Classification starts with the function, not the product
Four concepts frame the rest of this guide. Each is expanded in the sections below.
- Underwriting and pricing
- Risk assessment and pricing in life and health insurance is the use case Annex III names.
- Claims handling
- Triage, fraud flags, damage assessment, eligibility, and settlement each need a separate analysis.
- Article 6 analysis
- A written rationale per system, including any reliance on exemptions.
- Runtime evidence
- Logs, oversight records, and policy enforcement that show the system behaves as classified.
Claims and underwriting are different functions
Underwriting and pricing AI sits closest to the wording Annex III uses, because risk assessment and pricing in life and health insurance is the use case it names.
Claims handling covers a wider range of activity. Triage, fraud flags, damage assessment, eligibility, and settlement are distinct tasks, and each needs a separate analysis rather than a single label applied to the whole function.
What the regulation states and what requires interpretation
What the text names
- Risk assessment and pricing in life and health insurance is an Annex III use case.
What requires interpretation
- Whether a specific claims system falls within an Annex III use case.
- How Article 6 applies to that system, including any reliance on exemptions.
- Whether claims AI that influences benefit eligibility or settlement should be treated as in scope or out of scope, which needs its own documented analysis rather than an assumption either way.
Common claims and underwriting use cases: an indicative reading
The table below is an indicative reading only. It does not replace a system-level analysis.
| Use case | Indicative reading |
|---|---|
| Underwriting and pricing | Sits closest to the Annex III wording on risk assessment and pricing in life and health insurance. |
| Claims triage | Needs a separate analysis. |
| Fraud flags | Needs a separate analysis. |
| Damage assessment | Needs a separate analysis. |
| Benefit eligibility and settlement | Needs its own documented analysis rather than an assumption either way, particularly where the AI influences the outcome. |
Documenting a defensible classification rationale
Article 6 reasoning should be recorded for each system, not assumed for a function as a whole. A defensible record includes:
- A written rationale per system.
- Any reliance on exemptions, stated explicitly.
- The Annex III use case considered and how the system relates to it.
Turning classification into runtime controls
A classification is only useful if the system behaves as classified. Plan for runtime evidence that shows this:
- Logs of system behavior.
- Oversight records.
- Policy enforcement.