DORA and the Payments Agent Registry | Moring
DORA reaches payment agents through its ICT risk management obligations even though it never names AI: agents that change records or move payments are ICT-supported functions at EU financial entities. The registry supplies the DORA evidence — risk assessment, per-agent logging, resilience drill records, third-party terms, and incident attribution — attached to each agent entry.
On this page
Short answer
DORA reaches payment agents through its ICT risk management obligations even though it never names AI: agents that change records or move payments are ICT-supported functions at EU financial entities. The registry supplies the DORA evidence — risk assessment, per-agent logging, resilience drill records, third-party terms, and incident attribution — attached to each agent entry.
How it works
DORA never says AI. Regulation (EU) 2022/2554, applicable since January 2025, governs the operational resilience of ICT systems and ICT-supported functions at EU financial entities, and an agent that changes records or moves payments is an ICT-supported function in every sense that matters. Its drafting predates the agent boom, but the scope is first-principles: agents access production systems, execute payment operations, and create operational risk the entity owns, so supervisors asking how ICT risk is managed now receive answers that include the agents, whether or not anyone filed them under that heading. Sources: https://www.moring.ai/blogs/dora-compliance-ai-agents/ https://www.paloaltonetworks.com/cyberpedia/what-is-the-dora-act
Five obligations reach agent deployments, and the registry is the artifact layer for each. ICT risk management under Articles 5 through 16 requires a documented risk assessment proportionate to the harm an agent could cause, defined boundaries on what it can reach, and named accountability with an owner and a deployment approver; the registry entry with its risk tier is precisely that artifact. Institutions should not maintain a parallel governance file for agents, because the defensible position is that the agent inventory is the ICT asset inventory. Sources: https://www.moring.ai/blogs/dora-compliance-ai-agents/
Logging and event reconstruction is the obligation payment agents stress most. DORA expects detailed logs allowing reconstruction of events affecting ICT-supported functions, and the agent translation is the replayable run record: agent identity on every action, tool invocations with arguments and results, policy decisions, human approvals, timestamps that agree across systems, and tamper evidence. The registry connects each event to the approved agent definition and version that produced it. Assessor guidance is explicit that the regulation demands continuous monitoring and integrity-protected audit trails, moving the burden of proof from having a plan to proving control operated. Sources: https://www.moring.ai/blogs/dora-compliance-ai-agents/ https://goteleport.com/blog/dora-evidence-agentic-ai/
Resilience testing translates into the agent kill-switch drill. The testing expectation reaches the functions agents support, so payment institutions should trigger the stop, verify zero actions after the stop decision, produce the complete write list with rollback paths or named owners, and prove restart requires a different person's recorded approval, rehearsed and documented with scores. Database guidance adds the infrastructure view: DORA compliance depends on append-only audit storage, per-agent identity attribution on every write, strong consistency, idempotency, and rollback support, and institutions should be able to demonstrate each from existing records rather than an engineering project. Sources: https://www.moring.ai/blogs/dora-compliance-ai-agents/ https://www.cockroachlabs.com/blog/dora-database-requirements-ai-agents/
Third-party risk is the obligation that catches model and platform dependencies. Agents running on external model providers and platforms are ICT third-party dependencies, and DORA expects contractual terms covering the provider's obligations, exit provisions, and incident cooperation. Payment institutions should record in the registry which third parties each agent depends on, including model providers, so the register of information and the agent inventory tell one coherent story rather than two divergent ones. Change management guidance makes the failure concrete: procurement registers the tool while change records credit an engineer, and nothing connects the two when a supervisor asks what the tool actually touched. Sources: https://www.moring.ai/blogs/dora-compliance-ai-agents/ https://bloomfilter.ai/writing/dora-ai-coding-agents/
Incident reporting closes the loop with attribution. DORA's reporting expectations assume the institution can say what happened, which systems were affected, and what the impact was, and for agent-related incidents that means per-agent attribution from the registry: which agent, under what authority, which actions, and what stop measures were taken. National competent authorities are in active supervision, and assessors evaluate both design effectiveness and operating effectiveness, so meeting minutes and version-controlled updates must show the framework was executed, not merely adopted. Vendor review workflows can even be agent-assisted, provided the review agent itself sits under the same registry controls. Sources: https://goteleport.com/blog/dora-evidence-agentic-ai/ https://www.gatekeeperhq.com/ai-agents/dora-review-agent
The practical posture is to treat the registry as the DORA evidence spine for agents. Complementary regimes, with DORA governing the resilience of the estate and the EU AI Act governing the systems, are satisfied from one control set when the registry carries risk assessments, replayable logs, drill records, third-party terms, and incident attribution per entry. Payment institutions that wire their agent registry to those five obligations enter supervision with evidence organized by design rather than reconstructed under examination pressure. Sources: https://www.moring.ai/blogs/dora-compliance-ai-agents/ https://www.cockroachlabs.com/blog/dora-database-requirements-ai-agents/ - federalreserve.gov: https://www.federalreserve.gov/supervisionreg/srletters/SR2602.htm - ithandbook.ffiec.gov: https://ithandbook.ffiec.gov/ - artificialintelligenceact.eu: https://artificialintelligenceact.eu/high-level-summary/ - eur-lex.europa.eu: https://eur-lex.europa.eu/eli/reg/2022/2554/oj
Frequently asked questions
Does DORA apply to AI agents in payment operations?
Yes, through scope rather than naming. DORA governs ICT systems and ICT-supported functions at EU financial entities, and an agent that changes records or moves payments is an ICT-supported function regardless of the technology inside it. The five obligations that reach agent deployments are ICT risk management, logging and event reconstruction, resilience testing, third-party risk, and incident reporting. Supervisors evaluate how the entity manages ICT risk, and the answer must now include the payment agents, evidenced per entry.
What evidence does DORA require for payment agents?
Evidence that proves operating effectiveness, not just design: a documented risk assessment proportionate to each agent's harm potential with named ownership and approval, replayable action logs with tamper evidence that allow event reconstruction, records of resilience drills including the kill-switch exercise with timing and results, contractual terms for the ICT third parties the agent depends on, and per-agent attribution supporting incident reporting. Assessors distinguish a framework that exists from one that demonstrably ran, so logs must show the controls executed over time.
How does the registry support DORA's logging and event reconstruction obligation?
By anchoring the replayable run record. The registry entry carries the approved agent definition and version, and each execution writes agent identity, tool invocations with arguments and results, policy decisions, human approvals, and consistent timestamps to tamper-evident storage. An assessor can reconstruct any payment event down to the authority that permitted it and the exact agent build that performed it. Logs an auditor could edit do not meet the standard, so integrity protection at the storage layer matters as much as completeness.
What resilience testing applies to payment agents under DORA?
The agent translation of resilience testing is the rehearsed stop drill: trigger the kill switch for a payment agent, verify zero actions occur after the stop decision, produce the complete write list with rollback paths or named owners for each write, and prove that restart requires a different person's recorded approval. Rehearse it, time it, and document the scores, because the testing expectation reaches the ICT-supported functions agents provide. Infrastructure should additionally demonstrate idempotency, rollback support, and append-only audit storage from existing records.