Industries / Finance and business services / Insurance
Paying claims
Checking a reported loss against the policy and paying it, from a single broken window to a city flattened by a storm.
5
Milestones achieved, from 1906 to 2021.
1
To watch for: announced, not yet confirmed outside.
2
Open deliverables. Nobody has shipped them.
The road
AchievedWatch forOpen
Achieved
1906
Pay in full
An earthquake and three days of fire destroyed much of San Francisco from 18 April 1906. Earthquake exclusions were already standard, many insurers could not tell whether shaking or fire caused a loss, and at least 12 American insurers went bankrupt; of about $235 million insured, about $180 million was paid, by Insurance Journal's account. The Lloyd's underwriter Cuthbert Heath told his San Francisco agent to "pay all of our policyholders in full, irrespective of the terms of their policies", which cost Lloyd's over $50 million by its own account.
LessonWhen the cause of a loss cannot be proven, the policy wording decides who is paid, and an insurer that paid anyway won a market.
Credit
Cuthbert Heath, Lloyd's
2007
Wind or water
Hurricane Katrina in 2005 hit many homes with both high winds and flooding, which are often covered by different policies. The US Government Accountability Office (GAO) reported that disputes after the 2005 season showed how hard it is to tell how much damage each caused when a property is hit by both.
LessonWhen two policies cover one storm, someone has to divide the damage between them, often after the evidence is gone.
Credit
GAO, report GAO-08-28, December 2007
2010
Paid by measurement
After the earthquake of 12 January 2010, the Caribbean Catastrophe Risk Insurance Facility (CCRIF) paid the government of Haiti US$7,753,579, about 20 times its premium of US$385,500. CCRIF pays on parametric triggers: the measured strength and location of an event, not an assessment of damage on the ground. It decided on the full payout in less than 24 hours; the policy required a 14-day waiting period.
Credit
CCRIF
2016
When the model is wrong
Malawi's drought policy with the African Risk Capacity (ARC) at first did not pay for the 2015/16 season, because ARC's model showed few people affected while the government's estimate was much higher. The model assumed a maize crop of 120 to 140 days, but farmers had largely switched to a 90-day variety. Once the model was corrected, the policy paid US$8 million.
Lesson"Any model's ability to represent reality depends on the accuracy of starting assumptions and data," said ARC's chair, Lars Thunell.
Credit
African Risk Capacity
2021
Pandemic cover tested in court
The FCA brought a test case on business interruption policies after COVID-19 closures. On 15 January 2021 the UK Supreme Court ruled largely for policyholders: cover may apply to partial closures, and valid claims should not be cut because the pandemic would have caused the loss anyway. The FCA had identified about 370,000 policyholders holding 700 types of policy from 60 insurers who might be affected.
LessonOne court case can settle the meaning of a clause for hundreds of thousands of claims at once.
Credit
UK Financial Conduct Authority; UK Supreme Court
Watch for
Agents settle small claims
Allianz says Project Nemo, live in Australia since July 2025, uses seven AI agents (planner, cyber, coverage, weather, fraud, payout and audit) to handle food spoilage claims under AUD 500 after storms and power cuts. It reports an 80% cut in processing and settlement time, from several days to about a day or a few hours, and says a person makes every payout decision.
Achieved whenan outside party, such as a regulator or auditor, reports processing times and error or complaint rates for these claims before and after.
Claimed by
Allianz, November 2025
Open
01
Car damage priced from photos
A public answer key for repair estimates: damaged cars photographed to a fixed shot list, each paired with the final itemized invoice from the body shop that fixed it. Anyone can then score a photo-estimating AI in dollars.
The briefCredit · unclaimed02
A disaster claims tracker
One public table per major storm or wildfire that combines what each state regulator publishes on claims: number reported, closed with payment, closed without payment and the reasons, by line of insurance. It updates whenever a regulator posts new figures and links every number to its file.
The briefCredit · unclaimed
Open 01 · the brief
Car damage priced from photos
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
A public answer key for repair estimates: damaged cars photographed to a fixed shot list, each paired with the final itemized invoice from the body shop that fixed it. Anyone can then score a photo-estimating AI in dollars.
Why it is still open
The best-known public set of car damage photos, CarDD (2023), labels 4,000 images with six kinds of damage for detection and has no repair costs. A search for a public set that pairs photos with repair invoices found none.
Built with what exists
phone cameras; invoices shared with each owner's written consent, with plates, names and addresses removed; vision-capable AI models such as Claude, Gemini and GPT; CarDD for trial runs.
First step
Agree with one body shop to photograph 20 incoming cars to a fixed shot list and keep their final invoices.
Done looks like
300 cars, each with six photos and a final invoice. For at least three AI models, the median error in dollars and the share of estimates within 20% of the invoice are published.
Open 02 · the brief
A disaster claims tracker
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
One public table per major storm or wildfire that combines what each state regulator publishes on claims: number reported, closed with payment, closed without payment and the reasons, by line of insurance. It updates whenever a regulator posts new figures and links every number to its file.
Why it is still open
Florida's regulator publishes Hurricane Milton totals by line and county but not by insurer, keeps each storm on its own page, and states that the data "has not been audited or independently verified". As of 9 December 2025 it showed 134,177 of 385,146 Milton claims closed without payment.
Built with what exists
regulators' published spreadsheets and reports, starting with the Florida Office of Insurance Regulation; Python to read them; a language model to pull numbers from PDF reports, with a person checking each one.
First step
Put Florida's pages for Hurricanes Ian, Helene and Milton into one table with the same columns.
Done looks like
Ten events since 2017 from at least two states, every number traced to a regulator file, and closed-without-payment rates shown on the same basis across events, updated monthly for six months.