How to use AI in underwriting and claims without losing control
Short answer
Artificial intelligence adds most value at an insurer or brokerage when it removes repetitive work (reading documents, classifying claims, drafting replies) and when it helps prioritise (lapse risk, possible fraud, assisted underwriting). To stay in control, start with one specific use case, measure it, keep human review for decisions that affect the customer, and comply with the GDPR and the EU AI Act, which classifies risk assessment and pricing in life and health insurance as high-risk.
Where does AI add value in insurance?
| Process | What AI does | What the team gains |
|---|---|---|
| Underwriting | Extracts data from questionnaires and documents and flags what is missing or inconsistent | Less time per application and fewer data-entry errors |
| Claims | Reads claim forms, photos and invoices, classifies the claim and suggests the next step | Faster handling of simple cases |
| Fraud | Detects unusual patterns for a person to review | Investigation focused on the cases with most indicators |
| Portfolio | Estimates which policies are most likely to lapse and which clients may need another product | Retention and cross-selling actions in time |
| Customer service | Answers common questions and drafts replies | More time for cases that need judgement |
What data is needed?
Less than people often think, but it has to be organised. For most use cases the data already held in the management system and at the insurers is enough: policies, premiums, cancellations, claims and their cost, and customer interactions. What matters is that the data is identified consistently, that its origin is known and that there is a legal basis for using it for this purpose.
Reading documents does not require historical data: current models extract information from policies, claim forms or invoices without specific training, although they should be validated on a sample of real documents before going into production.
What do the rules say?
- Regulation (EU) 2024/1689 on artificial intelligence (AI Act). It classifies as high-risk the AI systems intended for risk assessment and pricing of natural persons in life and health insurance. These systems have stricter obligations on risk management, data quality, documentation, human oversight and transparency. The application timetable for these obligations has been under review at EU level, so check the date currently in force. Organisations using AI systems must also ensure their staff have a sufficient level of AI literacy.
- GDPR. People have the right not to be subject to decisions based solely on automated processing that significantly affect them, subject to exceptions and safeguards such as human intervention. Using customer data to train or feed models requires a legal basis, information and, in many cases, an impact assessment.
- Insurance distribution rules. AI does not change the duties to inform, advise and act in the customer’s best interest.
How to start without losing control
- Choose one specific, measurable use case, for example reducing the time to register a claim or anticipating lapses in one line of business.
- Measure the starting point: times, errors, lapse rate.
- Test with real data in a limited pilot and compare with the current process.
- Keep a person in the decisions that affect the customer: accepting, declining, paying or changing the price.
- Document what the system does, with which data and how it is reviewed.
- Train the team so they know when to trust the output and when not to.
- Scale only what has proven results.
Common mistakes
- Starting with the tool instead of the problem.
- Putting customers’ personal data into tools without a contract or data protection safeguards.
- Leaving decisions that affect the customer to a model without review.
- Not measuring before and after, and not knowing whether AI has improved anything.
This guide is for information only and does not replace professional advice for each case.