Where an agricultural insurer's technology decides real money

Remote sensing is now part of agricultural claims handling at most Chinese crop insurers, and rightly so: optical and radar imagery can map planted area, flooding, lodging and harvest progress across thousands of square kilometres in days. I treat it as a way to stratify and prioritise, not as a measurement of loss in itself. At Hesui, imagery assigns each insured plot to a stratum; drone flights and witnessed sample plots, measured by our surveyors with a village representative and a county agriculture technician present, set the loss rate for each stratum. Grain samples are tested where the damage is in quality rather than in the standing crop, which no satellite can see. In June 2024 this method covered 2.4 million mu of wheat across seventeen counties with 1,840 sample plots. Appeals ran at 1.8 per cent of households, and 39 per cent of them led to a higher payment.

Subsidised crop and livestock insurance in China is financed partly by premium subsidies from several tiers of government, and the insurer must account for cover and claims household by household and post them publicly in each village. Hesui's old system recorded group policies by village, with household lists attached as spreadsheets. From 2021 to 2024 I led a RMB 186 million programme to replace it with a ledger built around the household and each plot it farms, linked to land-contract records where counties make them available. The ledger now holds 4.6 million households and 38 million plots. It produces the village notices directly, detects plots insured twice, and lets us reconcile subsidies with county finance bureaus in 19 days rather than 74. I migrated it province by province in the quiet months after autumn sowing, with a full season of parallel running in each.

Fraud in agricultural and motor claims is real: reused ear-tag photographs, staged collisions, repair-shop inflation. Models help find it. But a model trained on historical decisions learns where records were incomplete as readily as where fraud occurred, and incomplete records cluster among small farms and rural repair shops. At Hesui our models route claims to reviewers with readable reasons; none holds or declines a claim on its own. Before any release we compare flag rates and post-review outcomes by farm size, county and channel, and a ratio above 1.5 between groups must be explained or fixed. In 2025 we tested a large language model for summarising claim files. It drafted useful summaries but invented a document reference in 3 of 200 test files, so it now drafts only with citations to source pages, and a reviewer signs every summary.

At Yunzhu Online I ran systems that issued over a billion policies a year at premiums often below two yuan: return-shipping, delivery-delay and screen-damage cover, embedded in a platform's checkout. The engineering problems are concrete. Issuance must answer within the platform's time limit or the sale is lost, and the busiest shopping festival can bring ten times an ordinary day's volume. Every policy, however small, still needs a stored wording version, a record the customer can retrieve and a claims route that works without a call centre. We moved 23 partner integrations from nightly batches to real-time issuance, settled 92 per cent of return-shipping claims automatically from carrier tracking data and sustained 41,000 policies a second at peak. Hesui's online book is far smaller, about RMB 600 million of premium in 2025, but it follows the same rules.

Agricultural claims involve identity numbers, bank accounts, land-contract records, photographs of homes and livestock, and medical information in accident cover. The Data Security Law and the Personal Information Protection Law, both in force since 2021, require us to classify data, collect no more personal information than a purpose needs, obtain proper consent and control what processors do on our behalf. In 2022 I found that a drone-survey contractor kept copies of household lists, with identity numbers, on its own servers. We ended the arrangement mid-season and moved imagery processing into a wholly owned subsidiary working inside our environment, at an extra cost of RMB 14 million that year. Hesui has since classified 1,140 data sets by sensitivity, and no contractor now receives an identity number where a household reference will do. I report data-protection exceptions to the board committee each quarter.