Regulation (EU) 2024/1689 · Artificial Intelligence Act ● Annex III: — days

EU AI Act for Insurance Companies: Claims, Underwriting and Risk Models

Insurance AI sits squarely in Annex III. This guide maps which insurer use cases are high-risk, what deployers must do, and how claims, pricing and chatbots are treated.

Renomira Team · 6 min read
Insurance policy paperwork on an office desk
Photo: Abiba Pauline / CC BY-SA 4.0 via Wikimedia Commons

The eu ai act insurance industry question has a sharper answer than most sectors get: at least one core insurance activity is named explicitly in Annex III, which means classification debates that take months elsewhere are settled by reading the list. Risk assessment and pricing in life and health insurance are annex iii insurance ai high risk uses — the Compliance consequence is immediate. This guide maps which insurer use cases fall where (claims automation, motor pricing, underwriting models, fraud detection, chatbots), what ai act insurance claims processing compliance requires when claims decisions are AI-assisted, and how deployer duties bite even when your model came from a vendor.

The Annex III hook: what is explicitly high-risk

Annex III(5)(c) covers AI systems intended to be used for risk assessment and pricing in relation to natural persons in the case of life and health insurance. Three observations decide most of the classification work:

  • Life and health only. Motor, property and liability insurance are not caught by this entry — their AI can still be caught elsewhere (a claims model making decisions about individuals touches transparency and, where solely automated with significant effects, the GDPR’s Article 22), but the high-risk designation does not attach on this limb.
  • Risk assessment and pricing. Mortality and morbidity models, health-risk scoring, lapse-risk assessment on natural persons — squarely in. Commercial-entity risk scoring is out of the Annex III(5)(c) wording; personal-lines pricing is in.
  • Natural persons. The entry keys on individuals throughout, consistent with the Regulation’s pattern: AI that assesses people carries the heavier regime.

For life and health insurers this collapses weeks of classification workshops into one citation. The next question is role: if the model is developed and placed on the market by an analytics vendor, the vendor is the provider and you are the deployer — but you both carry material duties.

Where other insurance AI lands

Insurance use case Classification Primary obligations
Life/health risk assessment and pricing on natural persons High-risk — Annex III(5)(c) Provider: Chapter II–III regime; deployer: Article 26 duties
Motor or property pricing Not high-risk on this limb Transparency where individuals are informed by AI; GDPR if solely automated decisions
Claims settlement automation (any line) Depends on effect, not label Art. 50 disclosure where chatbots interact; GDPR Art. 22 safeguards where solely automated significant decisions; documented human review
Fraud detection models Generally not Annex III — unless used for creditworthiness of natural persons Governance, GDPR, model-risk documentation
Customer-service chatbots Not high-risk Article 50(1) interaction notice, enforced the day you launch
Claims triage or document summarisation Not high-risk Human review before decisions; accuracy records; internal AI policy

The pattern: underwriting-adjacent AI on personal lines drifts toward high-risk; operational AI (triage, summarisation, fraud rules) mostly sits in transparency-plus-GDPR territory — but operational is not unregulated. A claims chatbot that quietly recommends settlement amounts has walked out of the operational category without anyone noticing; classify by what the system actually decides, not by what it was bought for.

Deployer duties: what the insurer owes directly

Even when the vendor holds the provider role, an insurer deploying a high-risk system owes Article 26 in its own name: use the system in accordance with the instructions for use; assign human oversight to natural persons with the necessary competence, training and authority; ensure input data relevance is appropriate to the intended purpose; monitor the system’s functioning and suspend use if risk arises; retain the logs under your control for at least six months; inform workers and their representatives before workplace deployment; and cooperate with the authorities. For an insurer, the operationally heavy items are oversight and inputs: claims handlers and underwriters using model output must be named, trained, and empowered to override — with the override recorded. An adjuster who “always accepts the model score” is a human-oversight failure even when the paperwork says otherwise.

Two further deployer duties deserve attention: registration for public bodies and private entities providing public services (not typical insurers, but relevant where you administer public schemes), and the fundamental rights impact assessment under Article 27 for deployers using Annex III(5) systems in their professional activity — the entry that catches life and health risk-pricing tools used by private insurers. The FRIA is a documented pre-deployment exercise on effects on natural persons and available remedies; if you have never run one, your Article 26(7) obligations and this one are the two items an authority will ask for first.

Vendor management: what to demand and verify

Your provider contract for in-scope systems should deliver: classification with article reference; instructions for use written for your claims or underwriting workflow; technical documentation held by the provider with cooperation duties on authority requests; data-governance statements covering the data your portfolio contributed; post-market monitoring and serious-incident procedures with notification commitments; and change control — advance notice if the model, training data or intended purpose changes, because a purpose change can move classification. Verify rather than assume: ask for the EU declaration of conformity where the provider confirms high-risk status, the register entry under Article 49, and their Article 50 marking behaviour for any generative component. A vendor questionnaire answer without a document reference is a gap, not evidence.

