Polysea
All posts
ai agentsregulationcommercial insurance

AI Agents in Insurance Workflows: What Regulators Should Know

Theodore Johnson-Kanu··8 min read

AI agents are already in insurance workflows. The regulatory framework has not caught up. Here is a practical map of where agents are operating, what the accountability gaps are, and what oversight should look like.

The state of play

Insurance is a document-heavy, workflow-intensive industry. It is also an industry where AI adoption has moved from pilot to production faster than most observers realize.

AI agents (software systems that can take actions on behalf of a human principal, not just analyze data) are already operating in:

  • Submission intake. Extracting structured data from email submissions, ACORD forms, loss runs, and supporting documents. Routing submissions to the appropriate underwriting team.

  • Underwriting triage. Pre-screening submissions against carrier appetite, flagging risks that require human review, and generating preliminary rating indications.

  • Quote generation. Producing indicative and firm quotes for standardized commercial lines, particularly in small commercial.

  • Endorsement processing. Reading broker endorsement requests, generating endorsement documents, and routing for binding.

  • Claims intake. First notice of loss processing, document extraction, and initial claim categorization.

  • Compliance monitoring. Scanning producer activity against licensing and appointment data, flagging exceptions.

Most of these applications are running at scale today. A few are still in pilot phases. All of them involve a software system taking actions that would historically have required a human with specific credentials and authority.

This is the gap that should concern regulators, and it is not a future problem. It is a current one.

The accountability question

The central regulatory question about AI agents in insurance is not whether they are accurate, fair, or explainable (all legitimate questions, addressed by other frameworks). It is whether the chain of accountability for an agent's actions is traceable.

When a licensed producer binds a policy, there is a clear accountability chain. The producer is licensed by a state. The producer is appointed by the carrier. The producer's actions are logged in carrier systems and, typically, in the producer's agency management system. If something goes wrong, the parties and the relevant authority can be identified.

When an AI agent takes an action in that same workflow, the accountability chain is less clear. The agent is operating under some delegation from a human (a producer, an MGA, a carrier employee), but the scope, conditions, and record of that delegation are rarely documented in a way that would survive regulatory scrutiny. The agent's actions are logged in whatever system is running the agent, which may or may not be connected to the systems that track the human principal's activity.

A regulator asking "who authorized this binding" in a post-incident review should be able to get a clear answer. Today, for AI-initiated transactions, the answer is often distributed across multiple logs, multiple vendors, and multiple document types that were never designed to be reconciled.

Addressing this accountability gap is exactly why Polysea is building neutral authorization chain infrastructure. Rather than leaving each vendor to document agent authority in their own way, we are creating the shared verification layer that lets regulators, carriers, and counterparties trace any agent action back to its authorized human principal through a single, auditable record.

The NAIC Model Bulletin and state-level developments

The National Association of Insurance Commissioners (NAIC) has moved to establish baseline expectations for how carriers govern AI use. The NAIC Model Bulletin on the Use of Artificial Intelligence Systems by Insurers, adopted in late 2023, has been adopted, in various forms, by a growing number of states.

The Model Bulletin's core expectations are governance-focused: insurers should have written AI programs, oversight mechanisms, testing and validation frameworks, and documented risk management. These are reasonable expectations at the organizational level.

What the Model Bulletin does not yet address in granular terms is the transaction-level accountability question. When an AI agent initiates a specific action (a submission, a quote, a bind, an endorsement), what is the record of that action, and what authority does it attach to?

Several states have moved beyond the Model Bulletin with more specific requirements. California, Colorado, and New York have issued their own guidance or regulations that touch on AI use in insurance. The pattern in these regulations tends to focus on bias, discrimination, and consumer disclosure, which are important but which leave the transaction-level accountability question largely unaddressed.

What regulators should know about current workflows

A few facts about how AI agents currently operate in insurance workflows that are relevant to regulatory design:

Agents typically operate under human delegation, but the delegation is implicit. A producer who uses an AI tool to process submissions has, in practice, delegated certain actions to the tool. There is rarely a formal, structured record of that delegation. The producer has signed up for a software product. The terms of service constitute a general authorization. The specific scope of what the agent is permitted to do on the producer's behalf is defined by the product's feature set rather than by a documented authorization.

