By the time most insurers go looking for fraud, the policy is already written. A claim comes in, an adjuster gets suspicious, and a special investigations unit starts working backward from a loss that is already on the books. That is fraud detection at the claims stage, and it is where the industry has spent most of its money and most of its technology. It is also the most expensive place to find the problem, because the risk was accepted, priced, and bound long before anyone asked whether it was genuine.
There is an earlier place to look. Fraud detection at underwriting scores the same signals while the risk is still being quoted, before a single policy is bound. The two are not rivals so much as two points on one timeline, and the difference between them is mostly a question of when the insurer pays for the miss. This article compares them, and shows where an external AI layer fits: on top of the systems an insurer already runs, never in their place.
Fraud detection at claims: finding it after the loss
Claims-stage fraud detection is the traditional model. The insurer writes the policy, the policyholder reports a loss, and the claim is examined for signs that it is exaggerated, staged, or entirely invented. Special investigations units, red-flag rules, and increasingly machine-learning models score each claim and route the suspicious ones for human review.
It works, and it is necessary. But it is reactive by design. The exposure is already on the book, the reserve is already set, and the investigation becomes a race to avoid paying a claim that should never have been priced in the first place. Detection at this stage limits the damage. It does not prevent it. An external AI layer here can score and flag a claim for the insurer's own investigators, but it does not settle claims or run the claims process end to end. That decision stays with the insurer.
Fraud detection at underwriting: finding it before the bind
Underwriting-stage fraud detection moves the same question upstream. As a submission arrives, the system reads it, enriches it against internal and external sources, and scores it for both risk and fraud signals before the insurer commits to a price. Misrepresented exposures, inconsistent histories, and identities that do not reconcile are caught while the insurer can still decline, refer, or reprice, rather than after it has taken the risk. This is why automating submission intake is usually where the earliest fraud signal appears.
This is the point in the lifecycle where an external AI layer is strongest, because reading and scoring the submission is exactly what the layer already does for underwriting. The fraud check is not a separate product bolted on at claims time. It is one output of the same risk engine that scores the submission against the insurer's appetite and underwriting manual. The enriched record and the audit trail it produces also carry forward, so if a claim does arrive later, the insurer's own investigators inherit a documented history instead of starting from a blank file.
Why the underwriting stage changes the economics
Fraud is not a rounding error in insurance. The Coalition Against Insurance Fraud estimated in 2022 that fraud costs the United States about $308.6 billion a year across all lines. The FBI estimates that non-health insurance fraud alone costs more than $40 billion a year. Separately, the FBI estimates that fraud adds $400 to $700 to the average U.S. family's annual insurance premiums, a general framing across insurance rather than a figure scoped to any one segment.
These are United States estimates, and they describe a problem that does not stop at a border. They also explain why the stage matters. Fraud that is not caught at underwriting does not disappear. It is priced into the book at a premium that assumed the risk was clean, and it resurfaces later as a paid or contested claim. In that sense it compounds on the loss ratio: the insurer collects too little premium for the true risk, then pays out more than the pricing assumed. Catching the same fraud at underwriting keeps it out of the book entirely, which is why the earlier check is the cheaper one.
Where an external AI layer sits
An external AI layer is software that sits on top of the insurer's existing core, reads and enriches each submission, scores it for risk and fraud against the insurer's own appetite, and returns an explainable decision through APIs. It is not the system of record and it does not replace the core. The policy administration system still issues the policy and holds the truth. The layer simply makes the risk and fraud judgment before the policy is bound, then writes the result back with a full audit trail.
For WIR, fraud detection is one function of the risk and fraud engine inside the underwriting journey, the same engine that scores each submission against the insurer's underwriting manual. Because the layer is external, there is no core migration and no load on the insurer's IT. Because every score is explainable and logged, a person can review why a submission was flagged, which matters as much for a declined applicant as it does for an auditor. WIR is not an insurer, a broker, or an MGA, and it does not carry risk. It automates the quotation and underwriting journey according to the insurer's own risk-acceptance policy.
What this means in Brazil: SUSEP and LGPD
WIR was born in Brazil, and two constraints shape any fraud automation there. The first is the Superintendência de Seguros Privados (SUSEP), the insurance regulator, which expects insurers to own and justify their pricing and acceptance decisions rather than defer them to a black box. The second is the LGPD, Brazil's general data protection law (Law 13.709 of 2018), whose Article 20 gives a person the right to request a review of decisions made solely by automated processing that affect their interests.
Both point to the same design rule. A fraud flag that changes whether someone is quoted, declined, or investigated has to be explainable and reviewable, traceable to the data and the logic behind it. That is why a well-built layer is designed to explain, to log, and to keep a person in the loop, not only to score. The Brazilian market treats fraud as a shared cost that honest policyholders ultimately pay for through higher premiums, and the industry, through bodies such as CNseg, coordinates prevention across carriers. The dollar figures above are United States estimates, but the underlying dynamic, fraud that is mispriced at underwriting and resurfaces as loss at claims, is the same in any market. Moving the check upstream, with an audit trail that a SUSEP review and an LGPD review can both stand behind, is what makes the automation defensible in Brazil.
The bottom line
Fraud detection at claims will always be part of insurance, because not every bad risk shows its hand at underwriting. But treating the claims stage as the only line of defense means paying to catch what better pricing could have kept out. Scoring fraud at underwriting, as one output of the same risk engine that already reads and prices the submission, moves the check to the cheapest point in the lifecycle. An external AI layer makes that practical, on top of the core the insurer already runs, never in its place, and with an explanation for every decision it returns.
Perguntas frequentes
What is the difference between AI fraud detection at underwriting and at claims?
At underwriting, AI scores fraud signals while a risk is being quoted, so misrepresented or fabricated exposures can be declined or repriced before a policy is bound. At claims, AI flags suspicious losses after they are reported, when the exposure already exists. Underwriting-stage detection keeps bad risk out of the book. Claims-stage detection limits the payout once the risk is already there.
Does an external AI layer replace the insurer's core system to detect fraud?
No. An external AI layer sits on top of the existing core and connects through APIs. It reads each submission, scores it for risk and fraud against the insurer's own appetite, and writes an explainable decision back to the policy system, which stays the system of record. There is no core migration, and the layer does not settle claims or carry risk.
Why is catching fraud at underwriting cheaper than catching it at claims?
Fraud that is not caught at underwriting is priced into the book as if the risk were clean, then resurfaces later as a paid or contested claim. It compounds on the loss ratio: too little premium collected for the true risk, and more paid out than the pricing assumed. Catching the same fraud at underwriting keeps it out of the book, so the earlier check avoids the loss instead of chasing it.
Is automated fraud scoring compatible with SUSEP and the LGPD in Brazil?
Yes, provided each decision is explainable and reviewable. Brazil's LGPD (Law 13.709 of 2018) gives individuals the right to request review of decisions made solely by automated processing, and SUSEP expects insurers to justify their acceptance and pricing. An external layer that logs its reasoning, returns an audit trail, and keeps a person in the loop is designed to meet both.