Photo-Based AI Car Damage Estimates: In Production, but Accuracy Is Unproven

The use case

A policyholder, a repairer or an insurer's own staff photographs a damaged car. A machine-learning model, meaning software trained on large archives of past damage photos and the estimates written for them, identifies the damaged parts, judges how severe the damage is and proposes repair operations, labour and parts⁸. Insurers use the output in four ways: to triage a claim as repairable or a likely total loss, to draft an estimate that an adjuster approves, to check an estimate a repairer has written, or, where the insurer's rules allow, to send an estimate straight to the customer¹³⁴.

Scope note. Only the step that reads the images and predicts damage and cost counts as AI here. Three older things are kept separate. First, the estimating databases that price parts and labour once repair operations are chosen; these are rule-based and long established. Second, photo estimating by human desk appraisers who work from claimant photos, which is now a large channel of its own and dominates the industry statistics on "photo estimates"²¹⁰. Third, the insurer-set rules that decide which claims are eligible for automation. Fraud scoring, total-loss valuation from market data and damage scanning at rental or lease return are out of scope.

Where it is running

The CCC figures are the broadest public measure of deployment, but they cover all of the company's AI products, including tools for bodily-injury claims, so they overstate photo-based motor estimating on its own¹. Roughly 10% of CCC's 2025 revenue, nearly $100 million, came from these AI products (vendor-reported, FY2025)¹.

Covéa is the clearest European case. Its 160,000 finalised claims (operator-reported, April 2026) accumulated over several years and sit against a customer base of 11.3 million (operator-reported, April 2026)⁴. The tool works through Covéa's approved repairer network, so it mainly speeds up the exchange between repairer and insurer rather than replacing the adjuster for every policyholder⁴.

The evidence is concentrated in the United States and France. This research found no deployment figures from the past three years outside these two countries.

What the evidence shows

No controlled field evaluation. This research found no randomised trial, quasi-experimental study, regulator audit or independent evaluation that measures how close deployed photo-AI estimates come to the final repair cost. Every performance figure below is either vendor-reported, operator-reported or from a laboratory test.

Laboratory accuracy is modest. A peer-reviewed 2025 conference study built a hybrid model that classified damage severity from user-uploaded images with 73.3% overall accuracy (measured on a test set, no field deployment, 2025)⁵. Studies of this kind do not test the commercial systems insurers use, but they show that judging severity from photos remains a hard task.

Speed. CCC's chief executive said in August 2025 that its photo-AI products had cut the time to identify a total loss by half for insurers using them (vendor-reported, no control group described, August 2025)³. Covéa says repairers receive assessments immediately, but publishes no cycle-time figure (operator-reported, April 2026)⁴.

Reach. Even at the largest vendor, AI handles a low single-digit to low double-digit share of claims, depending on the product (vendor-reported, February 2026)¹.

The baseline AI has to beat. Photo estimating of all kinds made up 26.4% of US repairable-claim inspections in 2025, behind direct-repair-programme shops at 46.7%, staff appraisers at 16.5% and independent appraisers at 6.9% (vendor-reported industry data, 2025)². The published summary does not separate estimates written by AI from those written by human desk appraisers². The most detailed comparison of photo and in-person estimates is older and concerns human appraisers: between 2013 and 2017, supplements on photo estimates averaged 52.9% of the original estimate's value, against 34.5% for staff field inspections and 30.5% for independent appraisers (vendor-reported, 2018, more than three years old)¹⁰. A supplement is an addition to the estimate once the car is opened up. The chance of a supplement was similar across channels, 28.6% for third-party photo estimates against 25.7% and 22% for the in-person channels (vendor-reported, 2018)¹⁰, and photo estimates were used mostly on lower-value repairs, which inflates their percentage¹⁰. No newer comparison was found, so there is no published evidence that AI has closed this gap.

What is missing is simple to state: a comparison, on the same claims, of the AI's first estimate against the final paid repair cost, published by someone other than the vendor or the insurer.

What it takes: cost, integration, governance

Cost. No public price per claim was found. The only spending signal is CCC's AI revenue of nearly $100 million a year across more than 125 insurers, which includes non-motor products (vendor-reported, FY2025)¹.

Data and integration. Models are trained on large sets of photos supplied by loss adjusters and repairers, labelled by people with damage type and severity⁸. The European case runs through an existing approved-repairer network rather than a standalone consumer app⁴. Photo quality matters: the system can only assess what the images show⁶⁸.

