Insurance AI Control Plane DORA Compliance | Moring
DORA applies to insurers since January 2025 and reaches AI agents by function, not by name. An agent that updates a claim record or touches a critical workflow is an ICT-supported function, so it needs documented ICT risk management, complete logging, tested resilience, third-party oversight, and incident reporting that a control plane is positioned to provide.
Short answer
DORA applies to insurers since January 2025 and reaches AI agents by function, not by name. An agent that updates a claim record or touches a critical workflow is an ICT-supported function, so it needs documented ICT risk management, complete logging, tested resilience, third-party oversight, and incident reporting that a control plane is positioned to provide.
How it works
DORA, the EU Digital Operational Resilience Act, has applied to EU financial entities including insurers since 17 January 2025. Its text never mentions AI. The obligations are function-based: any system, including an AI agent, that supports a critical or important insurance function falls inside the perimeter because of what it does. An agent that changes a claim record, moves a payment, or supports underwriting decisioning is an ICT-supported function, and its governance must survive the same scrutiny as any other part of the IT estate.
The ICT risk management framework under Article 8 requires identification of functions and information assets, classification by criticality, and controls proportionate to harm. For agent workflows that translates into documented risk assessments, defined boundaries on what an agent can reach, and enforcement of those boundaries at runtime. Article 9 adds detection of anomalous activities and data integrity checking, which in practice means audit logs, access monitoring, and controls that can catch unauthorized modification by an agent.
Logging is where most agent deployments fall short of DORA expectations. The regulation requires detailed records of events affecting ICT-supported functions. For an agent that means a complete action record: what the agent proposed, which policy version allowed it, what a human approved, and what changed in the source system. Logs that capture only model outputs leave an evidence gap, an attribution gap when actions cannot be tied to a specific agent session, and a state gap when the record cannot be reconciled with what the systems actually show.
Operational resilience testing gives the requirement a concrete agent translation. The drill is direct: press the stop control on one production agent and start a timer. Passing means zero actions after the stop decision, a complete list of what the agent wrote within ten minutes, a rollback path or named human owner for every write, and restart blocked until a named approver signs off. A deployment rollback is not a kill switch, and an agent that fails this drill is not resilient under the framework.
Third-party risk reaches agent dependencies too. Foundation model providers, orchestration platforms, and model-hosting infrastructure count as ICT third-party providers, so they belong in the register of information with criticality classification, audit rights, exit strategies, and subcontracting transparency. Incident reporting completes the picture: an agent that takes a wrong consequential action is an incident like any other, and the carrier must reconstruct exactly what happened, when, and who was accountable.
A control plane is the practical instrument for meeting these expectations because it concentrates the evidence DORA asks for in one place. Identity-bound records of every tool call, policy decisions retained with their versions, stop controls that have actually been tested, and vendor registers covering the model supply chain turn the regulation's function-based obligations into artifacts an examiner can review without an engineering project.
Frequently asked questions
Do AI agents used by insurers fall within DORA scope?
Yes, by function rather than by name. DORA regulates ICT systems and digital operational resilience, and it never mentions artificial intelligence explicitly. The moment an agent supports a critical or important insurance function, such as updating claim records, supporting underwriting decisioning, or touching payment flows, it is an ICT-supported function and inherits the framework's risk management, logging, testing, and third-party obligations. The agent's label does not matter; what it can do to production systems does.
What does DORA logging mean for an insurance agent workflow?
DORA requires detailed logging of events affecting ICT-supported functions, and for an insurer's agents that means a record complete enough to reconstruct a run: what the agent proposed, which policy allowed the action, what a human approved, and what changed in the source system. Three gaps show up repeatedly in agent deployments: evidence gaps where logs capture outputs but not the inputs behind them, attribution gaps where actions cannot be tied to a specific agent identity and session, and state gaps where records cannot be reconciled with what the systems show. Closing all three is what makes the log defensible.
How should an insurer test agent resilience under DORA?
An insurer should run the stop drill on a production agent: activate the stop control and start a timer. A passing result means no actions after the stop decision, a complete write list produced within ten minutes, a rollback path or named human owner for every write, and restart blocked until a named approver signs off. If the only answer to whether you can stop the agent is a deployment rollback, the workflow is not resilient. The drill should be repeated as agents, tools, and policies change, because resilience that was tested once is not evidence that it holds today.
How does a control plane help an insurer meet DORA expectations?
A control plane concentrates the evidence the regulation asks for. Identity-bound records tie every tool call to a specific agent session, policy decisions are retained with their versions so reviewers can see which rule applied at action time, stop controls exist as tested mechanisms rather than runbook text, and third-party dependencies such as model providers and orchestration platforms can be tracked in a register with audit rights and exit provisions. Instead of assembling these artifacts per agent during an examination, the carrier produces them from the layer that already governed the action.