Cyberpediaby Brain

AI in credit and insurance: transparency and fairness

What the AI Act classes as high-risk for credit and life or health insurance, and what the CNIL recommends for scoring.

On this page
  1. 1In 30 seconds
  2. 2What it is
  3. 3Straight answers
  4. 4How it works
  5. 5What you do
  6. 6Pitfalls
  7. 7Sources
  8. 8Cite this article

In 30 seconds

Where it happens
At a lender or an insurer, wherever a tool scores a customer's creditworthiness or assesses risk and sets the price of life or health insurance[1].
Why it is classed as high-risk
These tools determine access to financial resources or to essential services and may perpetuate past discrimination[2].
The reflex
Know whether the tool you use scores creditworthiness or prices insurance, and know that a person subject to a fully automated decision can ask for a human to intervene[4].
If a customer objects
They can ask for their situation to be reviewed, ask for an explanation of the decision, or contest the decision[4].

What it is

The European AI regulation (AI Act) classes as high-risk the systems intended to assess the creditworthiness of a natural person or to establish their credit score, with an exception for systems used to detect financial fraud[1]. It also classes as high-risk the systems intended for risk assessment and pricing in life and health insurance[1].

Recital 58 gives the reason for this classification: these systems may lead to discrimination between persons or groups, and in insurance to financial exclusion[2]. For credit, the CNIL published a recommendation on creditworthiness assessment on 7 May 2026, which recalls the safeguards attached to fully automated decisions[7].

This page explains what these texts say to an employee of a bank, an insurer or a customer service. It does not settle any particular case.

Straight answers

Is a credit score a high-risk system?

Yes, under Annex III, which covers systems intended to assess the creditworthiness of natural persons or to establish their credit score[1]. The European Commission gives as an example a system that produces a numerical representation of a person's creditworthiness, based on their payment behaviour and income, to support a consumer credit or mortgage decision[3]. It also cites the score produced by a credit agency and passed to a third party: the user of the score does not need to be the deployer of the system for the case to fall under the list[3].

The Annex III rules apply from 2 December 2027, after the postponement decided with the AI Omnibus[5]. The CNIL gives the same date[6].

Does insurance fall within scope?

For life and health insurance, the text covers risk assessment and pricing in relation to natural persons[1]. The European Commission cites an insurer that reviews life insurance applications based on age, health status, family history, lifestyle and occupation, and an insurer that predicts the annual medical cost of a risk group, with the result becoming the premium[3].

Are all of an insurer's or a bank's tools covered?

No. Recital 58 specifies that systems provided for by Union law for detecting fraud in the offering of financial services, and those that calculate the own funds requirements of credit institutions and insurance undertakings for prudential purposes, are not considered high-risk[2]. The Commission's information service also excludes, among others, customer classification for personalised marketing as long as it plays no role in the creditworthiness assessment, assistance that helps an applicant complete their file, the handling of complaints after a decision, the internal monitoring of credit exposures after the credit is granted, and health insurance claims management[3].

Can the person contest a scoring decision?

Article 22 of the GDPR gives, as a principle, the right not to be subject to a decision based solely on automated processing that produces legal effects or significantly affects the person[4]. Exceptions exist: explicit consent, necessity for entering into or performing a contract, specific legal provisions, and in these cases safeguards must apply[4]. The CNIL notes that the CJEU rulings in cases C-634/21 and C-203/22 confirm that scoring is an automated decision when it plays a determining role in granting credit, and establish a right to an explanation of how the mechanism that led to the decision works[7].

How it works

The AI Act does not replace the GDPR: the CNIL specifies that it complements it, and that the two texts apply together when a high-risk system uses personal data[6].

From 2 December 2027, Annex III high-risk systems are subject to obligations before they are placed on the market[5]. The European Commission lists seven: risk assessment and mitigation systems, high-quality datasets to limit discriminatory outcomes, logging of activity to ensure traceability of results, detailed documentation that lets authorities assess compliance, clear information to the deployer, human oversight measures, and a high level of robustness, cybersecurity and accuracy[5]. Human oversight and monitoring are the responsibility of deployers, that is, the organisations that use the tool[5].

