Banks and insurers in the EU meet DORA and the EU AI Act on the same systems, the same vendors and the same teams. The DORA AI Act overlap is not a problem to manage twice. It is one control layer: a registry, a policy gateway, an evidence log, tiered oversight and a kill switch that answer both rulebooks from one piece of infrastructure. This page shows how.
Since 17 January 2025, every financial entity in the EU has lived under DORA, Regulation (EU) 2022/2554: ICT risk management, incident reporting, resilience testing, a register of information for ICT third-party providers, contractual requirements for critical or important functions, ESA oversight of critical providers. None of it was written with language models in mind; all of it applies to them. The EU AI Act, Regulation (EU) 2024/1689 as amended by the Digital Omnibus on AI, looks at the same systems from the other side: DORA asks whether the system keeps running and whether you can prove what it did, the AI Act asks what it decides about people and whether that was controlled. I am an engineer, not a lawyer. What follows is my reading as of November 2026, to my knowledge, no legal advice: verify it with counsel.
Why banks and insurers face both regulations on the same systems
An AI system in a bank is an ICT system under DORA. If the model comes from a vendor, that vendor is an ICT third-party service provider and belongs in the register of information. If the function is critical or important, the contract needs the clauses DORA prescribes and the arrangement needs an exit strategy.
The AI Act adds classification by purpose. Three common use cases are high-risk under Annex III: creditworthiness assessment of natural persons (point 5(b)), risk assessment and pricing in life and health insurance (point 5(c)) and employment decisions (point 4). For those, the provider owes risk management per Art. 9, logging per Art. 12, human oversight per Art. 14, robustness per Art. 15 and serious-incident reporting per Art. 73; the deployer owes the duties of Art. 26. For everything else, Art. 4 AI literacy and Art. 50 transparency still apply.
Prohibitions and AI literacy apply since 2 February 2025, GPAI obligations since 2 August 2025, transparency duties under Art. 50 since 2 August 2026. The Digital Omnibus on AI (Regulation (EU) 2026/1744) moved the high-risk obligations: Annex III systems from 2 December 2027, Annex I products from 2 August 2028. DORA has applied in full since January 2025. That is the practical point: a bank that builds its AI controls into the DORA framework it already operates is ready for December 2027 without a second programme.
The AI Act expects exactly that. Art. 9(10) lets providers fold AI risk management into existing risk procedures, Art. 17(4) treats a financial institution’s internal governance as its quality management system, Art. 26(5) and (6) let financial institutions meet monitoring and log-keeping duties through documentation they already keep, and Art. 74(6) makes the financial supervisor the market surveillance authority. Whoever reads your DORA evidence will read your AI Act evidence too.
The DORA AI Act mapping: one table, two rulebooks
This is the table from our first workshop with risk, compliance and IT. Right column: the component of our governance layer that produces the evidence for both.
| DORA requirement | EU AI Act article | Governance layer component that serves both |
|---|---|---|
| ICT risk management framework (Art. 6 DORA) | Art. 9 risk management across the lifecycle; Art. 72 post-market monitoring | Registry record with owner, purpose, risk class, scope and expiry; quality gates; re-evaluation forced by expiry |
| Incident management and reporting of major ICT incidents (Art. 17 and 19 DORA) | Art. 73 provider reports serious incidents; Art. 26(5) deployer informs provider and authorities | Monitoring on gateway and evidence log; alerts to the business owner; incident record linked to decision IDs |
| Third-party risk, register of information, key contractual provisions (Art. 28 and 30 DORA) | Art. 25 role change by rebranding, modification or repurposing; Art. 25(4) written agreement with suppliers; Art. 26 deployer duties | System registry: one record per model and vendor with provider, versions, data zones, contract reference; feeds the register of information |
| Resilience testing and threat-led penetration testing (Art. 24 to 26 DORA) | Art. 15 accuracy, robustness and cybersecurity, including resilience against manipulated inputs | Quality gates: business-owned evaluation set, regression and adversarial tests, bias checks, three-way sign-off |
| Logging, detection and traceability of ICT activity (Art. 9 and 10 DORA) | Art. 12 automatic recording of events; Art. 26(6) deployer keeps the logs | Evidence log, append-only per decision: input hash, model and version, sources, policy version, output, signals, reviewer action |
| Business continuity, recovery and exit strategies (Art. 11 and 28 DORA) | Art. 14 oversight including the ability to interrupt or stop; Art. 15(4) backup or fail-safe plans | Kill switch held by the business owner, read by the gateway on every call; registered fallback: second model, rules engine or manual queue |
The right column repeats itself: registry, gateway, log, gates, kill switch. That is the point. The components in detail are on my AI Governance page.
The ESAs’ statement of 31 July 2026 on frontier AI models
On 31 July 2026, EBA, EIOPA and ESMA published a joint statement on ICT risks from frontier AI models. It calls for a cross-sectoral, risk-based and consistent supervisory approach and points to the AI Act’s framework for general-purpose AI models with systemic risk as a complement to DORA. It is a non-binding statement of supervisory expectation that creates no new legal duties, but it tells you where supervisors are looking. Law firm commentary reports that ECB-supervised significant institutions had to submit action plans on frontier AI model risks by 31 October 2026. I state that as reported, not verified.
What it means for model vendors
- Expect to be treated as an ICT third-party provider. If enough financial entities depend on your model for critical or important functions, designation as a critical provider under DORA, with ESA oversight attached, is on the table.
