AI for insurance claims: where to automate, and where humans still matter

Claims is where an insurer keeps or breaks its promise. It is also full of repetitive work: reading forms, checking policies, chasing documents, and routing files. That makes it one of the most natural places to apply AI.

Adoption is already broad. In surveys by the US National Association of Insurance Commissioners (NAIC), 88% of auto insurers and 70% of home insurers said they use, plan to use, or plan to explore AI or machine learning. In claims, they reported using it for accident image analysis, estimating final settlement values, and fraud detection.

The prize is real. The Coalition Against Insurance Fraud estimates fraud costs American consumers at least $308.6 billion a year and occurs in about 10% of property and casualty losses. Faster, more accurate claims handling also improves customer experience at the moment it matters most.

Not every claims task carries the same risk. The useful question is not "can AI do this?" but "what happens to the customer if AI gets it wrong, and how quickly would we notice?". That gives three levels.

The table below applies that rule to a typical claims process. Exact boundaries will vary by product, value, and jurisdiction, but the pattern holds across most lines of business.

Fraud detection is where AI can save the most money, and where it can do the most damage to honest customers. A model that flags too many genuine claims delays payments, damages trust, and can create unfair outcomes for particular groups of people.

Treat a fraud score as a reason for a closer look, never as a decision. Track how many flagged claims turn out to be genuine, and check whether flags fall unevenly across customer groups. The NAIC's model bulletin on AI reminds insurers that decisions made or supported by AI must comply with all insurance laws, including rules against unfair discrimination.

The message from regulators on both sides of the Atlantic is consistent: using AI does not change who is responsible. The NAIC states that insurers using AI remain responsible for complying with insurance laws, consumer protection rules, and requirements on fairness and accuracy, and that human oversight remains an important part of insurance decision-making.

The NAIC's 2023 model bulletin sets expectations for how insurers govern AI and lists the kind of information regulators may ask for in an examination. In 2026 the NAIC has been piloting an AI Systems Evaluation Tool with 12 states to help regulators review how insurers use AI. Regulators may also ask insurers to explain how AI is used in claims decisions.

In Europe, the insurance supervisor EIOPA published an opinion in August 2025 setting out how existing insurance rules apply to AI governance and risk management, using a risk-based and proportionate approach. Separately, the EU AI Act classes AI used for risk assessment and pricing in life and health insurance as high-risk. Claims systems are not named in that list, but they still sit under existing insurance conduct, data protection, and fairness rules.

Large language models are good at summarizing long claim files, drafting customer letters, and answering adjusters' questions about policy wording. These are strong uses, because a person reviews the result before it reaches a customer.

The risk is that these tools can produce text that sounds right but is wrong. The NAIC notes that they may generate information that sounds accurate but is incorrect, and that their output should be reviewed carefully for important decisions. Never let a generated letter state a coverage decision that a person has not made.

These controls apply whether the system is built in-house or bought from a vendor. Many claims tools are bought from vendors, which is why the NAIC formed a working group on third-party data and models in 2024. Responsibility does not transfer with the contract.

It is easy to report that claims are now handled faster. It is harder, and more important, to show they are handled correctly and fairly. The UK government's AI Playbook makes the point that model performance measures are not the same as service measures. In claims, a model can look accurate in testing and still frustrate customers in practice.

Track a balanced set of measures: time to settle, accuracy of payments, reopened claims, complaints, and customer satisfaction. McKinsey's 2026 research found that the organizations getting the most value from AI are the ones that redesign their workflows, rather than adding AI on top of the old process.

Start with intake and document reading. They are high-volume, easy to check, and free adjusters to spend time on the claims that need judgment. Next, add triage, so complex and sensitive claims reach experienced people faster. Only then consider straight-through payment, and keep it to simple, low-value claims with strict limits.

Before launch, put the basics in place. Our minimum viable AI governance pack lists the evidence to have ready, and our AI production release gate sets out the decisions to make before go-live. If you are buying a claims AI product, our third-party AI evidence pack for financial services covers what to ask the vendor.

Not sure your data or teams are ready? Our AI readiness assessment framework helps leaders score where they stand before committing budget.