Industries / Finance and business services / Insurance / InsurTech
Automating claims
Software that takes in a claim, estimates the damage, decides what is owed and pays it, from repair-cost databases to AI that reads photos.
6
Milestones achieved, from 1966 to 2024.
2
To watch for: announced, not yet confirmed outside.
2
Open deliverables. Nobody has shipped them.
The road
AchievedWatch forOpen
Achieved
1966
Repairs priced by computer
Audatex was founded in Minden, Germany, in 1966. It describes its system for estimating car repairs as "the world's first automated vehicle repair system".
Credit
Audatex, by its own account
2006
One price list for buildings
Xactware, founded in 1986, makes software for estimating repairs to buildings. When the Insurance Services Office (ISO), now Verisk, bought it in August 2006, its customers included 16 of the top 20 US property insurers and "approximately 80 percent of insurance repair contractors and service providers", according to Insurance Journal.
LessonOne vendor's price data became the shared measure that insurers and contractors both work from.
Credit
Xactware; ISO
2016
A claim paid in seconds
Lemonade says its claims bot, AI Jim, approved a claim for a stolen parka in 3 seconds in December 2016, after running "18 anti-fraud algorithms". It then sent payment of $729, the coat's $979 cost less a $250 deductible. In its annual report for 2025 Lemonade says AI Jim takes the first report of a claim without human help 96% of the time, and that "roughly 55% of our claims were automated".
Credit
Lemonade, by its own account; Claims Journal reported the record as the insurer's claim, and no outside measure of these shares was found
2020
Estimates from photos
Tractable sells software that estimates car repairs from photos of the damage. In March 2020 Insurance Journal reported that the UK insurer Ageas was using it to produce complete estimates, which Tractable called "the first time that a UK insurer is using AI to generate end-to-end estimates". In May 2020 CollisionWeek reported Tokio Marine's use of it as "the first time that a major Japanese insurer has deployed" such a system. From 2021 GEICO used Tractable's AI to check estimates already written by people.
LessonThe largest US user had the AI check people's estimates rather than replace them.
Credit
Tractable (Adrien Cohen); Ageas; Tokio Marine; GEICO
2022
Touchless estimates stall
In November 2021 CCC Intelligent Solutions launched Estimate-STP, named for straight-through processing, a service that writes a repair estimate from photos with no person involved. Insurers can set it to need an adjuster's approval, and USAA was the first user. A year later CCC had 14 clients and said use was "still at a tiny fraction of the potential". Its rival Mitchell said automation was "not as close to complete as the industry once thought". In November 2025 CCC reported generating AI-based estimates through another tool, Mobile Jumpstart, at a rate of over 1 million a year.
LessonThe software could write estimates before insurers were ready to let it decide alone.
repairerdrivennews.comrepairerdrivennews.comrepairerdrivennews.com
Credit
CCC, by its own account for the counts; USAA
2024
Algorithms may not decide alone
In November 2023 patients and the estates of patients sued UnitedHealth, alleging it used naviHealth's nH Predict algorithm to end care after hospital stays, and that "over 90% of claim denials are reversed on appeal". The court's opinion records that UnitedHealth "denies any use of nH Predict". In February 2024 the US Centers for Medicare & Medicaid Services (CMS) told Medicare Advantage plans that "algorithms or artificial intelligence alone cannot be used as the basis to deny admission". It also said that a predicted length of stay alone cannot be used to end care. A US Senate subcommittee reported in October 2024 that UnitedHealth's rate of refusing nursing-home care after a hospital stay rose from 1.4% in 2019 to 12.6% in 2022. In February 2025 the court let the patients' contract claims proceed.
LessonA decision about one person's care needs a person who looks at that case, and CMS now says so.
Credit
CMS; US Senate Permanent Subcommittee on Investigations
Watch for
Half of claims untouched
Travelers' chief executive Alan Schnitzer said more than half of its claims are eligible for straight-through processing, handled from start to finish without a person. He said "another 15% of all claims are processed with advanced digital tools", and that its "claim call center population is down by a third".
Achieved whenan outside source, such as a state market conduct exam, reports the share of claims settled without a person and how their outcomes compare.
Claimed by
Travelers, January 2026
AI in Medicare approvals
CMS's WISeR model has outside companies use tools "such as artificial intelligence" to review requests for some services in traditional Medicare before they are approved. It runs from 1 January 2026 to 31 December 2031 in New Jersey, Ohio, Oklahoma, Texas, Arizona and Washington. Denials must be reviewed by a licensed clinician, and the companies earn a share of the savings.
Achieved whenCMS's evaluator, the HHS Inspector General or the US Government Accountability Office publishes the model's denial and overturn rates.
Claimed by
CMS, July 2025
Open
01
Hail at an address, tested
An open tool that says whether hail of a given size fell at an address on a given date, from NOAA's public radar estimates of maximum hail size, the Multi-Radar Multi-Sensor (MRMS) system. It comes with a published test of how often it is right.
The briefCredit · unclaimed02
A register of claim algorithms
A public table of the algorithms named in court records, regulators' documents and insurers' own statements as used to decide or screen claims. Each entry gives the insurer, the kind of cover, what the tool is said to do, what the insurer says it does, and a link to each source.
The briefCredit · unclaimed
Open 01 · the brief
Hail at an address, tested
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
An open tool that says whether hail of a given size fell at an address on a given date, from NOAA's public radar estimates of maximum hail size, the Multi-Radar Multi-Sensor (MRMS) system. It comes with a published test of how often it is right.
Why it is still open
Commercial services exist. A public demo on Hugging Face reads NOAA's hail maps for an address, but it has no comparison with ground reports. A published check of radar hail estimates against "more than 10 000 storms" works at the level of storms, not addresses.
Built with what exists
NOAA MRMS data on the AWS open data registry, updated every 2 minutes; pyhail, an MIT-licensed Python library; NOAA's Storm Events Database as ground truth.
First step
Download one day of MRMS maximum-hail-size grids for a storm with known reports and check the tool's answer at ten reported places.
Done looks like
1,000 reported hail falls from 2020 to 2025 and 1,000 address-days with no report, run through the tool. The share of hits and of false alarms is published at three hail sizes. Code and data are released under open licenses.
Open 02 · the brief
A register of claim algorithms
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
A public table of the algorithms named in court records, regulators' documents and insurers' own statements as used to decide or screen claims. Each entry gives the insurer, the kind of cover, what the tool is said to do, what the insurer says it does, and a link to each source.
Why it is still open
In the cases found, insurers dispute what their tools do: UnitedHealth "denies any use of nH Predict", Cigna disputes that its PxDx process is AI, and in Huskey v. State Farm discovery about the screening tools was still under way in February 2026. KFF notes that public prior-authorization data give no breakdown of denials by reason.
litigationtracker.law.georgetown.eduinsurancenewsnet.comclearinghouse.netkff.org
Built with what exists
CourtListener and PACER for court filings; state regulators' AI bulletins and market conduct reports; a language model such as Claude to pull tool names and claims from filings, citing pages (see the Note "Seven more flagged").
First step
Read the published opinions in the nH Predict, PxDx and Huskey cases and list each tool named, with its page.
Done looks like
Every claims algorithm named in US court filings against the 20 largest US insurers since 2020, each linked to its source page. A second person checks 30 entries and agrees with at least 28. The table is updated every quarter.