AI for financial services
DORA, supervisory scrutiny and client data protection.
Research
Context
Banks, insurers and asset managers face the strictest ICT-risk regime in Europe. Supervisors expect clear answers: where AI runs, who can access client data and how to exit a provider.
Supervisors ask where each AI system runs, who can reach client data and how the institution would exit. Lindstead gives the board an independent, sourced answer and builds the first use case in-house.
Case study (anonymised): An insurer's AI assistant brought under DORA control A European non-life and pensions insurer, around 1,500 staff.
Need a first decision? The AI Sovereignty Scan covers readiness, regulation and hosting in ten working days.
AI is moving into credit decisions at large EU banks
ECB Banking Supervision collects yearly technology data from the banks it supervises directly. By 2024 nearly all used AI. What changed is what they use it for.
The share of ECB-supervised banks using AI for credit scoring more than doubled in 2024. Overall AI use barely moved.
Share of significant institutions supervised by the ECB using AI, in total and for two use cases, %
¹ Share of significant institutions reporting AI use in ECB supervisory data on innovative technologies: 107 banks in 2023, 110 banks in 2024. ² AI systems used to detect financial fraud are explicitly excluded from the high-risk credit category of the EU AI Act (Annex III, point 5(b)). ³ AI systems used to evaluate the creditworthiness of natural persons or establish their credit score are classified as high-risk under the EU AI Act (Annex III, point 5(b)).
Source: ECB Banking Supervision, Supervision Newsletter: AI's impact on banking, 20 November 2025
View data table
| Category | Series 1 | Series 2 |
|---|---|---|
| Any AI use¹ | 87% | 88% |
| Fraud detection² | 36% | 62% |
| Credit scoring³ | 14% | 30% |
Series 1: 2023. Series 2: 2024. Unit: %
Challenges
Credit scoring use rose from 14% to 30% in one year, and credit scoring of natural persons is high-risk under the EU AI Act. In September 2026 the ECB reported more than 90% of directly supervised banks integrate AI and 85% use generative AI. Boards now need answers on validation and explainability.
- Concentration on critical providers In November 2025 the European Supervisory Authorities designated 19 critical ICT providers, including AWS, Google Cloud and Microsoft. Building AI on these platforms deepens a dependency DORA already targets.
- AI exit plans on paper DORA requires a tested exit strategy for every ICT service behind a critical or important function. An assistant locked into one provider's stack has no alternative to test.
- Credit and pricing become high-risk From 2 December 2027, the EU AI Act treats AI that scores individuals' creditworthiness or prices life and health insurance as high-risk. Systems bought now must comply.
- Model risk meets language models EBA guidelines require banks to understand their automated credit models and keep outputs traceable. EIOPA set similar expectations for insurers in August 2025. An API model the provider can change or retire undermines that evidence.
- Client data and foreign access Client files and claims data are confidential and fall under GDPR. The US CLOUD Act can force a provider under US jurisdiction to hand over data it controls, wherever it is stored.
Fraud detection leads insurers' generative AI plans
EIOPA surveyed 347 insurers in 25 EU and EEA countries on generative AI as of the second quarter of 2025. Nearly two-thirds already use it, unevenly.
European insurers use generative AI most in customer service today. Fraud detection leads their plans for the next three years.
Share of insurance undertakings using or planning to use generative AI, by area of the insurance value chain, %
¹ Share of 347 responding insurance undertakings (148 non-life, 87 life, 112 both) in 25 EU and EEA countries, status as of Q2 2025. Self-use by staff of publicly available generative AI tools is out of scope. The remaining share in each area reported no plans to use generative AI. ² AI systems used for risk assessment and pricing in relation to natural persons in life and health insurance are classified as high-risk under the EU AI Act (Annex III, point 5(c)).
Source: EIOPA, Generative AI Market Survey: Outlook, Use Cases and Risk Management, February 2026 (survey July 2025)
View data table
| Category | Series 1 | Series 2 |
|---|---|---|
| Customer service | 40% | 48% |
| Claims | 32% | 59% |
| Sales and distribution | 22% | 54% |
| Pricing and underwriting² | 19% | 52% |
| Fraud detection | 15% | 64% |
| Product development | 8% | 51% |
Series 1: Using generative AI now¹. Series 2: Plan to use within 3 years. Unit: %
Key regulation
Today's use sits in customer service and claims, where staff check the output, and 64% of reported use cases are still proofs of concept. Plans shift to fraud detection, pricing and underwriting: decisions about individual policyholders. Life and health pricing is high-risk under the EU AI Act.
| Framework | Relevance |
|---|---|
| DORA | ICT third-party risk, contract and exit-strategy rules, applicable since 17 January 2025. |
| EU AI Act | Creditworthiness assessment and life and health insurance pricing are high-risk uses, with obligations from 2 December 2027. |
| GDPR | Applies in full to client data in AI systems. |
How Lindstead helps
- AI strategy AI workload map for the supervisor Lindstead rates each AI use case on client-data sensitivity, AI Act risk class and DORA criticality, and recommends where it runs: in-house, European cloud or external API.
