Insurance AI Agent Orchestration DORA | Moring
Under DORA, orchestrated AI agent workflows in insurance fall inside ICT risk management by function: any system supporting a critical or important function is covered whether or not it is called an agent. Carriers must fold orchestration into their documented framework, resilience testing, incident response, and third-party oversight from 17 January 2025 onward.
Short answer
Under DORA, orchestrated AI agent workflows in insurance fall inside ICT risk management by function: any system supporting a critical or important function is covered whether or not it is called an agent. Carriers must fold orchestration into their documented framework, resilience testing, incident response, and third-party oversight from 17 January 2025 onward.
How it works
DORA reaches insurance orchestration by function, not by name. Regulation (EU) 2022/2554 became applicable to in-scope financial entities, including insurance and reinsurance undertakings authorized under Solvency II, from 17 January 2025. Its text does not reference AI, machine learning, or AI agents explicitly; its obligations are function-based, so any system, including an AI model or an AI agent, that supports a critical or important function falls inside the ICT risk management framework. An orchestrated claims workflow that supports settlement, or an underwriting chain that supports quoting, is such a system.
The framework itself is demanding. Chapter II, Articles 5 through 16, requires a documented, board-approved framework covering identification, protection, detection, response, and recovery for ICT systems supporting critical or important functions. For orchestration, that means the multi-agent chain is in scope as a whole: the orchestrator, each agent, the tools they invoke, and the state they share. Mapping the workflow onto those five functions is the first compliance task, and it forces carriers to inventory orchestrated workflows they may have treated as experiments.
Operational resilience is where orchestration creates specific exposure. DORA compliance depends on timely, reliable evidence for incident investigation and reporting, and database-level guidance for AI agents under DORA emphasizes strong consistency to prevent stale reads, idempotency to prevent duplicate transactions, data residency, and tested failover. An orchestrated claims workflow that can leave a case half-processed after a partial failure, or that double-queues a settlement on retry, is precisely the weakness those requirements target. Defined state management is therefore a DORA control, not merely an engineering preference.
Traceability requirements also bind the orchestration layer. The workflow trace, which agent acted, under which policy version, on which data, with which tool, and what record was left, is the evidence base an incident investigation will need, and it must be available on a regulator's timeline, not an engineering team's convenience. Carriers should confirm that orchestrated workflows can be reconstructed after an outage, that logs survive the failure of any single component, and that incident response runbooks name the steps for pausing an agent chain safely.
Third-party risk management extends to the agent supply chain. Most enterprise agents now come from vendor platforms, with 61 percent of organizations using AI-vendor tools such as OpenAI Agent Builder, the Claude Agent SDK, and Google Workspace Studio in their workflows, and DORA's provisions on ICT third-party risk apply to those providers as service arrangements. A carrier deploying vendor agents inside critical workflows needs contractual and technical assurances covering availability, security, incident notification, and exit, and must be able to substitute or withdraw them without breaking the workflow.
The practical program is to treat orchestrated agent workflows as ICT systems from the start: add them to the inventory of critical functions, extend resilience testing to include agent-chain failure scenarios, wire their traces into incident reporting, and bring agent vendors under third-party oversight. The EU AI Act will add its own risk-based obligations on its timeline, but DORA obligations are already live. Carriers that integrated orchestration into their DORA framework early report the same benefit in both regimes: one evidence trail answering what the system did and why it was allowed.
Testing closes the loop. DORA's resilience expectations are meaningful only when exercised, so agent-chain failure scenarios belong in the testing calendar: orchestrator unavailability mid-claim, agent vendor outage, degraded dependency, and state corruption are all plausible and all testable. Each exercise should produce the same artifacts as any other resilience test, findings, remediation owners, and dates, and the results should feed the incident response runbooks. Insurers that fold orchestrated workflows into existing resilience testing this way satisfy the framework's intent without inventing a parallel process for AI.
Frequently asked questions
Does DORA apply to AI agent workflows in insurance?
Yes, by function rather than by name. Regulation (EU) 2022/2554 has applied to in-scope financial entities, including insurance and reinsurance undertakings authorized under Solvency II, since 17 January 2025, and its obligations are function-based: any system supporting a critical or important function is covered, whether or not it is called an agent. An orchestrated claims or underwriting chain is such a system.
What does DORA's ICT risk framework require for orchestrated workflows?
A documented, board-approved framework under Chapter II, Articles 5 through 16, covering identification, protection, detection, response, and recovery for systems supporting critical or important functions. The multi-agent chain is in scope as a whole, including the orchestrator, each agent, the tools invoked, and shared state. Inventorying orchestrated workflows against those five functions is the first compliance task.
How do database and state requirements relate to DORA for agent chains?
Directly: DORA compliance depends on timely, reliable evidence for incident investigation and reporting, and guidance for AI agents emphasizes strong consistency to prevent stale reads, idempotency to prevent duplicate transactions, data residency, and tested failover. A claims workflow that leaves a case half-processed after failure or double-queues a settlement on retry is precisely the weakness those requirements target.
How does DORA treat the vendors behind insurance agent workflows?
Under ICT third-party risk management. With 61 percent of organizations using AI-vendor tools such as OpenAI Agent Builder, the Claude Agent SDK, and Google Workspace Studio, those providers are service arrangements in scope. Carriers need contractual and technical assurances covering availability, security, incident notification, and exit, and must be able to substitute or withdraw vendor agents without breaking critical workflows.