AI tools save banking supervisors two hours a week
BDF Paper

AI tools save banking supervisors two hours a week

Banque de France reports that its supervisory AI assistant saves users around two hours weekly. Five priority applications rolled out in 2025 and early 2026 also support staff enquiries, household assistance, transaction checks and product mapping.

Experiments become operational tools

Jérôme Coffinet and Jean-Brieux Delbos describe over 60 AI and data-science experiments at Banque de France since 2015, including nine yielding operational applications by 2025.

Five priority uses followed in 2025 and early 2026: supervisory analysis, human resources enquiries, assistance with household applications, unusual-transaction detection and structured-product mapping.

Initial feedback records savings of around two hours weekly per supervisory user, response accuracy above 95 percent for HR and 90 percent for household assistance, and 95 percent analytical reliability for product mapping.

Transaction checks reduced false alerts.

Other applications support forecasting, inflation-perception indicators, fraud detection and statistical data quality.

Guardrails extend beyond accuracy

The article flags bias, hallucinations, model drift, confidential-data leakage, cyberattacks, environmental costs and systemic amplification from similar models.

Banque de France's AI Advisory Group, established in 2025, issued around 20 opinions in one year.

The Prudential Supervision and Resolution Authority (ACPR) is preparing financial-sector oversight under the European Union's AI Act.

The law entered into force on August 1, 2024, with phased implementation through 2027; bans on unacceptable-risk systems applied from February 2025.

Human oversight remains central

The article treats AI as support for central-bank expertise, not a substitute for human oversight.

Reported efficiency gains sit alongside requirements for confidentiality, explainability and reliable data.

Banque de France's 2026-28 strategy extends deployment to all staff and business processes.

Report an error