Part of the problem is the distribution of images (and text) you get from claims is not the same as what the model was trained on. A classic problem in ML.

Another part of the problem is that a model not specifically fine-tuned to make a total loss determination won't know the relevant factors, nor how an insurance company's concept of a total loss differs from the public's.

And still another part of the problem is that most total loss claims aren't what you, dear reader, are imagining: They are very rarely "the car is a thin pancake after being crushed by a meteor".

The much, much more common scenario is: "50% of the body panels sustained at least paint damage, both headlight modules need replacement, and the front wheels look funny. Given that the vehicle has an MSRP of $FOO, $BAR miles, no prior collision history on carfax, and is a popular color, is it cheaper to repair or total the vehicle?"

Of course, the model can turn over the hard cases to a human adjuster... but then what are we doing here? It only takes 10 seconds for the human adjuster to handle the "crushed by a meteor" case also.

Source: Listening to my SIL rant about being asked to stop training bespoke total loss models and just send it by 1-shotting a commercial LLM.