People. Spanish industry guidance describes a typical arrangement in which the AI gives a preliminary estimate and a loss adjuster reviews it to confirm or change the amount⁸. Negotiation with repairers over methods, costs and timing still needs an experienced adjuster⁸.

Rules. In the United States, the New Hampshire Insurance Department told insurers on 14 April 2025 that photographs and videos do not always show the full extent of damage, and that they must provide an in-person inspection within a reasonable time if the consumer asks, even after the consumer first chose a virtual process⁶. The bulletin applies to photo-based assessment in general, whether a person or software reads the images⁶. In the European Union, the AI Act lists AI for risk assessment and pricing in life and health insurance as high-risk, but not motor claims handling, so photo damage estimation does not by default carry the Act's heavier obligations⁷. A 2023 survey by the European insurance supervisor found 50% of responding non-life insurers already using some form of AI (self-reported, published April 2024)⁸.

Where it falls short

Hidden damage. Photos show the surface. Structural, mechanical and sensor damage behind a panel is not visible, and an engineering assessment from Centro Zaragoza, a Spanish vehicle research centre funded by insurers, names this as a core weakness, along with misidentifying specific parts⁸. The New Hampshire regulator makes the same point about photo assessment in general⁶.

Estimates change after the car is opened. More than 60% of US estimates now receive at least one supplement, by the vendor's own account (vendor-reported, August 2025)³. A fast first estimate that is later revised may move delay from the start of the claim to the repair stage rather than remove it.

Disputes and litigation. A lawsuit filed on 7 January 2026 in the US District Court for the Northern District of West Virginia accuses Liberty Mutual and several vendors, including CCC and the virtual-appraisal firm Snapsheet, of a scheme to underpay claims⁹. The complaint cites vendor marketing that allegedly promised insurers average indemnity savings of 15–20% (allegation in court filing, January 2026) and an AI-driven claims platform that allegedly rewarded fast, low settlements⁹. These are allegations; no ruling has been made⁹. The case mixes human virtual appraisal with AI tools, which is typical of how these systems are deployed and makes it hard to assign blame to the AI alone.

Fraud and manipulation. The same Spanish assessment warns that fraudsters may learn to submit images designed to fool automated systems, and that the savings on adjuster fees could be offset by higher claim payments⁸.

Challenge question. What would make this use case look less successful than it is usually presented? Four things. First, the headline speed gains are vendor- or operator-reported, from systems applied only to claims the insurer's rules select as simple, with no published control group¹³⁴. Second, no study has been replicated independently, and the only peer-reviewed accuracy figure found comes from a laboratory model, not a deployed one⁵. Third, much of the benefit depends on surrounding processes: repairer networks, adjuster review and eligibility rules, rather than on the AI itself⁴⁸. Fourth, the people least able to take good photos or to contest a low offer are the ones the regulator's in-person safeguard is meant to protect, and a regulator judged that safeguard necessary⁶. On current evidence, the gains are real enough for insurers to keep buying, but whether policyholders receive more accurate settlements is unmeasured.

Phase: In early use — why

The hypothesis was "In early use", and the evidence supports it. The systems are in production at named insurers under multi-year contracts, which rules out a pilot¹⁴. But AI handles only a low single-digit to low double-digit share of claims at the largest vendor, and the main European case has finalised about 160,000 claims in total over several years, which rules out wide use¹⁴. No independent outcome evidence exists that would support moving the phase either way.

What to watch next

• The West Virginia case against Liberty Mutual and its vendors, filed 7 January 2026: any ruling on the motions or disclosure of how vendor tools set estimates⁹.
• CCC's full-year 2026 results, expected around February 2027: whether the share of claims processed with AI rises beyond low double digits¹.
• CCC's next annual industry report, expected in spring 2027: whether photo-estimate statistics are split between AI-written and human-written estimates².
• Covéa's three-year renewal, running from April 2026: any published cycle-time, accuracy or complaint figures⁴.
• Further US state insurance bulletins following New Hampshire's April 2025 position on in-person inspection⁶.

Questions to ask before adopting

• On which claims, by damage type and value, will the AI's estimate be used without an adjuster's review?
• How far did the vendor's first estimates differ from final paid repair costs on our own past claims, measured by someone independent?
• What share of AI estimates later needed a supplement, and how does that compare with our in-person channel?
• How will a customer request and receive an in-person inspection, and how quickly?
• How will we detect manipulated photos and monitor whether settlements drift lower over time?