- Regulation and risk AI systems mapped to DORA and the AI Act Lindstead links each AI system to the DORA register, its AI Act role and risk class, and GDPR Article 22, with dates and a draft exit strategy per critical service.
- Model selection Open-weight shortlist for model risk Lindstead tests open-weight models on the institution's own documents and languages, checks licence and origin, and records the evidence for model risk teams.
- Infrastructure and cost Three-year cost case versus API spend Lindstead sizes the GPUs needed and compares owned servers, European cloud GPUs and the current API contract, showing every assumption and the break-even volume.
- Deployment First use case in production Lindstead puts one internal use case, such as a policy assistant, into production on infrastructure the institution controls. Access, logging and evaluation are documented for audit and the supervisor.
- Model stewardship Quarterly AI note for the risk committee Each month Lindstead tracks new models, licence changes and ECB, EBA, EIOPA and ESMA publications. The risk committee gets a quarterly note with actions.
Case study (anonymised)
An insurer's AI assistant brought under DORA control
A European non-life and pensions insurer, around 1,500 staff
Staff used an AI assistant from a US cloud provider, and the claims department wanted it on claims files. The risk function could not show where policyholder data went, how to exit or which AI Act duties applied.
The board wanted a decision before the next budget.
What Lindstead did
- Inventory and data flows Lindstead listed every AI system in use or planned and linked each to the DORA register. It traced where policyholder data went and which jurisdictions could reach it.
- Regulatory and exit mapping Lindstead mapped each system to its AI Act risk class, GDPR basis and DORA criticality. For the assistant, it named alternatives and drafted a testable exit plan.
- Models and cost Lindstead tested two or three open-weight models on anonymised claims documents and checked their licences. A three-year cost model compared the current service, European cloud GPUs and owned hardware.
- Board recommendation Lindstead wrote a decision paper with options, costs, risks and a phased roadmap. The chief risk officer and internal audit reviewed it first.
What the client received
- AI system register linked to the DORA register
- Regulatory map with AI Act, GDPR and DORA dates
- Tested open-weight shortlist with licence review
- Three-year cost model with break-even volume
- Board decision paper and phased roadmap
The decision
The board decided which AI workloads move to infrastructure the insurer controls and which stay with the current provider under a tested exit plan.
It also decided on funding a first in-house deployment.
Data and model risks top AI concerns in EU capital markets
Between June and September 2025, ESMA and 15 national authorities asked 728 securities market firms to name their main AI challenges.
Data protection, data quality and hallucinations top the AI worries of EU securities market firms.
Share of responding securities market firms citing each item as a main challenge from their use of AI, %
¹ Share of respondents, multiple answers possible; 345 of 728 entities did not answer this question. Respondents: 274 investment management companies, 262 investment firms, 106 credit institutions providing investment services, 77 market infrastructures, 8 credit rating agencies and 1 benchmark administrator in 19 EU countries. Participation was voluntary and the sample is weighted towards Portugal, Malta and Ireland, so results are not necessarily representative. ² Inaccurate or fabricated output from AI models. Hallucinations are also the top-cited generative AI risk among EU insurers in EIOPA's 2025 survey. ³ Microsoft was named as the leading AI provider by 47% of the 344 firms that named one, followed by OpenAI (20%) and AWS (8%). Only about 8% of all providers named are domiciled in the EU.
Source: ESMA, TRV Risk Analysis: AI adoption and trends in securities markets, EU evidence, 20 February 2026
View data table
| Category | Value | Group |
|---|---|---|
| Data protection¹ | 59% | 1 |
| Data quality | 55% | 1 |
| Hallucinations² | 51% | 1 |
| Data governance | 46% | 1 |
| Cyber security | 46% | 2 |
| Compliance with regulation | 45% | 2 |
| Third-party dependencies³ | 37% | 2 |
| Lack of skills or resources | 29% | 2 |
| High costs | 19% | 2 |
Group 1: Data and model risks. Group 2: Operational, regulatory and resource risks. Unit: %.
Opportunities
Firms worry most about their data and the reliability of AI output. Only 19% cite cost. Microsoft is the leading AI provider for 47% of firms that named one. Only about 8% of named providers are domiciled in the EU. Data governance and DORA exit planning belong on asset managers' board agendas.
- Claims file triage Summarising and routing claims documents for a claims handler, who decides. Claims files often hold health data, which favours a model inside the insurer's own environment.
- Credit file preparation Extracting and checking figures from annual accounts and borrower documents for a credit analyst, who decides. With the data in-house, the file stays traceable.
