Choices that made claims slower or costlier before they made them fairer

Pairing satellite strata with witnessed sample plots after harvest-week rain

In June 2024 eight days of rain fell across seventeen of our wheat counties in Henan and Shandong just as the crop was ready to cut. Within a fortnight 351,000 insured households had reported losses on 2.4 million mu. Much of the damage was sprouting and mould in the grain, which lowers its value without flattening the field. Households needed cash to plant maize within weeks. County governments sharing the premium expected prompt settlement, and after the 2021 floods our agricultural claims had taken a median of 63 days to pay.

Our head of claims proposed paying on plot-level loss estimates from satellite imagery alone, which our analysts said could be done in 18 days. Two branch general managers wanted conventional surveys, village by village. I supported neither. Imagery could identify lodged, waterlogged and unharvested fields, but it could not see sprouting in grain already cut, which was the larger part of the loss. Paying on imagery alone would have underpaid households whose fields looked fine from orbit. I proposed using imagery to divide plots into strata, then setting loss rates from witnessed sample plots and grain tests. The head of claims argued that this added two weeks and invited disputes. It did both.

Drone teams from our survey subsidiary and two contractors measured 1,840 sample plots, each with a village representative and a county agriculture technician present, and grain samples were tested for germination and weight. Every household received its stratum, its loss rate and an appeal route through its village coordinator.

We paid RMB 714 million with a median of 29 days from report to payment. About 6,300 households appealed, 1.8 per cent of the total, and 39 per cent of appeals led to higher payments, mostly in two strata where a river terrace drained differently from the land our model had grouped it with. Survey effort was about 7,600 staff-days, against an estimated 20,000 for a full conventional survey. The method is now written for each crop and peril and agreed with county bureaus before the season.

I would do two things differently. I would publish the method before the harvest instead of explaining it afterwards. And I would train more of our own drone pilots; contractors flew 40 per cent of the plots, and their footage was the hardest to audit.

Taking the sow-claims fraud model out of the decision to hold a claim

Hesui insures about 3.1 million breeding sows. A sow claim needs ear-tag identification and evidence that the carcass was disposed of safely. In October 2021 I approved the release of a machine-learning model that automatically held claims scoring above a threshold for investigation. In mid-2022 my data team broke down its first nine months by farm size. Claims from farms with fewer than 50 sows were held at 14 per cent, against 4.1 per cent for farms with more than 500. Of the small-farm claims held, 61 per cent were paid in full after review, on average 26 days later.

The cause lay in the records. Many small farms took carcasses to village collection points that issued batch disposal records without individual tag numbers, and the model had learnt to read the missing link as a fraud signal. Households that had lost one of twenty sows were waiting a month for nothing.

I recommended removing the model's power to hold any claim, using it only to route claims to reviewers with stated reasons, and making fairness tests a condition of every release. Our head of anti-fraud estimated that the change would let through about RMB 38 million of fraudulent claims a year, and our CFO supported him. The president approved a six-month trial with monthly reporting, and I agreed that we would revisit the decision if leakage appeared.

We rewrote the output as three or four plain reasons a surveyor could check, worked with 46 county disposal centres to add tag lists to batch records and retrained the model. No group may now be flagged 1.5 times as often as another without evidence.

By 2024 the ratio between small and large farms was 1.3. Median payment for small-farm sow claims fell from 31 days to 12, and livestock complaints fell 44 per cent. Confirmed fraud recoveries, RMB 44 million in 2021, were RMB 51 million in 2024, because reviewers spent their time on better-reasoned flags. Fairness results now go to the board's risk committee twice a year.

I would not release a claims model again without the group comparison; the review that found the problem used data we already held. And I would ask surveyors, who knew about the batch records, before the data scientists chose their features.

Deferring the motor claims platform to rebuild the agricultural ledger

When I joined Hesui in 2021, its agricultural business ran on a system built in 2009. It recorded group policies by village, with household lists attached as spreadsheets. Reconciling premium subsidies with county finance bureaus took 74 days, plots insured twice were found only by chance, and the village notices that list insured households and their claims were compiled by hand. The July 2021 floods showed how slowly such a system pays.

The budget for 2022 allowed one major programme. The head of our motor business wanted it for a new motor claims platform. The 2020 reform of motor insurance pricing had squeezed margins across the market, competitors were releasing faster claims and quotation tools, and he expected to lose renewals without one. I recommended the agricultural ledger first and the motor platform twelve months later. Agricultural cover was 38 per cent of premium, was financed partly with public money and was where our records were weakest. The executive committee divided. The president chose the ledger and committed to funding motor in 2023.

We built the ledger in-house for RMB 186 million over three years. Its core is the household and each plot it farms, linked to land-contract records where counties make them available. We migrated three provinces between November 2022 and February 2023 and six more a year later, always after autumn sowing and before spring underwriting, with one season of parallel running in each.

The ledger now holds 4.6 million households and 38 million plots. Subsidy reconciliation takes 19 days. In 2023 it found that 1.6 per cent of our insured area was insured twice, under different family members' names or with another insurer as well; we cancelled our duplicate cover and returned the premium, subsidy share included. Village notices are now produced from the ledger itself.

The cost fell on motor. Renewal rates in two branches slipped by about two percentage points in 2023, and the motor platform will not be complete until 2027. The head of motor was right about the effect, and I said so to the executive committee.

I underestimated the work of matching households to land-contract data, which consumed about 40 per cent of the effort. I would also bring county finance bureaus into the data design at the start, rather than at testing, when their reporting formats forced late changes.

Declining shoppers' return histories when pricing return-shipping cover

In 2019 return-shipping cover was the largest line at Yunzhu Online Property Insurance: about 870 million policies that year, at an average premium below one yuan, sold at the checkout of e-commerce platforms. The loss ratio was 71 per cent and rising as shoppers returned more goods. A large platform partner offered us a feed of each buyer's return history, so that we could price frequent and infrequent returners separately. Our head of pricing estimated a loss-ratio improvement of about 4.5 points, and competitors were said to price this way already.

I argued against accepting it. The Cybersecurity Law already required personal information to be collected lawfully, properly and only as far as necessary, and a national personal-information protection law was being drafted. A buyer's whole return history, across every merchant, was more than we needed to insure one parcel's return. Customers could not see why one paid three times what another did, and many frequent returners were simply buying clothing from merchants whose sizes varied. The feed would also have put our pricing at the mercy of a partner who could change or withdraw it. The head of pricing disagreed, and our chief executive initially favoured accepting. After two months of analysis he agreed to price on merchant, category and order characteristics and to decline the feed.

We built merchant-level return and complaint scores, priced heavier and more frequently returned categories higher, and described the price basis on the purchase page. We also built detection for merchant collusion, in which related accounts bought and returned goods to collect the shipping allowance.

The loss ratio rose to 74.8 per cent in 2020, and in two clothing categories we lost share to a competitor pricing on buyer history. Merchant-collusion detection recovered RMB 23 million in 2020, from a risk that a buyer-focused model would have missed. When the Personal Information Protection Law took effect in November 2021, after I had left, former colleagues told me there had been no buyer-level pricing data to unwind.

I would build the collusion detection first, before arguing the case, because its results were the strongest evidence for the merchant-level approach. And I would have involved the partner's own compliance team at the start; their support later shortened a difficult negotiation over commission.