What breaks when organisations approve AI in policy but do not measure usage?

Why This Matters for Security Teams

Approving AI in policy without measuring actual usage breaks the basic security assumption that governance can be enforced. Once employees can reach public tools, embedded copilots, browser extensions, or personal accounts, policy no longer tells security teams what is really happening. That gap creates shadow AI, weakens data handling rules, and makes audit evidence mostly declarative instead of operational. NIST’s Cybersecurity Framework 2.0 treats visibility and governance as inseparable, which is exactly where policy-only programs fail.

NHIMG’s Top 10 NHI Issues shows the same pattern across machine identities: if an organisation cannot see how an identity is used, it cannot manage it. The issue is not just compliance theater. Unmeasured use means risk teams cannot identify which models touch corporate data, which accounts are approved versus personal, or which workflows need compensating controls. In practice, many security teams discover the problem only after employees have already adopted unsanctioned tools at scale, rather than through intentional governance.

How It Works in Practice

Policy becomes enforceable only when it is paired with telemetry. For AI usage, that usually means collecting logs from sanctioned platforms, browser controls, identity providers, CASB or SaaS security tooling, and network paths that reveal where prompts and uploads are going. The goal is not surveillance for its own sake. It is to answer three operational questions: who is using AI, which instance they are using, and what data or accounts are involved.

Measured usage lets teams separate approved corporate access from personal access paths. That distinction matters because “approved AI” can still be misused if employees sign in with unmanaged accounts, forward sensitive text into public chat, or rely on plugins that bypass enterprise guardrails. Current guidance suggests combining policy with identity-aware controls, usage baselines, and periodic review of tool inventory. NIST’s CSF 2.0 supports that approach by tying governance to monitoring and continuous improvement, not just written rules.

• Define sanctioned AI services and require enterprise sign-in where possible.
• Log prompt, session, and file-transfer activity at the approved entry points.
• Correlate AI usage with user identity, device posture, and data classification.
• Review exceptions for personal accounts, shadow copilots, and unmanaged browser extensions.

NHIMG’s Ultimate Guide to NHIs — Lifecycle Processes for Managing NHIs is useful here because AI access behaves like other high-value NHI pathways: it needs lifecycle visibility, not just approval. Measured usage also reduces false confidence. In the secrets domain, NHIMG’s State of Secrets in AppSec research notes that 43% of security professionals are concerned about AI systems learning and reproducing sensitive information patterns from codebases, which is exactly the kind of risk that disappears when usage is invisible. These controls tend to break down in organisations that allow multiple unsanctioned entry points because the data needed for consistent monitoring never reaches a single control plane.

Common Variations and Edge Cases

Tighter AI controls often increase friction for staff, so organisations must balance visibility against productivity and privacy concerns. That tradeoff is real, especially when teams use mixed fleets of approved copilots, external SaaS tools, and local model gateways. The best practice is evolving, and there is no universal standard for this yet, but the consistent pattern is that “approved” without telemetry becomes a weak policy exception rather than a control.

Edge cases matter. Contractor access may be sanctioned but unmanaged. Business units may use approved AI through personal browsers where corporate logging never fires. Some regulated environments may also limit content inspection, which forces teams to rely more heavily on identity, device, and application metadata rather than full prompt capture. NHIMG’s Regulatory and Audit Perspectives is a strong reference point for building evidence that can survive review.

Where programs fail most often is in hybrid environments with both cloud AI and locally embedded assistants, because the same user can move between approved and personal paths faster than policy reviews can track.

Related resources from NHI Mgmt Group

• When should organisations use AI to help manage NHIs?
• How can organisations prevent AI agents from becoming overprivileged?
• When should organisations treat an AI agent as a privileged system?
• How can organisations govern AI agents that use service accounts and tokens?