Article 6 and Annex III in practice

Article 6 provides two routes into the high-risk tier. Annex I covers AI that is a safety component of a product already covered by EU product legislation, or is itself such a product. Annex III lists standalone use cases. For firms in financial services and insurance, Annex III is where the work is, and the discipline that matters is reading the entry rather than its shorthand.

The Annex III categories

Employment and worker management. Creditworthiness and credit scoring, at 5(b). Risk assessment and pricing in life and health insurance, at 5(c). Education. Essential services. Law enforcement. Migration. Administration of justice.

Three of those regularly catch firms in this sector, and two of them carry an additional consequence beyond the high-risk obligations.

The insurance entry is narrower than the industry assumes

The entry is drawn around risk assessment and pricing in life and health insurance. It does not extend to every AI system an insurer runs. Motor pricing, property pricing, claims handling, fraud triage, reserving, marketing and operations are outside it unless something else brings them in.

That is a real finding rather than a loophole, and it should be recorded as one. The corresponding trap is the assumption that being outside Annex III means being outside regulation. It does not. The EIOPA Opinion is aimed precisely at systems that are not high-risk, interpreting Solvency II, IDD, DORA and GDPR for them, and it is addressed to the supervisors who examine you. FINMA's expectations apply to a Swiss institution's AI applications generally. And an automated decision with significant effect engages revFADP Article 21 whatever the AI Act says.

The two entries that trigger a fundamental rights impact assessment

Article 27 applies to bodies governed by public law, to private entities providing public services, and to deployers of Annex III 5(b) creditworthiness and 5(c) life and health insurance pricing systems. That is the whole population, and it is narrower than commonly believed.

For a lender or a life and health insurer it means the deployer, not only the provider, owes a formal assessment with six prescribed elements: the deployer's processes in which the system will be used; the period and frequency of use; the categories of natural persons and groups likely to be affected; the specific risks of harm to those categories; how the human oversight measures will be implemented; and the measures to take if the risks materialise, including internal governance arrangements and complaint mechanisms. The results are notified to the market surveillance authority, and a FRIA may incorporate or cross-refer to a GDPR data protection impact assessment rather than duplicating it.

Notification to a supervisory authority is the element that changes the internal politics of the document. An assessment that will be read outside the firm is drafted differently from one that will not, and it is worth telling the business that at the start rather than at review.

Employment catches firms that are not employment businesses

The employment and worker management entry has nothing to do with your industry. Every firm hires, and a screening, ranking or scheduling tool in HR is inside the category. This is the Annex III entry most often missing from a financial services register, because HR systems are procured outside the technology governance perimeter and are rarely described as AI in the contract.

It also interacts with Article 26(7), the duty to inform workers' representatives and affected workers before a high-risk system is put into use at the workplace. That duty is performed before deployment. It cannot be remediated afterwards, which makes it one of the few AI Act obligations that a late discovery cannot fully repair.

Reading an entry properly

Three habits. Quote the entry in the classification record rather than paraphrasing it, because paraphrase is where scope creeps in both directions. Identify the specific decision the system makes and compare it to the decision the entry describes, rather than comparing business domains. And record what would move the system into the entry, because that sentence becomes the reclassification trigger.

Worked example: four systems at one insurer, read against the entries

Life underwriting risk assessment. Inside 5(c). High-risk. Article 27 applies to the insurer as deployer. Full Article 26 set. Dated to 2 December 2027.

Health product pricing. Inside 5(c). Same position. Note that the sensitive data involved also puts it squarely inside revFADP Articles 22 and 23 for the Swiss entity, and the two assessments should be run as one process producing both documents.

Motor claims fraud triage. Outside Annex III: not life or health, not risk assessment or pricing, no creditworthiness determination. Governed through the EIOPA-interpreted track, FINMA expectations, data protection law, and Article 21 revFADP if the triage outcome is a solely automated decision with significant effect on the claimant. Reclassification trigger recorded: extension to health claims, or use to set settlement amounts.

CV screening in group HR. Inside the employment and worker management entry. High-risk. Article 26(7) notification owed to workers' representatives and affected workers before use. Where used for New York City roles, an independent bias audit no more than one year old and candidate notice at least ten business days before use.

The insurer's instinct was that the insurance systems were the exposure and HR was administrative. The HR tool carries the one duty in the set that cannot be fixed retrospectively.

As of 24 September 2026.