Transparencia de la IA y conformidad con el AI Act de la UE

Regulatory framework and binding effect

This page provides transparency information under, inter alia, Articles 13, 26, 27 and 50 of Regulation (EU) 2024/1689 (the “EU AI Act”), read with the GDPR and French fairness principles.

It is not a certification, CE marking, notified-body attestation, hiring-performance warranty or waiver of Fleury Capital’s defences. Compliance follows the EU AI Act’s phased application timetable through continuous, proportionate compliance-by-design.

Qualification: high-risk AI system (Annex III)

CandidAct uses AI components intended, in recruitment and selection of natural persons, to assess candidates, produce scores, rankings and summaries, and assist recruiters. Such uses fall, as applicable, under high-risk AI systems in Annex III, point 4, of the EU AI Act (employment, workers’ management and access to employment), subject to authentic interpretation and forthcoming guidance.

That qualification triggers heightened duties on risk management, data governance, logging, transparency, human oversight and robustness. It does not itself prohibit the system, nor automatically classify the described uses as Article 5 prohibited practices where the safeguards below are observed.

Roles: provider, deployer, controller

Depending on contract and context, Fleury Capital may act as provider and/or deployer under the EU AI Act. A professional customer using CandidAct to assess its candidates generally acts as deployer and as GDPR controller of candidate data.

Duties (candidate information, operational oversight, documented instructions, records) are allocated by contract and law. Claims attributing to Fleury Capital failures solely attributable to the customer deployer will be contested.

Intended purpose, permitted use and prohibited misuse

The intended purpose is to assist hiring through job simulations, AI-assisted scoring against an approved rubric, an assisted voice interview, and a summary for a human decision-maker. Permitted use excludes: prohibited social scoring; prohibited sensitive biometric categorisation; solely automated rejection decisions with legal or similarly significant effects without meaningful human involvement; Article 5 manipulative practices.

Customer misuse (intentional discrimination, unnecessary special-category collection, off-purpose surveillance) is the customer’s sole responsibility and may lead to suspension.

Candidate information and operational transparency

Before the assessment starts, candidates are told clearly that AI components analyse the application, simulation and/or voice interview, that proctoring may apply (full screen, camera, relevant technical event logs), and that hiring decisions remain human.

The customer deployer must ensure its own channels do not contradict this information. Fleury Capital does not warrant the customer’s external communications.

Human oversight

In line with Article 14 EU AI Act, the system supports effective human oversight: recruiters can access the submission, rubric, justified scores, excerpts, transcript and integrity signals, and may advance, set aside, request review or disregard an automated recommendation.

Customers must not configure internal processes that nullify oversight where the law forbids solely automated rejection.

No prohibited solely automated decision-making

CandidAct is not designed to produce, by itself, a solely automated hiring-rejection decision under GDPR Article 22 and related prohibitions. Rankings and scores are decision-support. Marketing claims must be read in light of this page, which prevails on automation character.

Explainability, remedies and contact channel

Where reasonably feasible, assessments rely on an approved rubric, justified scores and excerpts. Candidates receive feedback and may request clarifications via través del formulario de contacto publicado en el Sitio.

Explainability does not require disclosure of source code, model weights, full system prompts or trade secrets where the information given allows understanding of the general logic, main criteria and human role. Abusive or competitive reverse-engineering requests may be refused or limited.

Logging, traceability and log retention

Relevant technical and business events may be logged for traceability, security, anomaly detection and diligence. Logs are retained proportionately and are not absolute proof of all real-world facts. Full log disclosure to unauthorised third parties may be refused where it would harm security, others’ privacy or technical secrets, subject to lawful process.

Accuracy, robustness, cybersecurity and limits

Fleury Capital implements proportionate accuracy, robustness and cybersecurity measures (Article 15 EU AI Act in principle). Language and speech models have intrinsic limits. Scores are assisted estimates, not metaphysical truths. No warranty of zero error, zero bias or certain on-the-job performance is given. Hiring decisions remain those of the human decision-maker/employer.

Fairness, non-discrimination and bias monitoring

The design aims at procedural equality: same exercise and rubric per role, explicit criteria, human review. Reviews may detect drift. Residual statistical or societal bias is not warranted to be fully eliminated. Customers remain responsible for employment and non-discrimination law in final decisions. Unlawful customer instructions may be refused.

Input and output data governance

Input and output data are processed under GDPR minimisation, purpose limitation and security, with governance aimed at reducing irrelevant or poor-quality assessment data. Third-party model training data remain under those providers’ documentation. Candidate content is not released publicly for third-party model training outside contract.

Liability, exclusions and governing law

Within mandatory law, Fleury Capital excludes indirect damages and loss of hiring opportunity, and any claim reading this page as a results obligation. French law governs; disputes follow the legal notice jurisdiction rules. Misleading partial quotation may be challenged. The French version prevails if locales diverge.

Regulatory contact

For AI transparency questions, explanation requests or reports, use exclusively el formulario de contacto publicado en el Sitio. Off-channel, anonymous or dilatory filings may not receive the same handling times.