What Does a Bank Do After Submitting Its ECB AI Cyber Action Plan?

7 min readAbleneo AI transformation team

Short answer. Around 110 significant institutions submitted an AI cyber action plan to the European Central Bank by 31 October 2026. The plan was the entry ticket, not the finish line. The ECB now expects 6 to 12 months of implementation: map every AI system that touches your defences or your attack surface, assign a named owner for each control, tighten third-party oversight, and build an audit trail a supervisor can read. Supervisors assess progress through the normal SREP cycle, so the measures you promised are now the ones you are graded against.

1What This Means in Practice

The ECB letter of 7 July 2026 (reference SSM-2026-0301) asked each significant institution to name concrete measures, allocate resources, assign roles, and set timelines. The deadline of 31 October 2026 covered the plan, not the completed work. The hard part starts now. A plan that reads well converts into controls that run in production, or it fails its first supervisory review.

In practice, the implementation phase breaks into four streams of work that run in parallel through the first half of 2027:

The common failure is treating the plan as a document instead of a backlog. The measure “strengthen detection against AI-generated attacks” has to become a monitored control with a dashboard, an owner, and a date. Ableneo puts 4 of every 5 AI projects into production, and that ratio comes from treating each promised control as a shippable unit of work, not a paragraph.

2Why This Matters for Regulated Industries

The ECB letter does not create a new regulation. It works through the Digital Operational Resilience Act (DORA), which has applied to financial entities since 17 January 2025, and it is assessed through the Supervisory Review and Evaluation Process. An unconvincing implementation becomes a supervisory finding, and supervisory findings feed capital and remediation demands. The letter accompanies a European Systemic Risk Board warning, adopted 25 June 2026, on systemic cyber risks from frontier AI models, so the concern sits at board level across the euro area.

For banks in Central Europe the reach is wider than the directly supervised list suggests. Slovakia and Austria are inside the euro area, so their significant institutions answer to the ECB directly. Czech banks sit under the national regulator, yet the large ones belong to euro-area groups such as Erste and KBC, so the parent-group expectation lands on them anyway. On top of the cyber letter, the EU AI Act classifies credit scoring as a high-risk use, with those obligations applying from 2 December 2027. A bank that builds one accountability model for AI cyber controls and reuses it for AI Act governance does the work once instead of twice.

The 31 October 2026 deadline was for the plan; the ECB now expects 6 to 12 months of implementation, assessed through SREP.

3What Does the ECB Actually Check After Submission?

The ECB has said it will analyse the submitted plans horizontally, comparing them across institutions to find common gaps. Expect follow-up questions from your Joint Supervisory Team that benchmark your plan against peers, not against an absolute standard. A plan that promised less than comparable banks, or that has no evidence of progress by the first review, is the one that draws scrutiny.

Supervisors look for three things: that the measures you named are actually moving, that the owners you assigned exist and are accountable, and that you can show evidence rather than assertion. The question is not “do you have a policy” but “show me the last time this control caught something, and who acted on it”. Prepare for the horizontal review by keeping a live progress record against each measure, with dates and proof, from the first week of implementation.

4How Do You Turn an Action Plan Into Production Controls?

Start with the asset register, because every other stream depends on it. You cannot oversee a vendor, assign an owner, or log a decision for an AI system you have not catalogued. List each model and agent, where it runs, what data it touches, and whether it sits on the defensive or the exposed side of the bank.

Then convert each plan measure into a control with three fixed attributes: an owner who can be named in an audit, a monitoring signal that proves it works, and a review date. A control without a signal is a hope. This is the discipline that separates a pilot from a production system, and it is the same discipline whether the AI runs a fraud model or a cyber-detection agent. Ableneo shipped 34 production AI projects in 2025 across 20 clients using exactly this conversion from intent to monitored control.

5Who Signs Off on AI Cyber Controls in a DORA Audit?

Under DORA the management body holds final accountability for the ICT risk framework, and that does not change because the risk is AI-driven. The board cannot delegate the accountability, only the execution. In practice each AI cyber control needs a first-line owner who runs it, a second-line risk function that challenges it, and a management-body member who answers for it to the supervisor.

The audit trail is what makes that accountability real. When a Joint Supervisory Team asks who approved a model update that changed the bank’s detection behaviour, the answer has to be a record, not a recollection. Version every model, log every material change with the approver’s name, and keep the evidence where an auditor can find it without a scavenger hunt.

6What Should a Mid-Market Bank Do First?

A smaller institution cannot staff every stream at once, so sequence by exposure. First, catalogue the AI systems that sit on the attack surface or inside the detection path, because those carry the risk the ECB letter is about. Second, fix third-party oversight for the AI vendors you already depend on, since that reuses DORA controls you have in place. Third, formalise the accountability model before you scale the control count.

Proportionality is on your side. DORA and its supervisory expectations scale with the size and risk profile of the entity, so a mid-market bank is not held to the control volume of a global systemically important bank. The goal is a defensible, owned, evidenced set of controls for the AI that matters most, not a complete catalogue on day one. Doing three streams well beats doing six badly.

7The Ableneo Perspective

Ableneo works inside regulated financial institutions in Central Europe, including banks and insurers facing exactly this supervisory cycle, and our record is production, not slideware: 4 of 5 AI projects reach live operation, 34 shipped in 2025 alone. That track record matters here because an ECB action plan is judged on whether the promised controls run, not on whether they were well described. Our work turns each measure into an owned, monitored, auditable control, and connects it to the DORA framework the bank already operates, so one governance model covers both the cyber letter and the AI Act obligations ahead.

Key takeaways

Sources

Planning AI in a regulated business? Ableneo takes systems from classification to governed production.

Talk to Ableneo