A fraud model should route a claim, never refuse it
Chinese insurers now score motor and livestock claims for fraud with machine learning. A score should send a claim to the right reviewer with its reasons attached. When it holds or declines a claim on its own, its mistakes fall on the policyholders least able to wait.
Claims fraud is not imaginary. In livestock insurance we see ear-tag photographs reused across claims and carcasses presented twice. In motor insurance we see staged collisions and repair estimates inflated with a workshop's help. Every yuan paid on a fraudulent claim is eventually paid by honest policyholders and, in subsidised agricultural cover, by the governments that share the premium. An insurer that does not look for fraud is not being kind. It is being careless with other people's money.
Machine-learning models have made the search more effective. They find patterns across thousands of claims that no individual investigator would notice: the same phone used to photograph losses in three counties, a repair shop whose estimates rise when a particular assessor is on duty. I have built such models at two insurers, and I would not run a claims operation without them.
My argument concerns what the model is permitted to do. A fraud score should decide which reviewer sees a claim, how soon and with what questions. It should never, by itself, hold a claim for investigation or decline it. The line sounds procedural. In practice it decides who pays for the model's errors.
Where the false positives land
In October 2021 Hesui put a sow-claims model into production that automatically held claims scoring above a threshold. I approved the release. The model performed well on the measures we had chosen: it found more confirmed fraud per investigator-hour than the rules it replaced.
Nine months later my data team broke its results down by farm size. Claims from farms with fewer than 50 sows were held at 14 per cent; claims from farms with more than 500 at 4.1 per cent. Of the small-farm claims held, 61 per cent were paid in full after review, on average 26 days later. A household that had lost one of its twenty sows waited almost a month longer than a large farm with a similar loss, and in most cases for nothing.
The cause was not malice in the model. A sow claim needs ear-tag identification and evidence that the carcass was disposed of safely. Large farms record disposal animal by animal. Many small farms take carcasses to village collection points that, at that time, issued batch records without individual tag numbers. The model had learnt that a missing link between tag and disposal record was associated with fraud, which was true in the historical data, and it applied that lesson to households whose records were incomplete for reasons that had nothing to do with honesty.
This is the general pattern. Fraud models learn from past investigations and the records around them. Records are thinner where policyholders are smaller, poorer, older or more rural, so a model trained on them will flag those groups more often. The insurer bears the cost of a false positive in reviewer minutes. The policyholder bears it in weeks.
Reasons a reviewer can read
In August 2022 I recommended removing the model's power to hold claims. It would instead route each claim to a reviewer, with the three or four features that most influenced its score written in words a surveyor could check: a batch disposal record without a tag list; photograph metadata matching a claim in another county; a claim reported within four days of the policy starting. Our head of anti-fraud estimated that the change would let through about RMB 38 million a year in fraudulent claims, and our CFO agreed with him. The president approved a six-month trial with monthly reporting.
Readable reasons changed the reviewers' work more than I expected. A reviewer who sees that a flag rests on a batch disposal record can telephone the collection point and resolve it in a day. A reviewer who sees only a score of 0.83 has to start an investigation from nothing. We also worked with 46 county disposal centres to add tag lists to batch records, which removed the largest source of false flags at its root.
Confirmed fraud recoveries rose from RMB 44 million in 2021 to RMB 51 million in 2024. The leakage our anti-fraud head feared did not appear in our results, though I cannot prove that none occurred. What I can say is that reviewers spent their time on claims with better reasons attached, and found more.
The same principle applies to newer tools. In 2025 we tested a large language model for summarising claim files for reviewers. Its summaries were fluent and usually accurate, but in 3 of 200 test files it cited a disposal certificate number that did not exist. A summary that invents evidence is worse than no summary. The model now drafts summaries in which every statement links to the page it came from, and a reviewer signs each one. It saves time. It does not decide anything.
Fairness tests as release criteria
The lesson I took from the sow model is that fairness tests have to be release criteria, not audits conducted afterwards. Before any claims model goes into production at Hesui, we compare flag rates and post-review outcomes across groups: farm-size bands, counties, sales channels and, in motor, the size and location of the repair workshop. If one group is flagged more than one and a half times as often as another, the release waits until we can either explain the difference with evidence of actual fraud or remove its cause. The results go to the board's Risk Management and Consumer Rights Protection Committee twice a year.
The threshold is a judgement, not a law of nature, and there are legitimate reasons for flag rates to differ. Some counties do have more organised fraud. The test does not forbid a difference; it obliges us to justify one. In two releases since 2023 the justification held. In three it did not, and those releases waited.
The tests also have limits. We can compare only groups we can identify lawfully and reliably, and we choose not to process some personal characteristics at all. A model can be fair across every group we measure and unfair across one we do not. That is an argument for keeping a person between the score and the decision, not for abandoning the tests.
By 2024 the ratio between small-farm and large-farm flag rates on sow claims was 1.3. Median payment for small-farm sow claims fell from 31 days to 12, and livestock complaints fell 44 per cent. None of this required a better algorithm. It required changing what the algorithm was allowed to do, and measuring who carried its mistakes.