meilynx

Reference · European Union

DORA and AI Workloads: What Regulation (EU) 2022/2554 Requires

Digital Operational Resilience Act, applied since 17 January 2025

Last reviewed September 4, 2026

Regulation (EU) 2022/2554 of 14 December 2022, the Digital Operational Resilience Act, sets ICT risk-management, incident-reporting, resilience-testing, and third-party-risk requirements for the EU financial sector. Under Article 64 it has applied since 17 January 2025. DORA is written for ICT risk in general, and it reaches an AI workload through its definitions: an AI system is a software asset in the entity's network and information systems, and a model provider delivering an API is an undertaking providing ICT services.

Who it applies to

Article 2(1) lists the entities in scope, from credit institutions, payment institutions, and account information service providers through investment firms, crypto-asset service providers, insurance and reinsurance undertakings, and crowdfunding service providers, with ICT third-party service providers as the final point. Points (a) to (t) are collectively the financial entities. Article 2(3) carves out several categories, including occupational pension schemes with fifteen or fewer members and insurance intermediaries that are microenterprises or small or medium-sized enterprises.

Article 3 supplies the definitions that pull AI into scope. An ICT asset is a software or hardware asset in the network and information systems used by the financial entity (Article 3(7)). ICT services are digital and data services provided through ICT systems to one or more internal or external users on an ongoing basis (Article 3(21)). An ICT third-party service provider is an undertaking providing ICT services (Article 3(19)).

What it requires

  • Governance and organisation (Article 5) and a sound, comprehensive, and well-documented ICT risk management framework as part of the overall risk management system (Article 6).
  • Identification (Article 8): identify, classify, and document all ICT-supported business functions and the information and ICT assets supporting them; identify all processes dependent on ICT third-party service providers; maintain inventories and update them periodically and on every major change.
  • Protection and prevention (Article 9): continuously monitor and control the security and functioning of ICT systems, and maintain high standards of availability, authenticity, integrity, and confidentiality of data at rest, in use, and in transit.
  • Detection (Article 10): mechanisms to promptly detect anomalous activities, with alert thresholds and criteria that trigger the incident response process.
  • Incident management (Articles 17 to 19): a process to detect, manage, and notify ICT-related incidents; a record of all ICT-related incidents and significant cyber threats; and reporting of major incidents to the competent authority as an initial notification, an intermediate report, and a final report once the root cause analysis is complete.
  • Testing (Article 24): a digital operational resilience testing programme.
  • Third-party risk (Articles 28 and 30): ICT third-party risk managed as an integral component of ICT risk, a register of information covering all contractual arrangements on the use of ICT services, an annual report to the competent authority on new arrangements, and contracts whose rights and obligations are allocated in writing.

How it reaches an AI workload

A model endpoint called from a firm's applications is an ICT service, the calling system is an ICT asset, and the provider behind the endpoint is an ICT third-party service provider. Each of the runtime obligations then attaches to that traffic:

  • Article 8(4) to (6): the AI systems in use, and the processes that depend on external model providers, belong in the asset and dependency inventories, updated on every major change.
  • Article 9(2): prompts and responses are data in transit; their confidentiality and integrity are the entity's to maintain wherever the data goes.
  • Article 10(1) and (2): anomalous activity in AI systems, including changes to their configuration, needs a detection mechanism with defined alert thresholds.
  • Article 17(2): an incident involving an AI system is an ICT-related incident and must be recorded; if major, Article 19(4) reporting follows.
  • Article 28(3): the contractual arrangement with each model provider is an entry in the register of information, distinguished by whether it supports a critical or important function, and reported annually.
  • Article 30(2): the contract with a model provider is expected to cover the description of services and any subcontracting, the locations where data is processed, data protection, access and return of data on termination, assistance during incidents, and cooperation with competent authorities.
The Article 28(3) register is the obligation most often missing an AI row. Standard API terms rarely address the Article 30(2) points, and a provider onboarded by an engineering team through a self-service console may never have reached the register at all.

What a runtime record does not cover

Governance accountability (Article 5), the risk management framework (Article 6), resilience testing (Article 24), incident classification and authority reporting (Articles 18 and 19), and contractual provisions (Article 30) are organisational acts. Evidence drawn from AI traffic substantiates the AI/LLM workload slice of the identification, protection, detection, incident-recording, and register obligations. The rest of DORA is answered elsewhere in the entity's programme, and no runtime record makes an entity DORA compliant on its own.

Control mapping

What a reviewer expects to be able to see.

ObligationWhat the system must doEvidence a reviewer expects
Identification (Art. 8)Carry every AI system and its provider dependency in the ICT asset and dependency inventories, updated on major changeInventory entries for models, calling systems, and providers, with change history
Protection and prevention (Art. 9)Monitor the security and functioning of AI systems continuously and protect the data they carry in transitMonitoring records and data-protection findings for AI traffic
Detection (Art. 10)Detect anomalous activity in AI systems promptly, against defined thresholds that trigger incident responseDetection rules, thresholds, and the alerts they produced
Incident management (Art. 17, 19)Record every ICT-related incident involving an AI system and report major ones in the three-stage formatIncident register entries and, where major, the initial, intermediate, and final reports
Register of information (Art. 28(3))Record each model-provider arrangement and whether it supports a critical or important functionRegister rows for AI providers and the annual submission to the competent authority
Contractual provisions (Art. 30)Hold contracts with AI providers that allocate rights and obligations in writing and cover the Article 30(2) pointsContracts, and a gap analysis against Article 30(2) for standard API terms

Key dates

  • 14 December 2022Regulation (EU) 2022/2554 adopted; published in the Official Journal on 27 December 2022.
  • 17 January 2025DORA applies to financial entities (Article 64).
  • AnnuallyReport to the competent authority on new ICT third-party arrangements, with the full register of information available on request (Article 28(3)).

Primary sources

Common gaps

Where the AI workload most often falls outside the DORA programme.

  • Model providers absent from the register. The register of information covers all contractual arrangements on ICT services. A model provider onboarded by an engineering team on standard terms is a contractual arrangement whether or not procurement saw it.
  • An inventory built from procurement records. Article 8(6) asks for inventories updated on every major change. Switching a model version or adding a provider is a change the inventory has to catch.
  • API terms treated as a DORA contract. Article 30(2) expects locations of data processing, incident assistance, and authority cooperation in writing. Standard API terms usually address none of them.
  • An incident process that has never seen an AI event. Article 17(2) requires all ICT-related incidents to be recorded. If leaked credentials in a prompt or an unapproved configuration change have never entered the incident register, the process has a blind spot.
  • A compliance claim resting on a tool. Runtime evidence covers a slice of the identification, protection, detection, and register obligations. Governance, testing, and contracts are organisational, and a claim of DORA compliance built on telemetry alone will not survive its first supervisory question.

Last reviewed September 4, 2026. This reference summarises publicly available regulatory guidance and is provided for general information. It is not legal advice. Obligations depend on an institution's charter, registration status, size, and activities. Verify against the primary sources cited above and consult counsel before relying on any summary here.

Regulatory updates

When a regulator changes what an AI examination asks for, hear about it first.

Short notes on SR 26-2, NYDFS 500, FINRA, the NAIC bulletin, the EU AI Act, and the employment-AI statutes, plus what we ship. A few emails a month.