Insurance Agent Identity Model Risk | Moring
Agent identity is now part of insurance model risk management. An agent that invokes rating, fraud, or claims models under borrowed credentials cannot be scoped, tested, or decommissioned like a model inventory entry. Treating each agent as an identified, owned, least-privilege actor lets carriers fold agent behavior into model risk governance and evidence for NAIC examination.
Short answer
Agent identity is now part of insurance model risk management. An agent that invokes rating, fraud, or claims models under borrowed credentials cannot be scoped, tested, or decommissioned like a model inventory entry. Treating each agent as an identified, owned, least-privilege actor lets carriers fold agent behavior into model risk governance and evidence for NAIC examination.
How it works
Model risk management in insurance was built around static predictive models: a model is validated, documented, approved, and monitored through periodic review. AI agents break that shape because they decide which models to call, chain calls together, and act on outputs without a human intermediary. Identity is the hinge. Identity governance research describes agents that authenticate programmatically, spawn child agents, and persist after the context that justified their privileges has changed. An actor with those properties cannot sit inside a model inventory unless it carries a defined, ownable identity of its own.
The quantitative picture explains why the gap grows. Survey data from the Cloud Security Alliance places non-human identities at more than 90 to 1 against human ones in many organizations, with 78 percent of respondents acknowledging no documented policies for creating or removing AI agent identities and only 28 percent able to trace agent actions to a human sponsor. For a carrier's model risk committee, those numbers mean the population of actors touching regulated models is expanding faster than validation and review cycles built for models, not agents.
Bringing agents into model risk discipline starts with identity. Each agent should hold a distinct non-human identity, scoped credentials limited to the models its task requires, and a named owner accountable for its behavior. Framework work on agent identity organizes the same requirements into pillars of identity, authorization, policy, and accountability, with an inventory, owner mapping, and approval thresholds among the operational components. Mapped onto a carrier's model risk program, the agent inventory sits beside the model inventory, and owner mapping gives the committee an accountable person for every actor that can trigger a regulated model.
The inherited credential problem is the most common failure. Agents are frequently provisioned with the service accounts or API keys already held by a development team, inheriting broad permissions that no single underwriting or claims task requires. In that posture, model risk controls lose resolution: the carrier can test a rating model, but it cannot bound which actor invoked it with which inputs. Runtime identity governance addresses this by checking each call at the moment it occurs, against the agent's declared purpose, rather than relying on a certification record that describes the agent as it was at approval time.
Regulatory context makes this more than internal hygiene. The NAIC bulletin expects risk-based evaluation of AI systems, accountability for third-party AI, and documentation an examiner can request, and it treats decisions made or supported by AI as subject to existing insurance law including unfair trade practices and unfair discrimination rules. The EU AI Act adds traceability obligations for high-risk systems, and DORA frames model dependencies as operational resilience matters. All three read more naturally when agent identities are scoped, because then the carrier can show the boundary around each model's use.
A practical integration path exists. Add agent identities to the model risk inventory alongside models and data sources. Require that any agent calling a regulated model hold scoped credentials and a named owner. Record per-call evidence of the policy applied and the decision made. Review agent identity grants on the same cadence as model performance monitoring. Carriers that follow this sequence convert agent-driven model usage from an unbounded exposure into a governable extension of their existing model risk framework.
Frequently asked questions
How does agent identity fit into an insurer's model risk framework?
Agents that invoke rating, fraud, or claims models are actors in the model risk surface, and they need the same treatment as models themselves: inventory, ownership, scoped access, and monitoring. A distinct non-human identity per agent lets the model risk committee bound which actor can trigger which regulated model, assign an accountable owner, and fold agent behavior into validation and review cycles instead of leaving it outside the framework.
What happens to model risk when agents use borrowed credentials?
Resolution is lost. An agent provisioned with a development team's service account or API keys inherits permissions broader than any single task requires, so the carrier can validate a rating model yet cannot bound which actor invoked it with which inputs. Identity governance fixes this by requiring per-agent credentials, least-privilege scopes, and runtime checks at the moment each call occurs rather than a certification snapshot.
Which statistics show why agent identity is a model risk concern?
Survey research reports non-human identities outnumbering human ones by more than 90 to 1 in many organizations, 78 percent of respondents lacking documented policies for creating or removing AI agent identities, and only 28 percent able to trace agent actions to a human sponsor. For insurers, those figures describe an actor population touching regulated models that grows faster than any review cycle designed for static models.
How do NAIC expectations connect to agent identity in model risk?
The bulletin expects risk-based evaluation of AI systems, defined accountability through named owners, and documentation retained for regulatory review. Scoped agent identities make each of those demonstrable: the carrier can show which identity invoked a regulated model, who owns that identity, and what policy constrained the call. Without identity scoping, the written program describes controls that the carrier cannot prove were applied to agent-driven model usage.