Financial AI: What Banks and Supervisors Actually Measure

European banks and card networks already rely on AI for part of their credit scoring and fraud detection. The measurable evidence of that spread sits in supervisory reports and threat surveys, far more than in published accounts.

A bank results release, ECB supervisory work, the first DORA incident report and the EU AI Act together separate what is measured from what is merely claimed.

The prevailing story about financial AI is a promise: efficiency gains that will eventually surface in bank margins. BNP Paribas’s second-quarter 2026 results show why that story is hard to check. The group’s cost-to-income ratio, operating expenses divided by revenues, came in at 56.7%, an improvement year on year. The chief executive points to the acceleration of the group’s transformation, notably in artificial intelligence. Nothing in the release says how much that lever contributed.

The silence says little about how far AI has spread. It says where to look. In finance, AI surfaces in risk registers long before it reaches the accounts.

What published accounts reveal

The cost-to-income ratio is the efficiency metric every listed bank discloses, and the one where an AI dividend would be expected to appear. At BNP Paribas, the second-quarter improvement came from a positive jaws effect: revenues rose faster than operating expenses year on year, according to the results release of 23 July 2026. Costs are broken down by business line. Never by technology. Attribution is genuinely hard: a model that cuts manual reviews in fraud operations also changes staffing, vendor contracts and loss rates, and none of these maps onto a single line of the income statement.

Euro area banking supervision reaches the same conclusion from the inside. In its Supervision Newsletter of 20 November 2025, the ECB notes that banks describe gains in accuracy and efficiency, but that the financial quantification of realised benefits remains a challenge, echoing repeated on-site findings about banks’ ability to track the financial impact of digital projects.

The macro setting blurs the picture further. Euro area real GDP grew 1.28% in 2025 (Eco3min calculation on the Eurostat series CLVMNACSCAB1GQEA19, sum of four quarters). The ECB deposit facility rate, raised in June and again in September 2026, has stood at 2.5% since 16 September (series ECBDFR). Rising rates support net interest income, hence revenues. Part of any improvement in the cost-to-income ratio comes from there, which makes AI’s contribution even harder to isolate.

Credit and fraud, the two uses in production

Two uses dominate. The same ECB publication reports that its annual data collection on innovative technologies, covering 107 significant banks in 2023 and 110 in 2024, shows a strong increase in AI use cases, notably for credit scoring and fraud detection.

The ECB then held workshops with 13 banks between May and August 2025. Ten used AI for credit scoring and ten for fraud detection (workshop annex). Most models are decision trees, with neural networks used mainly for fraud. No bank in the sample used generative AI to score borrowers, and none allowed self-learning after deployment. In lending, AI supports a human decision, except for small retail loans that are sometimes approved automatically. The ECB stresses that a sample this small cannot be generalised to the sector.

Fraud shows the other side. Visa’s Spring 2026 Biannual Threats Report, released on 20 May 2026, identified nearly $1 billion in scam-related activity between July and December 2025, the largest category of consumer payment fraud. Over the same period, fraud involving device tokens fell 9.6% year on year. Visa reads this as a sign that authentication is holding and that attackers are turning to manipulating the payer. AI makes those scams more convincing, and it also speeds up detection. The statistical limits of these scoring models are covered in our piece on automated risk scoring.

DORA, the first public count of operational risk

Since 17 January 2025, the Digital Operational Resilience Act has required banks, insurers, payment firms and market infrastructures to report major ICT incidents. The first joint report of the EBA, EIOPA and ESMA, published on 3 June 2026, counts 3,383 of them for 2025: 282 a month on average, or 0.18 per reporting entity.

More than 60% hit the banking sector and 16% the payments sector. Cyberattacks account for 10% of the total, with ransomware aimed mainly at insurers, which hold large volumes of sensitive health and financial data. System failures dominate, and 29% of incidents stem from a failure at a third-party provider. A third had a cross-border impact, while two thirds caused no or only minor disruption to clients. The authorities warn that incident counts are not a risk indicator in themselves and that reporting practices still diverge across countries. The register’s value lies less in its headline total than in its breakdown: for the first time, an outage at a provider shared by many institutions can be traced across the EU.

The report does not classify any incident as an AI model error. AI appears in another form: highly capable AI-driven tools, available to attackers, call for higher cybersecurity standards. For a scoring model hosted in the cloud, an outage at the provider is an ordinary DORA incident. The ECB workshops found exactly that pattern, with models often built in-house but run by cloud service providers.

A regulatory timetable pushed back sixteen months

Annex III of the EU AI Act classifies as high-risk the systems that assess the creditworthiness of individuals or set their credit score, except those used to detect financial fraud. The same annex covers risk assessment and pricing in life and health insurance, so insurers’ pricing models face the same deadline. Their risk management, data governance and human oversight obligations were due to apply from 2 August 2026.

Regulation (EU) 2026/1744, the Digital Omnibus on AI, published in the Official Journal on 24 July 2026 and in force since 27 July, moved that date to 2 December 2027, citing delayed standards and the late set-up of national authorities. The compliance cost is deferred, not removed. The extra sixteen months mostly buy time for model inventories and self-assessments, which the ECB workshops found under way but uneven across banks. The fraud exemption also creates an asymmetry. A model that grants a loan falls under Annex III. One that blocks a payment does not. How far rules can reach is discussed in our analysis of regulating financial AI.

The risk that is not yet counted

One risk still escapes the statistics. In chapter III of its Annual Economic Report of 30 June 2024, the BIS warns that institutions using the same algorithms, trained on the same data, could amplify procyclicality through herding, liquidity hoarding, runs and fire sales. Reliance on a handful of model providers adds third-party risk.

DORA’s provider failures offer an early hint. The mechanism is developed in our analysis of AI-driven concentration and set in the wider frame of AI as a financial stability issue.

The finding that gains remain invisible would be contradicted if large banks began disclosing AI-attributed savings on an auditable perimeter, or if the DORA report on 2026 singled out model-related incidents. Either would move the debate from stated intentions to the accounts.

Markers worth tracking

Three public indicators track diffusion without relying on announcements. The cost-to-income ratio and the jaws effect, quarter by quarter, are best compared with revenue growth: a lasting gap between costs and revenues while revenues are flat would be a first sign of genuine efficiency. The ESAs’ annual DORA report will show whether the share of third-party failures, 29% in 2025, keeps rising. Visa’s biannual reports give the trend in scams, where attackers’ AI meets the banks’ own.

That leaves 2 December 2027, a deadline already pushed back once. If it holds, banks will have to apply risk management and human oversight obligations to their scoring models, and supervisors could gain a first basis for comparing those models across institutions.

Last updated — 8 October 2026