How this maps onto model risk work you already do

Insurers have run formal model risk management for years — validation, challenger models, inventory, version control. The AI Act does not replace that programme; it re-badges parts of it and adds what model risk never covered. Direct correspondences: your model inventory becomes the AI register (add classification, role and article reference); validation reports feed the technical file and Article 10 data-governance evidence; change-control logs satisfy the provider duty to notify material modifications. The genuinely new pieces are transparency (Article 50 notices where people interact with the system), the deployer-side human-oversight assignment with authority to override, the FRIA for fundamental-rights effects, and incident reporting under Article 73 where you hold provider duties. Map once, document the mapping, and your model governance board and your AI governance rhythm become the same meeting — the single most common efficiency we see mid-market insurers miss.

The FRIA and the fairness file

Where Article 27 catches you — deployers using Annex III(5) systems professionally — the fundamental rights impact assessment becomes a document you produce before first use: description of the system and its purposes, the periods and categories of persons affected, the specific risks of harm to health, safety and fundamental rights (discrimination being the headline concern in personal-lines pricing), and the human oversight and governance measures answering them. Run it jointly with your model-risk and fairness testing rather than as a parallel legal exercise — the statistical work you already do on disparate impact supplies most of the evidence, and the FRIA is where it gets signed, dated and retained. Even where Article 27 does not formally bind you, producing the same analysis is the fastest answer to the two questions every regulator asks about insurance AI: does it treat people fairly, and who checked? Pair it with your Article 4 literacy record for underwriters and claims handlers, and you have covered the deployer side that authorities actually test.

How to check your own position

List every AI-assisted decision in your underwriting, pricing and claims flows, then classify each in one line: life/health risk-pricing on natural persons (high-risk, Annex III(5)(c)), other pricing (transparency plus GDPR where decisions are automated), claims decisions (human review documented, Art. 50 where a chatbot interacts), fraud and operations (governance, not high-risk). For each high-risk row, check the four deployer artefacts: instructions-for-use on file, named and trained overseers with documented override authority, six-month log retention configured, and the Article 27 FRIA signed before first use. For each vendor-supplied system, confirm you hold their classification reference and change-notification commitment rather than a sales assurance. The whole pass fits in a day for most insurers — and the questions it cannot answer are precisely the ones to escalate before a customer, a broker or a market surveillance authority asks them for you.

Frequently asked questions

Which insurance AI systems are high-risk under the EU AI Act?

Annex III(5)(c) names them: AI systems intended for risk assessment and pricing in relation to natural persons in life and health insurance. That covers mortality and morbidity models, health-risk scoring and personal-lines pricing in those categories. Motor and property pricing are not caught on this limb, though transparency, GDPR and internal governance still apply to them.

Does the EU AI Act apply to claims processing automation?

Not automatically as high-risk — claims automation is not an Annex III entry. But the surrounding duties still bite: chatbots handling claims carry the Article 50(1) interaction notice, solely automated claim decisions with significant effects trigger GDPR Article 22 safeguards and human review, and if your claims tool is repurposed for risk assessment or pricing on natural persons, classification moves with the intended purpose.

What must an insurer do as a deployer of a high-risk AI system?

Article 26 duties, in the insurer’s own name: use the system per its instructions; assign oversight to named, trained and empowered natural persons; ensure input data appropriateness; monitor functioning and suspend on risk; retain logs at least six months; inform workers before workplace deployment; cooperate with authorities — plus the Article 27 fundamental rights impact assessment before first use where the professional-activity scope catches you.

Do insurance chatbots need to disclose they are AI?

Yes. Article 50(1) requires that people are told they are interacting with an AI system unless that is obvious — a persistent label such as “AI assistant” at first interaction is the standard answer. If the bot quotes prices, triages claims or gives anything resembling advice, keep the human-review boundary visible too: disclosure answers the bot question, not the decision question.

How does the AI Act affect insurance pricing models?

For life and health personal pricing, the model is high-risk: provider documentation and conformity duties on the vendor side, oversight, FRIA, monitoring and logging on the insurer side. For other lines, the Act does not classify pricing as high-risk, but customers can still demand explanation under the GDPR where decisions are solely automated, and unfair outcomes remain actionable under consumer and non-discrimination law — classify correctly, then treat fairness as a standing obligation either way.

Next step

Put this guide to work on your own estate.

Run the free exposure scan to see which of these articles actually apply to your company — 12 questions, four minutes, no card. Confirm your email and the full report unlocks.

  • Free forever plan
  • No credit card
  • EU-hosted · GDPR-native
Run a free scan Sign up free

Build a live AI register in minutes — upgrade only when you need the audit pack.

eu ai act insurance industryai act life and health insurance risk assessmentinsurance claims automation ai regulationai pricing rules insurance eueu ai act underwriting ai systemsinsurance chatbot compliance eu ai actannex iii insurance ai high riskai act insurance claims processing compliance

Keep reading