Book a demo
Topics · GOVERNANCE-AGENTS

Payments Kill Switch Under DORA | Moring

DORA applies to a payments agent because an agent that moves money or changes payment records is an ICT-supported function at an EU financial entity, whether or not anyone filed it under that heading. A kill switch, a rehearsed stop drill, and reconstructable incident evidence are the concrete translations of DORA's ICT risk management, logging, resilience testing, and incident reporting obligations.

On this page
  1. Short answer
  2. How it works
  3. FAQs
  4. Related reading
  5. Sources

Short answer

DORA applies to a payments agent because an agent that moves money or changes payment records is an ICT-supported function at an EU financial entity, whether or not anyone filed it under that heading. A kill switch, a rehearsed stop drill, and reconstructable incident evidence are the concrete translations of DORA's ICT risk management, logging, resilience testing, and incident reporting obligations.

How it works

DORA, Regulation (EU) 2022/2554, has applied to EU financial entities since January 2025, and it governs the operational resilience of ICT systems and ICT-supported functions. Its drafting predates the agent boom and never uses the vocabulary of agents, but the scope is settled: a payments agent that accesses production systems, changes records, and initiates transfers sits inside it from first principles. When a supervisor asks how the entity manages ICT risk, the answer now includes the agents, and the stop control over those agents becomes part of the regulated estate.

Sources: https://www.moring.ai/blogs/dora-compliance-ai-agents/ https://www.cockroachlabs.com/blog/dora-database-requirements-ai-agents/

The five obligations that reach a payments agent deployment are ICT risk management, logging and event reconstruction, resilience testing, third-party risk, and incident reporting. The agent-specific translation is concrete. Identity must exist per agent, action logs must be complete, the stop drill must be rehearsed, contract terms must cover model providers, and incident evidence must reconstruct without archaeology. A kill switch is not an optional feature in this mapping; it is the artifact that proves the entity can contain a misbehaving payment agent inside an incident window.

Sources: https://www.moring.ai/blogs/dora-compliance-ai-agents/

Treasury practice illustrates why a stop authority has to be governed before a payment agent executes. A payment mandate is not a prompt. An agent may identify a recurring intercompany funding requirement, review the cash position, confirm the beneficiary, and prepare the transaction, only to find the bank details changed and the amount higher than expected. A framework separating intent, authorization, and settlement holds that authorization and settlement require deterministic controls, which is exactly where a kill switch lives: outside the agent's judgment, deciding whether the payment can proceed at all.

Sources: https://www.kyriba.com/blog/govern-ai-agent-payment-authority/

Architecture follows the same principle. The stop control has to sit on the payment path, between the agent and the rail, so that every payment the agent attempts is refused under a documented verdict while the agent is contained. Research on DORA and agentic infrastructure adds that the underlying data layer matters as much as the switch: append-only audit storage, per-agent identity attribution on every write, strong read consistency, idempotency, rollback support, and enforced data residency are the six database capabilities a compliance supervisor can test with existing records rather than engineering promises.

Sources: https://www.cockroachlabs.com/blog/dora-database-requirements-ai-agents/

Resilience testing is the obligation teams most often under-prepare. A kill switch that has never been triggered in a drill is a claim, not a control. The rehearsal should cover the payments-relevant failure path: the agent is suspended, in-flight payment messages are refused, queued settlement items are held, and human reviewers take over the funding queue. Observability research on agent estates notes that most organizations monitor AI traffic poorly end to end, which means the DORA evidence chain for a payment incident depends on controls designed to record what actually happened downstream.

Sources: https://www.moring.ai/blogs/dora-compliance-ai-agents/

Incident reporting closes the loop. When a payment agent causes or nearly causes an operational disruption, the report has to show when containment started, who invoked it, what the agent attempted before the switch engaged, and what compensating actions covered the funding window. Human oversight expectations reinforce this: reviewers must see agent intent, identity, permissions, and impact before action, and hold enforceable authority to pause or stop behavior. A payments program that can produce that record from a single control plane can answer DORA questions with evidence instead of narrative.

Frequently asked questions

Does DORA actually apply to an AI agent that initiates payments?

Yes. DORA regulates ICT-supported functions at EU financial entities, and an agent that accesses production systems, changes payment records, and moves money is an ICT-supported function regardless of whether it was classified as AI. The practical obligations that reach such an agent include ICT risk management, complete action logging with event reconstruction, resilience testing, third-party risk terms for model providers, and incident reporting supported by reconstructable evidence.

What does a kill switch have to demonstrate for a DORA examination?

It has to demonstrate that containment of a misbehaving payment agent is technically real and rehearsed. That means the switch sits outside the agent's own logic, refuses payment attempts on the rail path under a documented verdict, and has been exercised in a stop drill covering in-flight transfers and queued settlement items. Supervisors can test whether the data layer supports per-agent attribution and append-only audit storage using existing records rather than accepting engineering assertions.

How do DORA and the EU AI Act relate for payments agents?

They are complementary layers. DORA governs the resilience of the ICT estate, including agents that move payments, while the EU AI Act governs the AI system itself, its risk classification, and its oversight. A governed-agent architecture with per-agent identity, runtime policy enforcement, a kill switch, and complete decision logs satisfies both frameworks from one control set, so payments teams do not need to build parallel evidence chains for operational resilience and AI governance.

What evidence should exist after a payment agent is stopped?

The incident record should carry the stop request, the acknowledgements from each enforcement point, the payment actions the agent attempted before containment, any credential changes, the remaining exposure on the rail, and the compensating actions that covered the funding window while manual review took over. This record is what allows the incident report to reconstruct the event without archaeology and to demonstrate that the kill switch performed as designed within the required window.

Sources

  1. - federalreserve.gov
  2. - ithandbook.ffiec.gov
  3. - artificialintelligenceact.eu
  4. - eur-lex.europa.eu