence gathering is work they may not be able to do, and the fear of losing the payment changes behaviour.

Design decisions reduce that. Keep paying while the review runs unless a human decides otherwise. Write the opening contact as a request, not an accusation. Let the claimant explain before the case is characterised, because the explanation often dissolves the discrepancy in a sentence. Require corroboration from an independent source instead of letting weak signals stack into an apparently strong one. And watch who is actually being reviewed.

Assurance after the money has gone out

Post payment assurance has a property that pre payment work does not: the population is everyone who was paid. Nobody can review that, so the body samples, and the sample serves two purposes that get confused. One is measurement, producing the loss estimate the auditor relies on. The other is targeting, finding cases worth recovering.

They need different selection. An agent helping with targeting must not also select the measurement sample, or the estimate reflects what the agent looks for rather than what is there. Within that constraint it is strong, applying defined conditions across the whole paid population and preparing each sampled case with its evidence referenced.

The referral pack that has to survive an appeal

Whatever leaves this team may be read by an appeal officer, a tribunal, an auditor or a prosecutor, years later, by somebody hostile to it. That reader tests provenance before argument. Every element has to carry which system it came from, when it was retrieved, and which version of the scheme rules was in force on the date of the claim rather than today.

The agent drafts the pack: chronology, evidence with sources, and the material pointing the other way. Packs containing only what supports the finding are the ones that collapse. Keep what the agent proposed alongside what the officer changed: that override history is the strongest evidence that human control was real.

What the claimant is owed when the decision goes against them

A person told they were overpaid is entitled to know what was held against them, in plain language, with the facts relied on and what they can do next.

An agent drafts these well, which is where the danger sits: fluent letters read as reasoned whether or not the underlying match was ever checked. Constrain the drafting so it can only cite items present in the pack, and make the officer confirm each cited fact rather than the letter as a whole. If a fact cannot be put to the claimant in a sentence they could dispute, it is not yet evidence.

Recovery, debt referral and what is not worth chasing

Recovery starts with arithmetic: the period, the correct entitlement across it, what was paid, and how the cause affects what is recoverable. The agent can show its workings line by line, which matters because the sum is itself appealable.

Beyond the number is judgement the agent should only inform. Whether pursuit is proportionate when the balance is small. Whether deduction from an ongoing payment leaves the household able to eat. Whether the debt should be referred, paused or written off. An agent can surface hardship indicators the body already holds and draft the notice. It must never set a deduction rate.

Tip offs and the free text they arrive in

Public reports arrive constantly, in free text, mostly from neighbours, former partners and disgruntled colleagues. Most are not fraud. Some are malicious, and a body that acts on them uncritically becomes an instrument in somebody else's dispute. A minority are the best intelligence the team gets.

The agent turns them into something workable: who is alleged to have done what, under which scheme, over what period, and whether this reporter has sent it before. It then routes, and routes most of them to no action. A report on its own does not open a case.

Where AI fails here

Matching errors are not evenly distributed. They land on people with common names, unstable addresses, recent arrivals, shared accommodation and interrupted records, which is close to a description of the population most likely to be entitled and least equipped to challenge a wrong decision.

Fluency is the second trap. These systems write a confident narrative from thin material, and that narrative reads like evidence to a busy officer and to an appeal panel.

Third, an agent built on past cases reproduces past targeting. If certain neighbourhoods, surnames or scheme populations were reviewed more readily for reasons the department would not defend aloud, the agent does that faithfully and consistently.

Fourth, it only knows the fraud you caught, and organised abuse adapts to whatever is checked. Official error is close to invisible for the same reason: when the body's own record is wrong, matching against it confirms the mistake instead of finding it.

Finally, the quiet one. Automation raises the number of reviews a small team can open. Nothing raises the claimant's capacity to respond, and appeal capacity does not grow to match. Scaling the front of this process without scaling the defences around it is how a technically successful project produces real harm.

Where to begin when the money is public

Start where the output is evidence rather than judgement: assembly and provenance on a single scheme whose rules are written down, with an officer accepting or rejecting everything. Matching comes second, once its false positives have been examined on your own data. Decision letters come last.

Evaluate against cases you have already closed, and make sure the set includes those overturned on appeal. They are the most valuable evaluation data the body owns. Watch the direction of disagreement: an agent wanting to open cases you closed costs officer time, and an agent wanting to open cases a tribunal already told you were wrong is the one that hurts people.

Adopt AI Partners builds these systems inside the body's own case and payment systems, and stays until the counter fraud team runs them without us. To talk through where this would help first, describe your scheme to Adopt AI Partners at hello@adoptaipartners.com.