On documentation, the CNIL recalls that providers of high-risk systems supply their deployers with technical documentation and information that detail, among other things, the testing and conformity assessment procedures[6]. It adds that the data protection impact assessment can usefully draw on this documentation[6]. The text also allows, exceptionally, the processing of sensitive data to detect and correct bias, if strictly necessary and subject to appropriate safeguards[6].

What you do

The obligations of the AI Act weigh first on providers and deployers[6]. For an employee, the resulting actions are as follows.

  • Identify the nature of the tool. A score that assesses creditworthiness, or a pricing tool for life or health insurance, falls under the list[1]. A segmentation tool or a tool that helps complete a file is excluded if it does not take part in the assessment[3].
  • Know what the person must be told. They are informed, when their data is collected and at any time on request, of the existence of a fully automated decision, of the underlying logic, and of the significance and envisaged consequences of the decision[4].
  • Allow human intervention. Any person concerned can ask for a human to intervene, to obtain a review, express their point of view, obtain an explanation or contest the decision[4]. The CNIL lists transparency, human intervention and explainability among the safeguards for a fully automated credit decision[7].
  • Use relevant data. The CNIL recommendation asks to limit data to what is relevant and strictly necessary, to use objective information for past repayment incidents, and sets retention periods[7]. It targets private bodies that grant consumer or mortgage credit and their intermediaries[7].
  • Keep documentation. The CNIL recommends establishing and keeping available documentation on how the dataset was built (source, sampling, integrity checks, cleaning) and regularly reassessing the risks of discrimination and bias[8].
  • Consult the CNIL checklist, intended for data protection officers, compliance officers and legal teams[7].

Pitfalls

  • Confusing profiling with automated decision-making. Profiling evaluates a person, a fully automated decision is taken without any human intervention, and one can exist without the other[4]. A credit refusal based on the automatic application of criteria to the applicant's financial situation is the example the CNIL gives[4].
  • Believing a recommendation is the law. The CNIL states that it is possible to depart from its advice, provided the choices made are justified, in particular in the event of an inspection[7].
  • Stretching the fraud exception. Annex III excludes systems used to detect financial fraud, not those that assess creditworthiness[1]. If a complaints-handling or monitoring tool also assesses creditworthiness or establishes a credit score, the exclusion does not apply[3].
  • Getting the calendar wrong. The AI Act has been applicable since 2 August 2026, but the Annex III rules apply from 2 December 2027[5]. The AI regulation does not replace the GDPR, it complements it[6]. The CNIL also updated its recommendation to anticipate the entry into application, in November 2026, of a legal authorisation for fully automated decisions on consumer credit, recalling the safeguards attached to it[7].

Sources

  1. European Commission, “ANNEX III”. https://ai-act-service-desk.ec.europa.eu/en/ai-act/annex-3 (accessed 14 November 2026)
  2. European Commission, “Recital 58”. https://ai-act-service-desk.ec.europa.eu/en/ai-act/recital-58 (accessed 14 November 2026)
  3. European Commission, “Essential Services”. https://ai-act-service-desk.ec.europa.eu/en/essential-services (accessed 14 November 2026)
  4. CNIL, “Profilage et décision entièrement automatisée”, 29 May 2018. https://www.cnil.fr/fr/profilage-et-decision-entierement-automatisee (accessed 14 November 2026)
  5. European Commission, “AI Act”. https://digital-strategy.ec.europa.eu/en/policies/regulatory-framework-ai (accessed 14 November 2026)
  6. CNIL, “Entrée en vigueur du règlement européen sur l’IA : les premières questions-réponses de la CNIL”, 12 July 2024. https://www.cnil.fr/fr/entree-en-vigueur-du-reglement-europeen-sur-lia-les-premieres-questions-reponses-de-la-cnil (accessed 14 November 2026)
  7. CNIL, “Octroi de crédit : la CNIL publie sa recommandation sur l’utilisation de données personnelles pour l’évaluation de la solvabilité”, 7 May 2026. https://www.cnil.fr/fr/recommandation-octroi-de-credit (accessed 14 November 2026)
  8. CNIL, “IA : comment être en conformité avec le RGPD ?”, 5 April 2022. https://www.cnil.fr/fr/intelligence-artificielle/ia-comment-etre-en-conformite-avec-le-rgpd (accessed 14 November 2026)