You called the claim fraud. If you were wrong, that is bad faith.
Fraud is real, organised, and expensive — it inflates premiums for everyone. But a fraud finding is also an accusation against a policyholder, and a wrong one is a bad-faith repudiation. The special-investigations desk now carries both the loss from the fraud it misses and the legal exposure from the honest claim it wrongly accuses.
The scale of the problem is documented and rising. Kenya's regulator recorded a near-threefold increase in reported insurance fraud in a single year, with fake motor-insurance certificates and forged documents the most common forms, and the industry estimates that fraud inflates premiums by as much as a quarter. The public health insurer lost an estimated 11 billion shillings to fraudulent claims over a six-month period, much of it fake billing and phantom treatments from colluding providers. Motor and medical are the most exposed classes, and organised cartels rather than opportunistic individuals are increasingly the source.
Detecting it is a natural job for a model, and the insurers know it. A fraud model reads a claim against patterns — the staged-accident signature, the provider that bills impossible volumes, the multiple policies on a single vehicle — and flags the ones that look wrong. The capability is genuine and the loss it can prevent is large. The danger sits on the other side of the flag, and it is specific to fraud in a way it is not to ordinary claims handling.
Calling a claim fraudulent is the strongest adverse thing an insurer can say about a customer, and the law treats it accordingly. A fraud finding that turns out to be wrong is not a mere claims error; it is a bad-faith repudiation, defamatory in tendency, and in the conduct framework a serious breach. The special-investigations desk therefore faces an asymmetry sharper than the rest of the claims function: clear the fraud and the insurer eats the loss and the inflated premiums; accuse the honest claimant and the insurer faces a dispute, a conduct complaint, and a reputational and legal exposure that the original claim value does not begin to measure.
The model makes both errors easy to commit silently. A fraud score with no reasoning behind it gives the investigator no way to tell a strong case from a coincidence of risk markers, and a great many honest claims carry the same surface features as fraudulent ones — the late-night accident, the recently increased cover, the claim soon after inception. An investigator who declines a claim on an unexplained fraud score has made an accusation they cannot substantiate, and when the claimant disputes it, the basis has to be reconstructed after the fact, which is exactly the position a bad-faith allegation cannot be defended from.
There is a second problem the industry has long recognised: the strongest fraud signals are cross-insurer. The staged-accident ring, the serial claimant, the colluding garage or clinic move between insurers, and no single carrier sees the whole pattern. The obvious answer is a shared fraud database, and the obvious obstacle is data protection: pooling claims data across insurers means exposing identifiable policyholders to competitors, which the data-protection regime constrains. The pattern is shared across the market; the identities cannot be.
The two largest fraud classes show why reasoning matters as much as detection. In motor, fake insurance certificates and staged accidents dominate — the regulator recorded fake-certificate cases rising from a handful to twenty in a single year — and the markers of a staged accident overlap heavily with those of a genuine one in a high-risk transport sector. In health, the fraud is provider collusion: phantom treatments and inflated billing of the kind that cost the public health insurer billions in a matter of months, where the accusation falls not only on a claimant but on a hospital with the standing and the motive to fight back. In both classes, a flag is the start of a case, not the end of one.
A fraud flag the investigator cannot substantiate is not loss prevention. It is a bad-faith accusation the model made on the insurer's behalf.
Where each sits.
Akki ingests the signals a fraud decision rests on — the claim, the policy history, the provider record, the device and incident data — as a governed substrate, and logs every input to a fraud disposition. When a fraud finding is disputed, the basis can be reconstructed exactly rather than approximately, which is the difference between defending an accusation and retracting it.
Solva structures the fraud reasoning and refuses to assert fraud on thin or circumstantial signal, surfacing what would be needed to substantiate the finding rather than letting a score become an accusation. Underneath each fraud disposition sits the basis — the markers weighed, the reasoning, the confidence — which is precisely what a bad-faith challenge or a conduct review demands. The refusal to call fraud the insurer cannot prove is the integrity, and here it is also the defence against a defamation and bad-faith claim.
This is one of SyniSense's strong homes. When insurers need to detect the cross-insurer patterns — the ring, the serial claimant, the colluding provider — SyniSense anonymises policyholder and claimant identity at the perimeter so the shared pattern can be found across the market without any insurer exposing identifiable customers to its competitors. The fraud is caught in the pooled pattern; the honest policyholder's identity never leaves the carrier that holds it.
For the special-investigations lead, the fraud finding carries its own substantiation. An accusation is made only where the evidence supports it and is recorded when it is, so the desk no longer accumulates undefended fraud allegations that a single disputed case could turn into a bad-faith and defamation exposure.
For the honest policyholder wrongly caught by a risk-marker coincidence, the protection is structural: the model cannot turn a thin score into a repudiation, because the reasoning that would be required is surfaced as absent rather than assumed as present.
For the provider relationship in health insurance, the discipline matters commercially as much as legally. Accusing a hospital of fraudulent billing on a thin model signal damages a provider network the insurer depends on to serve its members; a finding the insurer can substantiate, set against one it cannot, is the difference between cleaning the network and losing it.
For the market, cross-insurer fraud detection becomes possible without the data-protection obstacle that has held it back. Carriers can pool the patterns that catch organised fraud while keeping their policyholders' identities inside their own perimeters, which is the only basis on which a shared fraud capability is lawful.
For the conduct regulator, the insurer presents a fraud process that is evidenced and proportionate — catching real fraud while refusing to convert honest claims into accusations. That is the posture that keeps a fraud-control programme from becoming a conduct problem of its own. A fraud function that catches the cartels while sparing the honest claimant is the one that lowers premiums without raising complaints — the outcome the regulator and the policyholder want from it alike.