Insurance claims handler — tasks, one by one

Taking the claim

Structured intake against a policy document is the ideal case for automation: the questions are known in advance, the answers are checkable against records the insurer already holds, and a wrong answer surfaces immediately rather than years later. Lemonade's own annual report states that as of the end of 2023, 98% of the time its claims bot takes the first notice of loss and pays or declines without human intervention.

That figure is one direct-to-consumer insurer writing simple personal lines, and the same filing's next sentence says claims the bot is not authorised to settle, or where it identifies concerns, are routed to human claims experts. So it establishes that the intake end is automated at that company, not that claims work no longer needs people. It also says nothing about traditional insurers running older systems, which is most of the market by premium.

Deciding whether it is covered

Where the facts are clean and the policy is standard, this is rule application and machines do it consistently — more consistently than a tired person at the end of a shift. What does not transfer is the case where the facts are contested, because deciding which version of events to believe is not a rule, and the person deciding has to be answerable for having believed it.

A judgement about the structure of the work, not a measurement of how often facts are contested — and that share differs enormously by line of business. It also does not say consistent is the same as correct: a rule applied consistently to a badly written policy produces consistent unfairness.

Putting a number on the loss

Photo-based damage estimation and parts pricing are now ordinary and genuinely fast — the cost side of this task is well-suited to a model trained on millions of repairs. What is left is the argument: a repairer who says the estimate is too low, a claimant whose item has no market price, and the decision about who absorbs the difference.

Says nothing about accuracy on unusual items or in markets where parts supply is volatile, and nothing about whether faster estimates changed what claimants receive. Neither has been measured publicly.

The claim that does not fit

Automation gets better at the cases that look like previous cases, which is the definition of what it cannot do here. The exception is also where an insurer's money and reputation actually sit: a handful of mishandled unusual claims costs more than thousands of routine ones, and knowing which unusual claim is the expensive one is judgement about this book of business.

A judgement about where the risk sits, not a measurement of how many claims are exceptions. It also does not claim humans handle exceptions well — a tired handler at volume is exactly how an expensive claim gets missed.

Telling someone no

A system can produce a decline; it cannot be the party that answers for it. In most markets a declined claim comes with a right to an explanation, a complaint route and a regulator, and each of those needs a person who can be asked why. This is the task that turns a claims department from a processing function into an answerable one.

It protects the answerability, not the headcount. One person can answer for many automated declines, and in several markets the explanation given is itself templated. What would settle how much the duty weighs is an enforcement action against an insurer over an automated decline, which has not surfaced.

Spotting the invented claim

Pattern detection across a whole book is something people cannot do at all, so this is one of the clearest cases where the machine adds capability rather than replacing it. What stays human is the decision to act on a flag: accusing a customer of fraud is a legal act with consequences, and a model's confidence is not evidence.

Nothing here measures false positives, which are the whole cost of this task — a wrongly flagged claimant is a person accused of a crime by a statistical model. Fraud-detection performance is also rarely published by insurers, so this rests on the shape of the problem rather than on a measurement.