Why Sector-Wide Statistics Are Hard to Find
Insurance brokerages have moved quickly to deploy AI agents across underwriting support, claims triage, and client communications. Leadership teams and boards are asking for benchmarks: how many peer firms have deployed AI agents, how many have documented governance gaps, and how exposed the sector is to oversight failures. As of 2026, no verified, methodologically sound survey data specific to insurance brokerage AI governance has been published with the rigor needed to support enterprise risk decisions. Broader insurance regulatory frameworks, including NAIC model guidance, address AI use in principle, but they do not yet provide the quantitative adoption or incident benchmarks that risk committees are asking for. Any statistic circulating without a named source, methodology, and sample size should be treated as unverified, regardless of how frequently it is cited in industry commentary.
The Risk of Using Unverified Numbers
It is tempting to cite a round number, such as a percentage of brokerages using AI agents in claims workflows, because it makes an internal business case easier to build. This is a mistake. If a risk leader presents a figure to their board or a regulator and cannot trace it to a primary source, the credibility of the entire governance proposal is at risk. Regulators reviewing AI governance programs are increasingly likely to ask where adoption or risk figures originated. A governance program built on a defensible internal measurement framework, even a modest one, will hold up better under scrutiny than one built on borrowed statistics from an unrelated industry survey or an unattributed blog post.
What Can Be Measured Internally, Starting Now
In the absence of sector benchmarks, risk leaders can still build a credible evidence base using their own environment. Three categories of internal metrics are directly actionable and do not depend on external data. First, audit-trail coverage: the percentage of AI agent tool calls, across underwriting, claims, and client-facing systems, that produce a reviewable log entry. Second, permission scope: whether agents operate with standing, broad access to systems and data, or whether access is scoped to the specific task at hand and expires afterward. Third, approval friction: which categories of agent action, such as issuing a quote, modifying a claim record, or sending client communications, require a human approval step before execution. These three metrics form a practical starting governance scorecard that a risk committee can review quarterly, independent of any published industry benchmark.
Regulatory Context Without the Guesswork
State insurance regulators, working through NAIC model frameworks, have signaled active interest in how insurers and brokers govern AI systems. What has not yet solidified is a uniform, quantified compliance standard specific to brokerage AI agent deployments. This means risk leaders currently operate in a period where governance expectations are directionally clear but numerically undefined. The practical response is not to wait for regulators to publish specific thresholds, but to document current practice against the general principles already visible in model bulletins: explainability of automated decisions, human oversight of consequential actions, and record-keeping sufficient to reconstruct how an AI-assisted decision was made. Firms that can demonstrate this documentation will be better positioned regardless of how specific quantitative rules eventually evolve.
Where Runtime Governance Fits
Regardless of what sector-wide statistics eventually emerge, the underlying operational problem is consistent across insurance brokerages deploying AI agents: agents are being granted tool access to underwriting systems, claims platforms, and client communication channels, often without consistent enforcement of least privilege or complete audit logging of what the agent actually did. Runtime governance addresses this at the point of execution, enforcing permission boundaries, requiring approval for sensitive tool calls, and generating audit records as agents act, rather than relying on after-the-fact review. This does not require waiting for an industry benchmark to justify investment. It requires an honest internal assessment of current agent permission scope and audit coverage, which most brokerages can produce today without external data.