For a bank's AI agents, the EU AI Act and DORA overlap on risk management, logging, security, incidents and vendors. The AI Act asks whether the agent is safe and fair for the people it affects; DORA asks whether it is resilient and controlled as ICT. Run one agent register, one control library tagged to both and one incident process.
What does each regime require, in plain words?
The EU AI Act (Regulation (EU) 2024/1689) entered into force on 1 August 2024 and applies in phases. For banks, its sharpest edge is Annex III, which lists creditworthiness assessment and credit scoring of natural persons as high-risk. High-risk systems bring provider duties on risk management (Article 9), data governance (Article 10), logging (Article 12), information to deployers (Article 13), human oversight (Article 14) and accuracy, robustness and cybersecurity (Article 15). Deployers carry Article 26 duties and, for credit scoring, an Article 27 fundamental rights impact assessment. Transparency under Article 50 and AI literacy under Article 4 apply more widely. The bank-specific view is in EU AI Act for banks deploying AI agents.
DORA (Regulation (EU) 2022/2554) applies from 17 January 2025. It requires an ICT risk management framework, ICT incident management and reporting, resilience testing and management of ICT third-party risk. For background, see what DORA is.
Who has to meet both?
Banks and other financial entities in the EU that use AI agents. Every agent is within DORA as an ICT asset. Only some are high-risk under the AI Act, but all carry its transparency and literacy duties where relevant. Under the AI Act, a bank using a vendor's agent is a deployer, a role explained in what a deployer is under the EU AI Act; under DORA, the same vendor is an ICT third-party provider.
The AI Act also recognises that banks are already heavily governed. In substance, it allows certain duties, such as aspects of quality management and deployer monitoring, to be met through the internal governance arrangements financial services law already requires, and it generally leaves supervision of AI used by financial institutions with their financial supervisors. That is a strong signal to integrate rather than duplicate.
Where do the two overlap for an agent that acts?
Five places.
Risk management. AI Act Article 9 asks for a risk management system across the lifecycle of a high-risk system; DORA asks for an ICT risk framework covering every asset. One risk register, with agent capabilities as entries, can serve both if each risk is tagged to the regimes it concerns.
Records. Article 12 asks for automatic logs to trace operation; DORA needs detection, incident reconstruction and evidence for testing. One append-only action record, capturing each step and verdict, meets both.
Oversight and access. Article 14 asks for effective human oversight; DORA asks for access limited to approved functions and controlled privileged actions. Holding high-impact agent actions for a named person satisfies both at once.
Security. Article 15 asks for cybersecurity appropriate to the risks, including manipulation of inputs; DORA asks for protection and testing. One testing programme with agent scenarios, such as injected instructions, serves both.
Incidents and vendors. The AI Act asks deployers to report serious incidents and inform providers; DORA asks for major ICT incident reporting and detailed third-party management. One incident process can produce both reports, and one vendor file can hold AI Act provider information alongside DORA due diligence.
How do you run one programme?
| Requirement | Control | Evidence it leaves |
|---|---|---|
| AI Act risk management and DORA ICT risk framework | One register of agent capabilities with risks tagged to each regime | Register with regime tags |
| AI Act logging and DORA detection and reconstruction | Append-only action record of every step and verdict | Action record per case |
| AI Act human oversight and DORA privileged access | High-impact actions held for a named person; tested stop | Held-action records and stop tests |
| AI Act cybersecurity and DORA testing | Shared testing programme with agent scenarios | Test plans and results |
| AI Act serious incidents and DORA major incidents | One incident process with both classification tests | Incident records and both reports |
| AI Act provider information and DORA third-party risk | One vendor file with provider instructions, due diligence and contract | Vendor file per agent platform |
| AI Act FRIA and literacy | Run alongside DORA work for high-risk credit agents | FRIA and training records |
| Governance | One committee overseeing agents, reporting to the management body | Minutes and board papers |
What will supervisors ask?
- Do you run separate AI Act and DORA programmes for the same agents? Why?
- Show us one agent's entry and how it maps to both regimes.
- How would one agent incident be assessed under each regime?
- Where does the AI Act information from your vendor sit, and who reviews it?
- Who owns agent risk at management body level?
What does combining them not solve?
DORA does not address fairness, fundamental rights or explanations to customers; the AI Act and consumer credit law do. The AI Act does not address operational resilience to DORA's depth. Neither replaces GDPR. And UK banks outside the EU work to a different set of expectations, described in PRA expectations for AI agents in UK banks. Whether DORA reaches a particular agent is covered in does DORA apply to AI agents in financial services, and the full map is in AI agent regulation mapped to controls.
This is an explanation of the rules, not legal advice; confirm how they apply to you with counsel.
Questions people ask
Who supervises AI Act compliance for a bank?
In substance, the Act generally leaves supervision of AI used by financial institutions with their financial supervisors, so expect familiar supervisors to ask AI Act questions alongside DORA ones.
Is a credit scoring agent's incident reported under both?
It may be. An agent failure can be a major ICT incident under DORA and a serious incident under the AI Act. Assess both tests in one process.
Should the AI governance committee and the ICT risk committee merge?
Not necessarily, but agent risk needs one owner and one view. Shared membership or a joint agenda item is often enough.