AI in Finance: The Rules Behind the Agent
DORA, NIS2, the EU AI Act and the UK rules that shape deployment. Regulatory snapshot: 17 November 2026.
Executive summary
- Classify the financial service and the decisions AI influences, then map the rules. A model supplier's assurances cannot cover the whole deployment.
- Separate customer protection, data rights and operational resilience. Passing one test does not satisfy the others.
- Build evidence into the service: decision records, effective human intervention, supplier oversight and tested recovery.
Buying an AI model does not buy permission to use it for a lending decision.
An assistant explaining a lending policy and an agent assessing a customer's creditworthiness may use the same model. Their regulatory implications are different.
For financial institutions, the useful question is: what authority are we giving this system, and what evidence makes that authority defensible?
1. DORA: can the service survive its dependencies?
The EU's Digital Operational Resilience Act has applied since 17 January 2025 to financial entities within its scope. It addresses ICT risk, incidents, resilience testing and third-party dependencies. ESMA
For an AI service, the dependency map should reach beyond the model: cloud hosting, retrieval, identity, tools and subcontractors can each interrupt delivery. Financial entities remain responsible when using ICT providers. Central Bank of Ireland
Practical consequence: ask procurement for evidence of service levels, incident cooperation, audit rights and workable exit arrangements. For services supporting critical or important functions, an exit plan should be tested, not merely attached to the contract. DORA third-party policy requirements
Switching models is only a recovery option if the replacement can perform the approved task safely. A fallback that changes lending outcomes needs validation, not just a successful API connection.
2. NIS2: map the boundary before duplicating controls
NIS2 is an EU cybersecurity directive implemented through national law. For covered financial entities, DORA takes precedence over overlapping cybersecurity risk-management and reporting requirements. Treating both as identical checklists creates confusion. European Commission
NIS2 can still matter to cloud providers, managed-service suppliers and other group entities within its scope. Check the legal entity, activity and national implementation. A bank's supplier requirements do not automatically make every supplier directly subject to NIS2.
3. The AI Act: classify the use, not the brand
The EU AI Act lists natural-person creditworthiness and life/health insurance risk assessment or pricing as high-risk uses, subject to Article 6's classification conditions. Financial-fraud detection is excepted from the creditworthiness entry, not from all regulation. Annex III, Article 6
The Commission's current timetable places Annex III high-risk obligations at 2 December 2027, following the AI Omnibus amendments. Transparency rules started applying on 2 August 2026, with specific transitional provisions. Do not interpret the high-risk extension as a general compliance holiday. Current timeline
For relevant credit and insurance deployments, prepare the fundamental-rights impact assessment alongside any required data-protection assessment. Article 27
Also determine whether the institution is a deployer or a provider. Substantial modification or repurposing can change its responsibilities. A UK location does not necessarily remove EU exposure when the system's output is used in the EU. Article 25, Article 2
4. The UK: existing rules already reach AI
The FCA says it will rely on existing frameworks rather than introduce an additional AI rulebook. Consumer Duty, governance and accountability still apply. Its October 2026 findings warn against treating human review as a box-ticking control. FCA approach, AI supervisory findings
Human oversight needs time, information and authority to change the outcome. Measure overrides, escalations and customer outcomes, not simply whether someone clicked “approve”.
For firms in scope, operational-resilience requirements already demand delivery of important business services within impact tolerances during severe but plausible disruption. FCA resilience rules
Banks within PRA SS1/23's scope also need model identification, governance and independent validation. This statement does not apply identically to every financial institution. PRA scope
5. Data rights: one platform, different decision rules
EU GDPR Article 22 restricts solely automated decisions with legal or similarly significant effects, subject to exceptions and safeguards. European Commission
The UK's Data (Use and Access) Act changed that framework. The relevant reforms commenced on 5 February 2026. Broader automated decision-making remains subject to lawful processing, safeguards and additional restrictions for special-category data. Individuals must be able to obtain information, make representations, seek human intervention and contest significant solely automated decisions. ICO commencement, Act, section 80
Design implication: shared technology can need different jurisdictional decision paths. A human rubber stamp is a weak basis for claiming that a decision is not solely automated.
Decision authority, data rights, recovery and accountability are complementary questions. Legal scope must be assessed separately.
Read the full reference map and primary sourcesWhat to build into the next release
Use one evidence record per AI-enabled service: intended purpose, affected customers, accountable owner, data permissions, model and tool versions, outcome tests, supplier dependencies, recovery and review triggers. This is an operating recommendation, not a regulatory certificate.
Watch adjacent changes too: the EU Data Act removes switching charges for covered data-processing services from 12 January 2027; CRA manufacturer reporting has already started; the UK's Cyber Security and Resilience Bill is still progressing through Parliament. Data Act, CRA, UK Bill status
The strategic advantage is a service whose decisions can be challenged, whose suppliers can be replaced and whose operation can be defended with evidence.
This reflects the regulatory developments I am paying attention to as of 17 November 2026. It is not legal or compliance advice. Regulations and interpretations change; confirm what applies to your institution, jurisdiction and use case.
REGULATORY REVIEW QUESTIONNAIRE
What should your institution
investigate next?
Answer nine questions about one AI service. Review the scope, evidence and primary sources without a compliance score or automated legal determination.
Explore the regulatory questions