Here’s a thing IE learned recently. For decades people believed that an ancient city State in the Sumerian civilisation had a sort of ritual suicide funeral for kings or queens. This was based on the work of one archaeologist, who found 74 skeletons in Ur in the 1920s. He speculated that the bones of musicians, servants etc were set in sleeping positions, as if they had swallowed poison and died peacefully to serve their queen in the afterlife. That theory became written history for decades. But a detailed scan of the skulls in the 1990s revealed a heavy blunt force trauma had killed them all, and then the bodies of the underlings were carefully arranged in position, according to the funeral rites.
There is a lesson here for any insurance brand using AI to settle all their minor motor claims; images or video only scratch the surface. They don’t offer a detailed scan of the car’s underside, the engine mounts, or the battery pack mounting/compartment in hybrids and EVs. Then there’s steering and brakes. All this stuff is hidden and policyholder images might not tell the whole story. Here’s the word from Laird;
Visual artificial intelligence has an important role to play in motor claims, but insurers risk expensive and potentially unsafe decisions if image recognition is allowed to substitute for engineering assessment, according to independent vehicle engineering specialist Laird.
The warning follows Laird’s analysis of more than 250,000 supplementary accident damage assessments as part of its ongoing Project X-Ray research into the damage and repair operations most commonly identified after an initial estimate.
The analysis shows a recurring problem: some of the items that materially change the severity, cost and even roadworthiness of an accident-damaged vehicle simply cannot be reliably identified from external photographs.
Laird says both visual AI and LLMs can be extremely effective where the task genuinely is visual. Vehicle condition monitoring, de-fleeting, off-hire inspections and identifying obvious external panel damage are all areas where image recognition can deliver significant speed and efficiency.
Accident damage assessment, however, presents a different problem.
Nik Ellis, Director of Laird, said:
“AI can be good at telling you that a bumper and a wing are damaged. What it can’t necessarily tell you from those photographs is why the wheel is sitting 20mm further back than it should be.
“The first example might be a straightforward cosmetic repair. The second could involve suspension or structural damage, affect whether the vehicle is roadworthy and potentially change the entire economics of the claim.
“If the system can’t see it, it shouldn’t be allowed to assume it isn’t there.”
Laird’s Project X-Ray work has examined patterns across supplementary assessments, looking at items subsequently identified following initial estimates, further inspection or strip-down.
The analysis found recurring omissions across comparable accident types, with approximately one in eight items reviewed being flagged as worthy of additional engineering consideration. Among strip-down related items analysed, around 60% were absent from comparable initial assessments.
Laird says the findings should not be interpreted as an argument against AI.
The company’s own engineering workflow uses artificial intelligence, automation, fraud filters, image assessment and vehicle data alongside established estimating platforms. The distinction, it says, is that the technology supports the engineering decision rather than being allowed to create an unsupported one.
Ellis continued:
“We use AI every day, so this certainly isn’t an anti-AI argument.
“The problem starts when clever image recognition gets confused with a complete vehicle damage assessment. They’re not the same thing.
“A photograph shows you the outside of the car. Accident energy doesn’t politely stop at the bumper skin.”
The issue has become increasingly important as insurers look to artificial intelligence to increase claims handling speed and reduce assessment costs.
According to Insurance Times’ AI Claims Report 2025/26, 38% of surveyed insurance organisations were already using AI and a further 38% were piloting or testing it. Of those already using AI, 35% were deploying it for image recognition and damage assessment.
Laird argues that the biggest opportunity lies in combining AI speed with appropriate engineering controls.
Visual systems can identify obvious damage, assist triage, compare imagery, identify inconsistencies and highlight areas requiring attention. Engineers can then assess impact direction, likely transferred forces, hidden damage, repair methodology, roadworthiness and the wider consequences for the claim.
Ellis added:
“The objective shouldn’t be to choose between AI and engineers. The sensible model is to let each do what it is good at.
“Use the machine to process enormous amounts of information quickly. Use engineering judgement where the answer depends on something the camera cannot see.
“AI getting an answer in ten seconds isn’t much of an efficiency saving if somebody discovers three days later that it was the wrong answer.”
Laird is continuing to develop Project X-Ray by using historic supplementary assessment data to identify patterns of commonly missed damage according to impact area and accident type.
The objective is to use those patterns proactively, prompting engineers to consider potential hidden damage earlier in the assessment process rather than discovering it after repair has begun.
Be the first to comment