DORA does not mention AI agents, but every agent a financial entity runs is an ICT asset and usually depends on ICT third parties. So DORA's four core duties apply: bring agents into the ICT risk management framework, classify and report agent incidents, include agents in resilience testing, and manage agent platforms and model providers as ICT third-party arrangements.
What does DORA require, in plain words?
The Digital Operational Resilience Act (Regulation (EU) 2022/2554) applies from 17 January 2025. It aims to make sure financial entities can withstand, respond to and recover from ICT disruption. Its main duties, in substance:
- ICT risk management. A documented framework, owned by the management body, to identify, protect, detect, respond to and recover from ICT risk.
- ICT incident management and reporting. A process to detect, classify and handle ICT incidents, with major incidents reported to the competent authority.
- Digital operational resilience testing. A testing programme proportionate to the entity, with advanced threat-led testing for those identified by authorities.
- ICT third-party risk. Managing reliance on ICT third-party service providers, including contract terms, a register of arrangements and exit plans.
DORA also encourages sharing of threat information among financial entities. For background, see what DORA is.
Who does it apply to?
A wide range of EU financial entities: credit institutions, payment and e-money institutions, investment firms, crypto-asset service providers, trading venues, insurers and reinsurers, and others. Smaller entities benefit from proportionality and some from a simplified framework. Critical ICT third-party providers designated by the European supervisory authorities face direct oversight.
The short answer on scope for agents is in does DORA apply to AI agents in financial services. Outside the financial sector, essential and important entities face similar cybersecurity duties under NIS2, set out in NIS2 and AI agents.
What does each duty mean for an agent that acts?
Risk management. An agent is an ICT asset that also makes decisions about which other assets to use. It must be inventoried, its dependencies mapped, its access tightly limited and its behaviour monitored. The detail is in DORA ICT risk management framework for AI agents.
Incidents. Agent failures include availability problems, such as the model provider going down, and integrity problems, such as an agent applying the wrong rule to a batch of transactions. Both can be ICT incidents. Classification uses DORA's criteria: clients affected, duration, geographical spread, data losses, criticality of services and economic impact.
Testing. Resilience testing must include scenarios particular to agents: model provider outage, manipulated inputs that push the agent outside its scope, and mass loss of authority if certification fails.
Third parties. Agent platforms and model providers are ICT third-party service providers. If they support a critical or important function, DORA's stronger contract and oversight requirements apply.
The four duties also connect. A model provider outage is a third-party event, an incident, a testing scenario and a risk in the framework all at once. Entities that run four separate workstreams for agents usually end up with four inconsistent inventories. One register of agents, with each capability, its depth, its owner, its sources and its providers, should feed all four duties, so that the answer to a supervisor's question about any agent is the same whichever team is asked.
How do you meet DORA for agents?
| Requirement | Control | Evidence it leaves |
|---|---|---|
| ICT risk framework covers agents | Agents and their capabilities inventoried with owner, dependencies and access | Asset register entries per agent |
| Protection and prevention | Least-privilege agent identities; every action checked against its bounds | Access reviews and filed verdicts |
| Detection | Monitoring of verdict rates, refusals, errors and provider availability | Alerts and monitoring reports |
| Response and recovery | Ability to drop an agent to read-only or stop it; fallback to staff | Runbooks and recovery test results |
| Incident classification and reporting | Agent incidents assessed against DORA criteria | Classification records and reports filed |
| Resilience testing | Agent scenarios in the testing programme | Test plans and results |
| Third-party risk | Agent platforms and model providers assessed, contracted and in the register | Due diligence, contracts, register entries |
| Exit | Records, briefs and configurations exportable | Tested exit plan |
| Management body oversight | Reports on agent risk and incidents to the management body | Board papers and minutes |
What will a supervisor ask?
- Where do agents appear in your ICT asset inventory and risk assessment?
- Which agents support critical or important functions?
- How would you classify an incident where an agent processed transactions wrongly for an hour?
- Have you tested a model provider outage? What happened?
- Are your agent platform and model providers in the register of information?
- What can your management body tell us about the risks agents bring?
What does DORA not cover?
It does not regulate whether an agent's decisions are fair, accurate or lawful; the EU AI Act, GDPR and conduct rules do. It does not prescribe agent design or autonomy levels. It does not apply to entities outside its scope. And it does not stand alone: national supervisors apply it alongside banking, insurance and payments rules. For the supervisors involved, see who regulates AI agents in financial services, and for the full map, AI agent regulation mapped to controls.
This is an explanation of the rules, not legal advice; confirm how they apply to you with counsel.
Where Pryme fits
Pryme keeps one record for each agent capability: its depth, its certificate, its sources and the bound that decided every action. That single record can feed the ICT asset inventory, incident classification and reconstruction, resilience test evidence and the third-party register, rather than each team building its own.
Questions people ask
Is an AI agent an ICT asset under DORA?
In substance, yes. It is software running on ICT systems that processes information and supports business functions, so it belongs in the ICT asset inventory and risk framework.
Does DORA require human oversight of agents?
Not in those words. But its requirements on access control, change management and response make it hard to justify letting an agent take high-impact actions without a person. The EU AI Act addresses human oversight directly for high-risk systems.
Do small payment institutions need to apply all of this?
DORA applies proportionately, and some smaller entities use a simplified framework. The principles still apply: know your agents, limit their access, handle their incidents and manage their providers.