Book a demo
Topics · AI-CONTROL-PLANE

Insurance AI Control Plane Risks | Moring

The main risks of an ungoverned insurance AI estate are shadow deployments, over-permissioned agents, and decision chains that span claims, underwriting, and policy systems without a record. A control plane addresses them by placing identity, policy enforcement, and evidence between insurance agents and the systems they act on.

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

Short answer

The main risks of an ungoverned insurance AI estate are shadow deployments, over-permissioned agents, and decision chains that span claims, underwriting, and policy systems without a record. A control plane addresses them by placing identity, policy enforcement, and evidence between insurance agents and the systems they act on.

How it works

Insurers now run AI across underwriting, pricing, claims handling, marketing, and fraud detection, and a 2025 survey conducted by the NAIC across 16 states found that 84 percent of health insurers were already using AI and machine learning in some capacity. Adoption at that pace produces a specific failure mode: individual security and compliance controls may each work as designed while the combined chain of access, data, models, and actions remains poorly understood. Risk often sits between systems, which is exactly where an insurance agent operates when it reads a policy file, calls a claims API, and writes to a system of record in a single run.

The shadow dimension is measurable. In the 2026 IBM Cost of a Data Breach Report, 43 percent of breached organizations traced incidents to shadow AI, up from 20 percent a year earlier. For an insurer the stakes are higher than a security incident alone. An ungoverned agent that touches claims medical detail or rating variables can produce outcomes that regulators treat as unfair trade practices or unfair claims settlement practices, the two harms the NAIC model bulletin expects governance frameworks and risk management protocols to prevent.

The risk categories themselves are established. A risk and governance white paper from the Wharton school groups them as data related risks, AI and ML attacks, testing and trust, and compliance, and notes that AI can produce privacy issues or discriminatory outcomes when deployed without care. Applied to insurance, those categories map onto proxy variables that correlate with protected characteristics in rating models, prompt injection against claims assistants, and testing gaps where a triage agent was never adversarially evaluated before it touched live claim files.

What separates managed risk from unmanaged risk is whether controls observe, enforce, and leave evidence, rather than whether a policy document exists. The control set that does this work includes a continuously refreshed inventory covering internal models, vendor-embedded AI, agents, and datasets; risk classification tied to use rather than to the model; pre-deployment adversarial testing; runtime monitoring with retention long enough to reconstruct an incident months later; and enforced data boundaries. Every one of those controls is scoped by the inventory, which is why an estate nobody has recorded cannot be governed at all.

The regulatory exposure is concrete. State insurance regulators oversee insurer use of AI and may require companies to explain how these tools are used in underwriting, pricing, marketing, or claims decisions, and supervisory expectations are moving toward examination-ready documentation. The EU AI Act classifies AI systems by risk and places obligations on providers and deployers, which reaches insurers serving the European market, while DORA frames the operational resilience duty of EU insurance entities that increasingly depend on automated agents. Fragmented evidence across five disconnected systems is difficult to reconcile with any of those regimes.

The practical mitigation is architectural rather than documentary. An insurer should inventory every agent and the systems it touches, assign scoped identities, require approval for consequential writes such as payment initiation or coverage changes, and retain logs that tie each action to the agent and the principal that triggered it. Where those controls live in a control plane rather than in individual agent code, the exposure of the whole estate becomes reportable, and the risks that remain are ones the insurer has chosen and can defend.

Frequently asked questions

What are the biggest risks of AI agents in insurance operations?

The largest risks are unrecorded shadow deployments, agents holding permissions beyond their function, and decision chains across claims, underwriting, and policy systems that leave no attributable evidence. Because insurance agents act on regulated decisions, an ungoverned action can also become an unfair trade practice or unfair claims settlement practice finding, which turns a technology gap into a regulatory matter.

How does a control plane reduce risk for insurance claims and underwriting agents?

A control plane sits between the agent and the systems of record, so every tool call, data access, and write is checked against a scoped policy before it completes. A claims agent can be limited to the claim file it is servicing, and an underwriting agent can be blocked from rating variables it has no mandate to touch. Each allowed action is logged, which converts risk from an assumption into an evidence base.

What do insurance regulators expect regarding agent risk?

State insurance regulators may require insurers to explain how AI is used in underwriting, pricing, marketing, or claims decisions, and examination activity increasingly asks for documentation rather than assurances. The NAIC model bulletin frames the expectation as governance frameworks and risk management protocols that prevent unfair outcomes, while the EU AI Act adds classification and oversight duties for insurers operating in Europe. Evidence produced by a control plane maps directly onto those requests.

How should an insurer begin managing agent risk?

Begin with an inventory of agents, models, vendor-embedded AI, and the datasets behind them, then classify use cases by the decisions they influence. Apply scoped identities and least privilege, require human approval for consequential writes such as payment initiation, and retain logs long enough to reconstruct an incident. Piloting the controls on internal workflows before extending them to consumer-facing claims and underwriting decisions keeps the program credible.

Sources

  1. domain
  2. domain
  3. domain
  4. domain
  5. domain
  6. domain
  7. domain
  8. domain
  9. - content.naic.org
  10. - artificialintelligenceact.eu
  11. - eur-lex.europa.eu