Insurers rushing to deploy AI tools for catastrophe claims assessment are trading accuracy for speed – and may not realize the cost until a major event exposes the gap.
That is the warning from David Heathcote, head of intelligence at McKenzie Intelligence Services and a former British military imagery analyst, who argues that the AI assessment tools now being marketed to carriers are producing error rates that would be unacceptable from any human analyst.
"If I had an analyst who 20% of the time was wrong, I would fire that person," Heathcote said. "So I don't know why people would be content to utilize a mechanical solution" that requires the same level of scrutiny.
The first issue is explainability. When a regulator challenges a portfolio-level assessment or a policyholder disputes a total loss decision, a carrier needs to show why it believes a property was or was not damaged. AI-only workflows make that difficult.
"It can be a bit of a black box," Heathcote said. "Information goes in, a lever is pulled, an assessment pops out at the other end. If you were to query that assessment, because there's been no human that's been involved in it, it's very hard for a company to turn around and really explain [their decision] in detail."
The explainability problem sits on top of a more basic one: AI does not actually understand what it is looking at.
"We're essentially talking about a probability machine," Heathcote said. "We are talking about something that is trained to identify what a property is. Even though it doesn't really understand what a house is, it has no concept of what housing is and of itself – it's just looking at pixels and then trying to compare pixels to other pixels."
Train a model on US suburban housing and point it at dense neighborhoods in South America, and the results deteriorate fast.
"I have seen a tool just looking at an undamaged, a pre-event image of a city in South America and it's basically saying almost everything is damaged because it is looking at housing that it can't even compute," he said.
The errors run in both directions. False positives inflate loss estimates and undermine trust with capital providers. False negatives delay recoveries and create litigation risk.
The LA wildfires provided a live test. With smoke hampering aerial collection and pressure mounting for quick damage views, AI tools were pushed into action. The results, in Heathcote's telling, were not reassuring.
"There were some wildly inaccurate AI assessments," he said. It was looking at cars, recognizing cars as being a regular shape, assuming therefore that it was property, and then making assessments of whether that property had been damaged or not, he added.
“Baseball diamonds in the US is a classic, because they have a regular shape. And so you have AI saying this baseball diamond has been destroyed — when it's just on the ground, nothing's happened to it."
These are not edge cases. They are the kind of errors that surface precisely when insurers are under the greatest scrutiny.
Heathcote's answer is not to reject AI outright but to keep it in a supporting role. MIS builds its assessments around trained human analysts – people who completed the same six-month military imagery course he did.
The difference shows up in how analysts handle ambiguous evidence. In one flood scenario, post-event imagery showed no obvious standing water. A model might conclude the area escaped damage. Heathcote's team noticed something else.
"In this instance, you can't see any flooding in the image. But if you look at the center of the image, you've got a swimming pool. And that swimming pool is not the normal color of a swimming pool. That swimming pool is very dark. And what that indicates is that floodwater has gone through there," he said. "That's the kind of contextual stuff that the AI just doesn't consider."
Unlike a model, a human analyst can be questioned – why was this property flagged and not that one, what cues were used, how confident are they. Those conversations form the backbone of defensible claims positions.
James Doe, head of business development at MIS, frames the broader lesson plainly: "There's no magic bullet. It's still hard, it's still difficult to do. There's still lots of different angles and complexity."