Cross-party agents are increasingly common. An agent operating inside a broker's workflow may interact with a carrier's system on the broker's behalf. The carrier may not know it is interacting with an agent rather than a human. The authorization for the agent to interact with the carrier is inherited from the broker's existing appointment, but there is no explicit record of this inheritance.

Logs are fragmented. An AI agent's action log typically lives inside the software product running the agent. The corresponding log in the carrier's policy administration system records the result of the action but not the agent-specific context. Reconstructing a full transaction trail across both systems is a manual exercise.

Error recovery is unclear. When an agent takes an incorrect action (an unauthorized bind, an incorrect endorsement), the rollback process is not well-defined. In a human workflow, the producer who made the error is clearly the responsible party. For an agent-initiated error, the accountability can fall anywhere between the software vendor, the producer using the tool, the carrier who accepted the transaction, and the insured who may not have known an agent was involved.

What oversight should look like

A regulatory framework adequate to current AI agent use in insurance should address, at minimum, four requirements.

1. Explicit delegation records. When an AI agent acts on behalf of a licensed producer or carrier representative, there should be a structured, verifiable record of the delegation. The record should specify what actions the agent is authorized to take, under what conditions, and for what duration. This record should be available to regulators and counterparties, not buried in a software product's terms of service.

2. Attribution at the transaction level. Every insurance transaction should be attributable to a specific authority chain. For a human-initiated transaction, this is the producer's license and appointment. For an agent-initiated transaction, this should include the agent's delegated authority and the human principal from whom that authority derives.

3. Cross-party visibility. When an agent acts across organizational boundaries (a broker's agent interacting with a carrier's system), the counterparty should be able to verify that the agent is operating within an authorized scope. This requires infrastructure that works across party boundaries, not just internal logging.

4. Auditability of agent actions. The action log for AI-initiated transactions should be designed to be auditable by regulators, not just reviewable internally. This implies cryptographic integrity, standardized formats, and retention policies that match regulatory requirements for the underlying line of business.

These requirements do not require a new regulatory regime from scratch. They extend existing producer accountability frameworks to cover non-human actors operating under delegated authority.

What this means for the industry

For carriers, the regulatory direction means that accepting AI-initiated transactions without a verifiable delegation chain is an increasing risk. A carrier that cannot answer "who authorized this action" for an AI-initiated submission is exposed in a way that did not exist when all submissions came from human producers.

For brokers, the same applies in reverse. A broker using AI tools to process submissions should have a clear record of what the tools are authorized to do on the broker's behalf. In a dispute, "the software did it" is not a defense that existing E&O frameworks handle well.

For MGAs and program administrators, the challenge is amplified. Binding authority programs involve multiple layers of delegation. When an AI agent enters the workflow at any layer, the delegation chain needs to extend cleanly through the layer below and above. This is a structural requirement, not an optional feature.

For software vendors building AI tools for insurance, the implication is that products need to be designed with regulatory accountability in mind. Products that operate as black boxes, with no structured way to document delegation or produce auditable action logs, will become increasingly hard to deploy in regulated workflows.

Infrastructure requirements

A practical regulatory framework for AI agents in insurance depends on infrastructure that does not currently exist in a standardized form. Specifically:

  • A standard for structured delegation records that can be issued by producers, carriers, and other authorized principals to software agents.

  • A verification mechanism that allows counterparties to check an agent's delegation in real time, without relying on a specific vendor's system.

  • An audit standard for agent-initiated actions that produces records suitable for regulatory review.

These are infrastructure-level problems. They cannot be solved inside a single software product, because the whole point is that the infrastructure needs to be trusted across parties and across vendors.

Conclusion

AI agents are operating in insurance workflows today, at scale. The existing regulatory framework focuses on governance at the organizational level, which is necessary but not sufficient. The transaction-level accountability question, specifically "who authorized this action and where is the record," is largely unaddressed.

The infrastructure to answer this question is buildable. It requires structured delegation records, cross-party verification, and auditable action logs. The industry and its regulators have a window to define this infrastructure deliberately, before a high-profile failure forces a reactive framework. That window is open now.

Polysea is building neutral infrastructure for the commercial insurance ecosystem, including shared exposure data management, authorization chain tooling, and automated loss run extraction. If the problems described in this article are relevant to your work, we would like to hear from you at hello@polysea.ai.