Short answer. DORA (Regulation (EU) 2022/2554), in force since 17 January 2025, requires every in-scope financial entity to place each production AI system inside a documented ICT business continuity policy, with response and recovery plans, backups, a defined recovery time objective, and a test at least once a year. Treat the model, its prompts, its retrieval index, and its configuration as critical assets, not just the application around them. The plan is how a bank proves it can keep the AI running, or fail over cleanly, when something breaks.
A business continuity plan for an AI system is the written, tested answer to one question: when this model stops behaving, what keeps the bank running? DORA does not treat AI as a special case. It treats a credit-scoring model or a document-processing agent as an ICT asset that supports a business function, and it asks for the same resilience evidence it expects for a core payment system.
In practice that means three things the pilot stage usually skips. First, the AI system gets a recovery time objective and a recovery point objective tied to how critical its function is. Second, every component the model depends on is backed up and restorable, not just the code. Third, the plan is tested on a schedule, and the results are written down.
What the plan has to cover looks like this:
For a bank, insurer, or telecom operator in Central Europe, two regimes now point at the same AI system. DORA has applied since 17 January 2025 and governs operational resilience. The EU AI Act adds its own obligations: Article 50 transparency and general-purpose model rules have been live since 2 August 2026, and high-risk stand-alone systems such as credit scoring and insurance pricing (Annex III) carry obligations from 2 December 2027. Article 15 of the AI Act requires high-risk systems to be accurate, robust, and resilient to faults, which is the AI Act restating in product terms what DORA asks for in operational terms.
The practical consequence is an audit. A supervisor can ask a Slovak or Czech bank to show the continuity plan for a named AI system, the date it was last tested, and the recovery time it achieved. A model that works in a demo but has no restore procedure and no test record is a finding waiting to happen. The continuity plan is how a regulated institution turns a working model into a system it can defend.
DORA (Regulation (EU) 2022/2554) has required an ICT business continuity plan for every production AI system since 17 January 2025.
DORA sets the obligations in two articles. Article 11 requires a documented ICT business continuity policy, with response and recovery plans that are specific, documented, and regularly tested, and it requires those plans to cover functions outsourced to third-party providers. Article 11(6) requires financial entities to test the business continuity plans at least once a year, and for all but microenterprises those tests must include cyberattack scenarios and a switchover between primary infrastructure and backup or redundant capacity.
Article 12 governs the data side. It requires backup policies that state what is backed up and the minimum frequency, set by how critical the data is, plus restoration and recovery procedures. It requires a recovery time objective and a recovery point objective for each function, and it requires restoration to run on systems physically and logically segregated from the source. After an incident, the entity must run checks and reconciliations so data integrity holds.
An AI system is more than its application code, and a backup that captures only the app will fail the restore test. The components that need a version and a restore path are the model weights or the pinned version of a hosted model, the prompts and system instructions, the retrieval index or vector store, the feature pipelines, and the configuration that wires them together. For a fine-tuned model, the training and evaluation data lineage belongs in scope too, because you cannot reproduce the model without it.
Each component gets a recovery point objective that matches its change rate. A vector index that updates hourly needs a tighter backup frequency than a model retrained twice a year. The test that matters is not whether the backup exists. It is whether you can restore the whole set on segregated infrastructure and get the same answers you had before the failure.
Most banks run at least part of their AI on a provider they do not control. DORA treats that dependency as part of the continuity plan, not outside it. Article 11 requires the continuity arrangements to cover functions outsourced to ICT third-party providers, and DORA’s third-party rules require an exit strategy for critical providers so the bank is never stranded.
In practice that means three defenses, decided before the outage rather than during it. A fallback path: a second provider, a self-hosted model, or a rules-based mode that keeps the critical function alive. A degraded-service definition: what the system still does when the primary model is gone, and what it stops doing. And an exit plan: the steps and the time needed to move off a provider for good. Each one is tested, because an untested fallback is an assumption.
At least once a year, under Article 11(6). A test is not a document review. For a bank above microenterprise size, DORA expects scenarios that include a cyberattack and a real switchover to backup or redundant capacity. For an AI system that means restoring the model and its index on segregated infrastructure, running the switch to the fallback provider, and measuring how long the function was degraded against its recovery time objective.
The output of a test is evidence: what was tested, when, how long recovery took, what failed, and what changed as a result. That record is what a supervisor reads in an audit, and it is what turns a continuity plan from a policy document into proof the system can be defended in production.
Accountability sits with the bank, not the vendor. Under DORA the management body is responsible for the ICT risk management framework that the continuity plan lives inside, and that responsibility cannot be handed to a supplier. The ICT risk or operational resilience function usually maintains the plan, the business owner of the AI function signs off on the recovery objectives, and the model owners keep the technical restore procedures current. A vendor can run infrastructure, but the bank owns the obligation to show the plan works.
Ableneo builds AI systems for regulated institutions in Central Europe to survive production, not just a demo. Roughly 4 of 5 AI projects we delivered in 2025 reached production, across 34 production systems, because resilience is designed in from the first architecture review rather than bolted on before an audit. For T-Mobile Czech Republic we put a document-validation system into production that runs 14 times faster than the manual process at over 90% accuracy, with the monitoring and recovery path a regulated operation needs. We treat the continuity plan as part of the build, so the model a bank ships is a model it can defend. See how we approach AI in production for regulated industries.
Key takeaways
Planning AI in a regulated business? Ableneo takes systems from classification to governed production.