What it means for banks and insurers
- Put the frontier model into the register of information now, with the function it supports, its criticality and the subcontracting chain behind the API. A few providers behind hundreds of institutions is the concentration pattern DORA’s oversight regime was written for.
- Treat the model as a component you cannot inspect. Scope enforced in the gateway, evaluation on your own cases, logs you own, a tested exit. The statement does not require this. A supervisor reading your evidence will expect it.
A worked example: an insurer, a vendor LLM and claims summaries
An insurer wants a hosted language model to summarise incoming claims for its handlers: notification, attachments and prior correspondence on one page. The model decides nothing; the handler decides.
Under DORA
- Register entry. The vendor goes into the register of information with the service, the data location, the subcontractors behind the API and the function it supports. Claims handling will in most insurers be an important function.
- Contract clauses. Service levels, incident notification, audit rights, data location, subcontracting, termination and exit. Most standard LLM contracts lack several of these; the gap analysis is usually the first deliverable.
- Testing and continuity. What happens when the API is down, slow or wrong, and whether the fallback, here the manual summary plus a second registered model, carries the volume.
Under the EU AI Act
- Classification first. Summarising property or motor claims for a human handler is not an Annex III use case as I read it. If the same assistant feeds risk assessment or pricing in life or health insurance, point 5(c) applies. The registry record carries the classification and the reasoning.
- Art. 50 transparency. Where generated text reaches the policyholder, or the policyholder interacts with the assistant, people must be told they are dealing with AI. For a purely internal summary the duty does not bite; we mark generated text as generated anyway.
- Art. 26 deployer duties. They apply in full only to high-risk systems: use per the provider’s instructions, competent people assigned to oversight, relevant input data, monitoring, log retention. We take them as the design target regardless, so the system can become high-risk tomorrow without a rebuild.
What the evidence log holds
Per summary: the hash of the pseudonymised input, the claim ID, the model and prompt version, the document versions the model read, the policy version, the summary, the signals and the handler’s action. Names never reach the model: IDs in, IDs out. The test is the same under both regulations: nine months later, can an auditor reconstruct why this handler saw this summary without asking anyone? If yes, DORA’s traceability, Art. 12 and Art. 26(6) are answered by one record.
A DORA AI Act twelve-month plan
The sequence I recommend to an institution that has DORA running and AI systems appearing faster than the register can absorb them.
- Months 1 to 3: inventory and alignment. Find every AI system in use, including the ones IT does not know about. Classify each under the AI Act, tag its DORA criticality, and give the system registry and the register of information one source of truth.
- Months 4 to 6: the only door. Put the policy gateway in front of every model call: no registration, no route. Run the contract gap analysis for every vendor model. Map AI incidents into the existing DORA incident process.
- Months 7 to 9: evidence and testing. Switch on the evidence log per decision. Set the oversight tier per system. Add adversarial tests and fallback drills to the resilience testing programme.
- Months 10 to 12: dry run. Reconstruct one decision from each high-risk candidate from the log as a supervisor would. Close the gaps. Plan the Annex III backlog for 2 December 2027 with dates and owners.
The AI Act side in depth, Art. 9 to Art. 72, is on my EU AI Act consultant page and in the article hub.
Typical mistakes
- Two programmes, two registers. DORA lives in operational risk, the AI Act in compliance, and nobody notices that both describe the same vendor model until the supervisor asks.
- A fallback on paper. “We would switch providers” is a sentence, not an exit strategy.
- Waiting for December 2027. The Omnibus moved the high-risk dates; it removed no obligation. DORA already applies, and the components that satisfy it are the ones the AI Act will ask for.
About Umair Zaffar
I am the founder and CEO of sifamo GmbH, an AI consultancy for regulated industries with offices in Berlin and Frankfurt am Main. Founded in 2019, self-financed and without loans, 28 people today. I come from enterprise architecture and spent more than ten years building large platforms for banks, insurers and the public sector before we focused on AI in production in 2023. Our clients include Commerzbank, VW Bank, Bosch, DHL, Zurich, VGH and the German federal administration. We work remotely across the EU from Frankfurt with on-site phases, in English and German, and with Gulf institutions that serve EU customers (see AI governance consultant in Dubai). More on my AI expert page and my founder page.
Frequently asked questions
Does DORA already cover what the AI Act will ask for?
Partly. DORA covers resilience, incidents, third parties and traceability, not whether a decision about a person was fair, explainable or overseen by a competent human. The same registry, log and gates carry the extra fields the AI Act needs: risk class, intended purpose, evaluation results, oversight tier.
Who reports an incident, the vendor or the bank?
Under DORA, the financial entity reports major ICT-related incidents to its competent authority; the vendor notifies the entity, which is why that clause belongs in the contract. Under the AI Act, Art. 73 places serious-incident reporting on the provider of a high-risk system and Art. 26(5) obliges the deployer to inform the provider and the authorities. Both paths start from the same alert.
Does this apply to a bank outside the EU with EU customers?
DORA applies to financial entities authorised in the EU. The AI Act reaches further: under Art. 2(1)(c) it covers providers and deployers in third countries where the output is used in the Union, so a Gulf bank scoring EU customers is in scope. Details: EU AI Act for UAE and GCC companies.
Want one control layer that your DORA auditor and your AI Act supervisor can both read? Reach me via the contact page, by e-mail at contact@sifamo.com or on LinkedIn. Background: my guide to the EU AI Act and the governance layer deep dive.