Why This Gap Is Hard to Quantify Publicly
Vendor surveys, analyst reports, and conference talks frequently cite figures on how many organizations have deployed AI agents versus how many have implemented runtime controls for them. In practice, most of these figures fail basic evidentiary tests: undisclosed sample sizes, vendor-funded methodology, or definitions of "governance" that vary so widely between sources that comparison is meaningless.
This does not mean the underlying concern is unfounded. It means the concern is currently better addressed through internal measurement than through citation of external benchmarks that cannot be verified or reproduced.
Deployment Metrics Are Not Governance Metrics
"Percentage of organizations using AI agents" and "percentage of agents operating under enforced runtime policy" are two different measurements answering two different questions. The first describes adoption. The second describes control. Conflating them produces a false sense of security: an organization can have widespread agent deployment and near-zero runtime governance simultaneously, and a single blended statistic will not reveal that.
Governance teams should track these as separate, named metrics internally, even before any external benchmark exists to compare against.
Runtime Governance Is Architecturally Distinct From Model Evaluation
Pre-deployment evaluation and runtime governance address different points in an agent's lifecycle. Evaluation asks whether a model behaves acceptably before it reaches production. Runtime governance asks what a deployed agent is actually permitted to do while it is operating: which tools it can invoke, what data it can access, and whether its actions are logged and auditable.
Treating evaluation results as a substitute for runtime controls leaves a structural gap. An agent that passed evaluation can still be granted excessive standing permissions in production, with no mechanism to detect or constrain misuse at runtime.
Adapting Least Privilege to Non-Human Identities
Least privilege is a familiar principle in identity and access management, but most existing IAM systems were designed around relatively static human or service account access. AI agents require dynamic, task-contextual permissions that can expand or contract based on what a given task requires, not a fixed role assigned at provisioning time.
Extending or complementing existing IAM to represent non-human identities with this kind of contextual scope is a prerequisite for meaningful runtime governance, not an optional enhancement.
Building a First-Party Measurement Baseline
In the absence of a trustworthy external benchmark, the most defensible number a governance team can produce is its own. That means tracking, at minimum, two figures: the total number of deployed agents, and the subset of those agents operating under active runtime policy enforcement with scoped permissions and audit logging.
This baseline does not require waiting for industry consensus on definitions. It requires an internal inventory process and a consistent definition of what counts as "governed" inside your own environment.