- KYC and AML casework Assembling due diligence files and drafting case summaries for compliance officers. Identity and transaction data are highly sensitive, which suits in-house or European hosting.
- Policy and regulatory search An internal assistant for procedures, supervisory letters and regulatory texts. Confidential documents and steady use make own infrastructure easy to size.
Frequently asked questions
-
The main frameworks are:
- DORA: ICT third-party risk, contract and exit-strategy rules, applicable since 17 January 2025.
- EU AI Act: Creditworthiness assessment and life and health insurance pricing are high-risk uses, with obligations from 2 December 2027.
- GDPR: Applies in full to client data in AI systems.
-
Organisations in this sector most often face three issues:
- Concentration on critical providers: In November 2025 the European Supervisory Authorities designated 19 critical ICT providers, including AWS, Google Cloud and Microsoft. Building AI on these platforms deepens a dependency DORA already targets.
- AI exit plans on paper: DORA requires a tested exit strategy for every ICT service behind a critical or important function. An assistant locked into one provider's stack has no alternative to test.
- Credit and pricing become high-risk: From 2 December 2027, the EU AI Act treats AI that scores individuals' creditworthiness or prices life and health insurance as high-risk. Systems bought now must comply.
- Model risk meets language models: EBA guidelines require banks to understand their automated credit models and keep outputs traceable. EIOPA set similar expectations for insurers in August 2025. An API model the provider can change or retire undermines that evidence.
- Client data and foreign access: Client files and claims data are confidential and fall under GDPR. The US CLOUD Act can force a provider under US jurisdiction to hand over data it controls, wherever it is stored.
-
The use cases with the clearest value are:
- Claims file triage: Summarising and routing claims documents for a claims handler, who decides. Claims files often hold health data, which favours a model inside the insurer's own environment.
- Credit file preparation: Extracting and checking figures from annual accounts and borrower documents for a credit analyst, who decides. With the data in-house, the file stays traceable.
- KYC and AML casework: Assembling due diligence files and drafting case summaries for compliance officers. Identity and transaction data are highly sensitive, which suits in-house or European hosting.
- Policy and regulatory search: An internal assistant for procedures, supervisory letters and regulatory texts. Confidential documents and steady use make own infrastructure easy to size.
-
Yes. Open-weight models can run on-premise or in a European private cloud, which keeps data and model access under the organisation's control. Whether that is cheaper than an external API depends on sustained volume, and whether it is required depends on data sensitivity and regulation. Lindstead's break-even analysis shows the volume at which own infrastructure pays off.
-
Yes. AI for credit scoring of natural persons is high-risk under Annex III, point 5(b), except for fraud detection. So is risk assessment and pricing of individuals in life and health insurance, under point 5(c). After the Digital Omnibus, Regulation (EU) 2026/1744, obligations apply from 2 December 2027. Deployers must also carry out a fundamental rights impact assessment under Article 27.
-
Yes, if the service supports a critical or important function. DORA Article 28(8) requires documented, sufficiently tested and periodically reviewed exit strategies, with alternatives and transition plans. The contract also goes in the register of information under Article 28(3). An open-weight model on infrastructure the institution controls is one concrete alternative.
-
Yes. In Case C-634/21 of 7 December 2023, the Court of Justice held that a credit score is an automated individual decision under Article 22 GDPR when it plays a determining role in granting credit. Such decisions are prohibited in principle. Where an exception applies, safeguards such as human intervention are required. Lenders need real human review and records of how scores are used.
Related research
- European cloud GPU prices and provider ownership
- Open-weight models a bank can run on its own servers
- Self-hosting or API: where is the break-even point?
Sources
- EIOPA, European Supervisory Authorities designate critical ICT third-party providers under DORA, 18 November 2025
- ESAs, List of designated critical ICT third-party service providers, November 2025
- EIOPA, Digital Operational Resilience Act (DORA), applicable since 17 January 2025
- EUR-Lex, Regulation (EU) 2022/2554 (DORA), Article 28, 14 December 2022
- European Commission, AI Omnibus enters into force, 27 July 2026
- European Commission AI Act Service Desk, Annex III: high-risk AI systems, points 5(b) and 5(c)
- European Commission AI Act Service Desk, Article 27: fundamental rights impact assessment
- EBA, Guidelines on loan origination and monitoring (EBA/GL/2020/06), section 4.3.4, 29 May 2020
- EIOPA, EIOPA publishes Opinion on AI governance and risk management, 6 August 2025
- Court of Justice of the EU, Press release 186/23, SCHUFA Holding (Scoring), C-634/21, 7 December 2023
- US Department of Justice, CLOUD Act white paper, April 2019
Discuss the priorities in financial services.
A 30-minute meeting covers the use case, sector requirements and next step.