# Insurance claims handler

> Decides whether a policy pays, how much, and who has to be told no — on the worst day someone has had in a while.

Page: https://flyvolo.ai/en/careers/insurance-claims-handler
Impact index: 63/100 · confidence Low confidence · assessed 2026-09-13
This is not a probability of losing a job.

## Scope

Written for claims handling in personal lines — motor, home, health, travel — where volume is high and most claims look like each other. Complex commercial, liability and large-loss adjusting is a different job with different exposure, and so is underwriting, which decides who gets a policy rather than what a policy pays. Regulation matters here: what an insurer must disclose about an automated decision differs sharply by market.

## Tasks

### Taking the claim

- Direction: Automating
- Basis: evidenced (verified records attached)
- Why: 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.
- What this does NOT establish: 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.
- Records:
  - 2023-12-31 · Deployment · Lemonade's annual report states that as of 31 December 2023, 98% of the time its claims bot takes the first notice of loss and pays or declines the claim without human intervention — https://flyvolo.ai/en/changes/ev-20231231-insurance-claims-handler-1 (source: https://www.sec.gov/Archives/edgar/data/1691421/000169142124000026/lmnd-20231231.htm)

### Deciding whether it is covered

- Direction: Being augmented
- Basis: evidenced (verified records attached)
- Why: 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.
- What this does NOT establish: 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.
- Records:
  - 2025-08-06 · Constraint · EIOPA's Opinion says EU insurers should keep effective internal control and human oversight over AI under existing law; the high-risk use it names is life and health risk assessment and pricing — https://flyvolo.ai/en/changes/ev-20250806-insurance-claims-handler-2 (source: https://www.eiopa.europa.eu/publications/opinion-artificial-intelligence-governance-and-risk-management_en)

### Putting a number on the loss

- Direction: Being augmented
- Basis: evidenced (verified records attached)
- Why: 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.
- What this does NOT establish: 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.
- Records:
  - 2025-12-31 · Deployment · Ping An's 2025 annual report: AI reading medical documents settled nearly a million non-auto claims end to end, 70% of personal injury claims automatically, and fraud detection saved RMB10.51 billion — https://flyvolo.ai/en/changes/ev-20251231-insurance-claims-handler-5 (source: https://group.pingan.com/resource/pingan/IR-Docs/2026/pingan-ar25-report.pdf)
  - 2021-10-28 · Deployment · US state regulators' survey of 193 large auto insurers: AI claims models mostly inform adjusters, 55 use them on images of loss, 83 to refer suspected fraud, and none to deny claims — https://flyvolo.ai/en/changes/ev-20211028-insurance-claims-handler-3 (source: https://content.naic.org/sites/default/files/inline-files/PP%20Auto%20Survey%20Team%20Report%20120822.pdf)

### The claim that does not fit

- Direction: Still human-led
- Basis: evidenced (verified records attached)
- Why: 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. Even the most automated insurers route it to people: Lemonade says about 55% of its claims are fully automated, and its system sends the claims it is not authorised to settle, or has concerns about, to human claims experts.
- What this does NOT establish: 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.
- Records:
  - 2025-12-31 · Deployment · Lemonade says about 55% of its claims are fully automated, and its system routes the claims it is not authorised to settle or has concerns about to human claims experts — https://flyvolo.ai/en/changes/ev-20251231-insurance-claims-handler-6 (source: https://www.sec.gov/Archives/edgar/data/1691421/000169142126000016/lmnd-20251231.htm)

### Telling someone no

- Direction: Still human-led
- Basis: evidenced (verified records attached)
- Why: 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.
- What this does NOT establish: 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.
- Records:
  - 2021-10-28 · Deployment · US state regulators' survey of 193 large auto insurers: AI claims models mostly inform adjusters, 55 use them on images of loss, 83 to refer suspected fraud, and none to deny claims — https://flyvolo.ai/en/changes/ev-20211028-insurance-claims-handler-3 (source: https://content.naic.org/sites/default/files/inline-files/PP%20Auto%20Survey%20Team%20Report%20120822.pdf)

### Spotting the invented claim

- Direction: Being augmented
- Basis: evidenced (verified records attached)
- Why: 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.
- What this does NOT establish: 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. Regulators' survey counts how many insurers use such models and one insurer reports what it saved, but none reports how many flagged claimants turned out to be honest.
- Records:
  - 2026-08-03 · Deployment · Two MS&AD insurers put a tool into claims payment that flags claim photos likely made or altered with generative AI, as reference for handlers who keep the final decision — https://flyvolo.ai/en/changes/ev-20260803-insurance-claims-handler-4 (source: https://www.aioinissaydowa.co.jp/corporate/about/news/pdf/2026/news_2026080301510.pdf)
  - 2025-12-31 · Deployment · Ping An's 2025 annual report: AI reading medical documents settled nearly a million non-auto claims end to end, 70% of personal injury claims automatically, and fraud detection saved RMB10.51 billion — https://flyvolo.ai/en/changes/ev-20251231-insurance-claims-handler-5 (source: https://group.pingan.com/resource/pingan/IR-Docs/2026/pingan-ar25-report.pdf)
  - 2021-10-28 · Deployment · US state regulators' survey of 193 large auto insurers: AI claims models mostly inform adjusters, 55 use them on images of loss, 83 to refer suspected fraud, and none to deny claims — https://flyvolo.ai/en/changes/ev-20211028-insurance-claims-handler-3 (source: https://content.naic.org/sites/default/files/inline-files/PP%20Auto%20Survey%20Team%20Report%20120822.pdf)

## How to cite this

- The unit is the task, not the job title: say which tasks are changing and which are not.
- Carry each task's basis and its limits with it; an inferred judgement is not a finding.
- This page gives no number of years. Do not derive one from it.
- More: https://flyvolo.ai/llms.txt