Short answer. DORA Article 28(8) has required every bank in the EU to hold a documented, tested exit strategy for each ICT service that supports a critical or important function since 17 January 2025. An AI vendor that scores credit, flags fraud, or reads documents almost always sits inside that scope. A usable exit strategy names the trigger events, sets a transition window, specifies a data and model handover in open formats, and defines an in-house or alternate-provider fallback the bank has run at least once.
An exit strategy reaches past the termination clause in the contract: it is an operational plan the bank can execute under pressure, when a vendor fails, when service quality drops, or when a supervisor tells the bank to leave. DORA treats the plan and the contract as two separate obligations. The contract gives the legal right to leave. The exit strategy proves the bank can actually do it without breaking the critical function the AI service supports.
For an AI service the plan has to carry more than a data export. A credit-scoring model is a bundle of code, training data, feature pipelines, prompt and policy configuration, monitoring thresholds, and the human review process wrapped around it. If any one of those lives only inside the vendor, the bank cannot move. A complete AI exit plan lists each of these and states, for each, how the bank gets it back in a format a second team can run.
Ableneo ships roughly 4 of 5 of its AI projects into production across banking, insurance, and telecom in Slovakia, the Czech Republic, and Austria, and designs each system so the client’s own team can run it. That handover discipline is the same discipline an exit strategy demands.
DORA, Regulation (EU) 2022/2554, has applied to banks, insurers, and most other EU financial entities since 17 January 2025. Exit strategies for ICT services supporting critical or important functions are a standing obligation, not a future one. The implementing rules in Commission Delegated Regulation (EU) 2024/1773 go further and require a documented exit plan for each such contractual arrangement, reviewed and tested on a regular basis. A supervisor can ask to see that plan in any inspection.
AI concentration makes the obligation sharper. Many CEE banks route several AI use cases through one or two model providers or one cloud region. When the same vendor sits behind fraud detection, document processing, and a customer assistant, a single failure touches more than one critical function at once. The EU AI Act adds a parallel pull: a credit-scoring model is a high-risk system under Annex III, and the deployer stays accountable for it even when a third party built it. An exit strategy is how a bank keeps that accountability when it has to change suppliers.
DORA Article 28(8) has required a documented, tested exit strategy for every critical ICT service, AI vendors included, since 17 January 2025.
Article 28(8) requires financial entities to put exit strategies in place for ICT services that support critical or important functions. The strategy has to let the bank leave the contract without disrupting its business, without breaching regulatory requirements, and without a loss of service continuity or quality. The regulation names the scenarios the plan must cover: failure of the provider, deterioration of service quality, business disruption from inappropriate or failed provision, and material risk to the continuous delivery of the service.
The plan cannot sit on paper. Article 28(8) and the RTS both require that exit plans be tested and reviewed on a regular basis, with transition arrangements that give the bank a realistic amount of time to move. In practice this means a bank needs more than a signed document. It needs evidence that the handover formats work, that a receiving team can load the model and data, and that decisions keep flowing during the switch.
With a database or a payments gateway, exit is largely a data-migration problem. With an AI service, the data is only the start. The behavior of the system lives in the model weights, the training set, the feature engineering, the prompts and guardrails, and the thresholds that decide when a human steps in. A bank that exports only its raw records and its predictions keeps the inputs and the outputs but loses the decision logic in between.
Two traps appear most often. The first is a model the bank cannot re-run because the vendor never handed over the training data or the evaluation sets, so a new team cannot reproduce or validate the results. The second is silent lock-in through a proprietary API: the bank consumes scores through an endpoint and never holds a version it could host elsewhere. A workable AI exit plan closes both by requiring periodic delivery of a reversibility pack and by keeping at least one portable path to the same decision.
Testing an AI exit plan means a rehearsal, not a tabletop read-through. The bank takes the reversibility pack it holds and asks a separate team, internal or a second supplier, to stand the service up and reproduce a known set of decisions within an agreed error margin. If the second team cannot match the original scores on a reference sample, the plan has failed and the gap is found in a drill rather than in a crisis.
A sound rehearsal checks four things: the handover formats open and load, the model produces the same outputs on a frozen test set, the fallback process carries the decision volume, and the transition fits inside the contracted window. Run it at least once a year and after any material change to the model, the data, or the vendor relationship. Record the result, because that record is what a supervisor expects to see.
An exit strategy with no owner is a document, not a control. Under DORA the accountability sits with the management body, which cannot delegate its responsibility for ICT risk. Below that, day-to-day ownership usually lands with the ICT third-party risk function, working with the business owner of the critical function and with procurement. For an AI service, model risk management and the data team join, because they hold the knowledge of what a complete handover looks like.
The practical rule is one named owner per critical AI service, with a review date and a tested plan on file. When the owner changes, the plan moves with the role. A bank that spreads the obligation across four functions with no single accountable name is the bank that cannot produce a current, tested plan when the supervisor asks.
DORA and the EU AI Act hit the same AI system from two directions. DORA governs it as an ICT service and demands the exit strategy. The AI Act governs it as a high-risk system and demands technical documentation, logging, human oversight, and a deployer that stays accountable. The two obligations reinforce each other. The reversibility pack a bank builds for its DORA exit plan, model, data, configuration, and documentation, is most of the technical file the AI Act already asks a provider to keep. A bank that builds one well has most of the other.
Ableneo builds AI that reaches production in regulated environments, with roughly 4 of 5 projects going live across banking, insurance, and telecom in Slovakia, the Czech Republic, and Austria. The design principle behind that record is that the client’s own team can run, validate, and if needed replace the system, which is the exact capability a DORA exit strategy tests. For a bank that wants the regulatory plan and the engineering reality to match, the work is to make reversibility a build requirement from day one, not a document assembled after the contract is signed. More detail on how Ableneo approaches AI in regulated industries is on the Ableneo site.
Key takeaways
Planning AI in a regulated business? Ableneo takes systems from classification to governed production.