A UK financial firm can read the Digital Operational Resilience Act as somebody else's law, since it is an EU regulation. That reading misses two things: DORA reaches UK firms that serve EU clients, and the UK runs its own operational resilience regime that asks the same question of your AI in different words.
The short answer
- DORA has applied across the EU since January 2025, setting rules on ICT risk, incident reporting, resilience testing and third-party oversight for financial entities.
- It reaches UK firms that operate in the EU or provide services to EU financial entities, and it covers ICT providers, which is what an external AI service is.
- The UK has its own regime: the FCA and PRA operational resilience rules, fully in force since March 2025, plus a new oversight regime for critical third parties.
- Adding AI does not create a separate duty. An AI system becomes an ICT asset and an AI vendor becomes an ICT third party, both inside the resilience rules you already meet.
- The exposure is concentration: many firms leaning on the same few AI providers is exactly the third-party risk both regimes are built to surface.
Why an EU regulation is a UK firm's problem
DORA, the Digital Operational Resilience Act, has applied since 17 January 2025. It requires financial entities to manage ICT risk, report significant ICT incidents, test their digital resilience, and manage the risk from ICT third-party providers, with an EU-level oversight framework for the most critical of those providers. Its reach does not stop at the EU border. A UK firm with EU operations, or one providing services to EU financial entities, can be caught, and because the Act governs ICT providers as well as financial entities, a UK technology or AI provider serving EU finance can find itself in scope as a third party.
This is the same market-based reach UK boards met with the GDPR and the EU AI Act, covered in who the AI Act actually binds. The presence of an EU nexus, rather than the location of the head office, decides whether DORA applies.
The UK asks the same question in its own words
A UK firm outside DORA's reach is not outside operational resilience regulation. The FCA's policy on building operational resilience and the PRA's supervisory statement set a principles-based regime built on important business services, impact tolerances, mapping of the resources each service depends on, and testing against severe but plausible disruption. That regime reached full implementation on 31 March 2025. Alongside it, the UK has stood up an oversight regime for critical third parties that provide material services to regulated firms, giving the regulators direct reach into the providers the sector concentrates on.
The two regimes differ in mechanics and share an aim: a financial firm must understand what its important services depend on, and must be able to withstand and recover when one of those dependencies fails. An AI system that sits inside an important service is one of those dependencies.
Where AI lands in the resilience picture
Adding AI to a financial service does not open a separate regulatory front. It adds to the ICT estate the resilience rules already govern, in four concrete ways.
| What AI adds | How the resilience regime already treats it |
|---|---|
| An AI system inside an important service | An ICT asset whose failure has to be mapped, tolerated and tested like any other. |
| An external model or AI vendor | An ICT third party, subject to the same due diligence, contract terms and exit planning. |
| A model that can fail in new ways | Drift, degraded output and unavailability are disruption scenarios the testing has to cover. |
| Reliance on a few large providers | Concentration risk, which is precisely what the third-party oversight regimes exist to address. |
The concentration point is the one boards underrate. When much of a sector builds on the same handful of AI providers, a single provider outage becomes a systemic event, and both DORA and the UK critical-third-party regime are designed to make that dependency visible and governed. A firm that cannot say which of its important services rely on which AI provider, and what happens if that provider fails, has a resilience gap the regulator is now equipped to find.
The constraint
The binding limit is a complete map of which important services depend on which AI systems and providers, rather than any single resilience control. Testing individual systems while that dependency map is incomplete leaves the concentration risk unseen, because the deciding question the resilience regimes ask is what your important services rely on and how they recover when a dependency fails, and only the map answers it for AI.
What a firm should do now
Treat every AI system as part of the resilience estate from the start. Record which important business services depend on it, name the third party behind any external model, and set the impact tolerance and the recovery plan for its failure. Include AI-specific failure modes, such as model drift and degraded output, in the scenario testing rather than only outright unavailability. And map the concentration: if several important services lean on one provider, that is a dependency to govern deliberately. A firm that does this is meeting the resilience rules it already answers to, whether the binding regime is DORA, the UK framework, or both at once, and the delivery of that operational picture is the same discipline set out in how AI lands inside critical-infrastructure duties.
Sources
- Regulation (EU) 2022/2554 (DORA), Digital Operational Resilience Act, applicable from 17 January 2025.
- Financial Conduct Authority, PS21/3 Building operational resilience, and the PRA supervisory statement SS1/21, fully in force from 31 March 2025.
- Bank of England, PRA and FCA, the oversight regime for critical third parties to the UK financial sector.