Governing AI use across employees and autonomous agents

At a glance

What this is: Lumia Security is framing AI governance as a real-time control problem for both human users and autonomous agents, with policies driven by content, context and business intent.

Why it matters: This matters because IAM and security teams now have to govern AI behaviour as an access and policy issue, not only as a data-loss or shadow IT problem.

By the numbers:

• The company emerged from stealth in December 2025 with $18 million in seed funding.

Context

AI governance is moving from policy documents to runtime control. As organisations adopt more copilots, chat interfaces and autonomous workflows, the real problem is no longer whether AI exists, but how much authority it has, what context it can see, and which actions it can take on behalf of the business.

That makes this an identity and access issue as much as a content issue. If employees and agents can both act through AI interfaces, security teams need ways to distinguish who or what is acting, what it is allowed to do, and how those permissions change across applications and workflows.

Key questions

Q: How should security teams govern employee AI use without blocking productivity?

A: Start with visibility into sanctioned and shadow AI use, then apply runtime policies that inspect intent and context rather than only keywords. The goal is to allow legitimate work while preventing sensitive data from leaving controlled boundaries. Teams usually need ownership, approved models, and enforceable logging before they can scale access safely.

Q: Why do static data classifications fall short for AI governance?

A: Because many AI risks are not about a single data type, but about the meaning of the request in a specific business context. A prompt may be harmless in one workflow and disallowed in another. Policy needs to reflect intent, not just content, if organisations want controls that match real usage.

Q: What breaks when organisations monitor AI but cannot intervene in real time?

A: They get visibility without control. If the security layer can only observe, then risky prompts, agent actions and downstream decisions can still complete before anyone can react. Effective governance needs enforcement that can allow, block, redact or modify an interaction while it is happening.

Q: What is the difference between governing human AI use and autonomous agent actions?

A: Human use is usually bounded by awareness and direct accountability, while agent actions can execute independently across systems with delegated authority. That means agent governance has to focus on permission scope, revocation, and runtime controls. If an agent can act without supervision, it should be governed like an identity with tightly constrained authority.

Technical breakdown

How protocol analysis changes AI governance

Protocol analysis looks at the content, context and intent of AI interactions across the systems where those interactions occur. In practice, that means the control point is not just the model or the chat interface, but the flow of requests moving through applications, environments and agent actions. The value of this approach is that it can apply controls consistently even as new AI applications and protocols appear, which is important because point-in-time classifications age quickly in fast-moving AI estates. See the Agentic AI Security Guide.

Practical implication: map AI policy enforcement to interaction flows, not just to approved model endpoints.

Why business-context policies matter for AI and agents

Traditional security controls are good at defining broad data categories, but they often struggle with business-specific intent. A topic engine lets organisations define the subjects, tasks or decisions that should trigger policy, such as forbidden advice, restricted business topics or acceptable use boundaries. That is especially important for autonomous agents, because the risk is not only what they read, but what they decide to do with that context. Policy has to follow meaning, not just data labels. See the AI Agent Authorisation Guide.

Practical implication: define AI policy in business terms that match the decisions users and agents are actually making.

How real-time interventions reshape access control for AI

Real-time interventions such as allowing, alerting, redacting, modifying, triggering or blocking move AI governance from after-the-fact review to active control. That is a meaningful shift for identity programmes because it treats AI actions as enforceable events, not only as audit records. When those controls extend across devices, browsers, native applications and coding environments, they start to resemble a policy layer for machine-mediated access rather than a narrow content filter. See the AI Agent Observability, Audit and Incident Response Guide.

Practical implication: design approval, redaction and blocking decisions as runtime controls that can be applied before an AI action completes.

NHI Mgmt Group analysis

AI governance is becoming an access-control problem, not a content-moderation problem. The article’s central shift is from classifying prompts after the fact to controlling what AI is allowed to do in the moment. That is closer to identity governance than traditional DLP because the question is not only what data was seen, but whether the acting entity, human or agent, had legitimate authority to use it. Security teams should treat AI policy as a runtime authorization layer, not as a reporting layer.

Business context will become the deciding control plane for AI risk. Static labels cannot express whether a request is acceptable in a legal, financial, engineering or customer-facing context, which is why plain-language policy definition matters. The industry needs policy engines that understand business meaning, not just content type, because AI risk is increasingly about permitted intent rather than prohibited keywords. Practitioners should expect governance to move from generic controls to context-specific decision rules.

Agent actions need to be governed as first-class identities in enterprise control models. The article explicitly spans employees and autonomous agents, which is the right framing for the next phase of AI governance. Once agents can act across multiple environments, the control problem becomes one of authority, delegation and revocation, not merely monitoring. AI agent identity security will matter wherever AI can initiate actions without direct human supervision, and practitioners should align agent policy with identity governance, not ad hoc exception handling.

Real-time intervention is the operational difference between AI visibility and AI control. Logging alone will not contain AI misuse if the enterprise cannot allow, block or modify the action before it completes. That distinction matters because many AI governance programmes over-invest in observability and under-invest in enforcement. Security leaders should judge any AI control layer by whether it can interrupt risky behaviour in motion, not only by whether it can explain it afterward.

From our research library:

• According to Gartner, worldwide spending on generative AI is set to reach $644 billion in 2025, a nearly 77% year-over-year increase.
• Read next: AI Agent Identity Security Buyer's Guide

What this signals

Policy has to move closer to the action. As AI use expands across employees and autonomous agents, runtime enforcement becomes more important than retrospective review. Organisations that keep governance at the level of awareness and logging will struggle to shape what AI actually does in business workflows.

AI adoption is outpacing confidence. Only 13% of organisations feel extremely prepared for the reality of agentic AI despite the majority racing toward autonomous adoption, according to the 2026 Infrastructure Identity Survey. That gap suggests the next control challenge is not experimentation, but governable scale.

Identity teams will be pulled into AI policy ownership. As AI systems start to mediate decisions, approvals and execution across applications, the distinction between AI governance and access governance will keep narrowing. Practitioners should expect AI controls to be evaluated alongside privilege, delegation and lifecycle management, not as a separate niche.

For practitioners

• Define AI policy in business terms Write rules around the tasks, topics and outcomes that matter to the organisation, not just around broad data classes or generic acceptable-use language.
• Map where AI actions are actually enforced Inventory the applications, browsers, coding environments and agent workflows where AI interactions occur, then identify where policy can be applied before a decision completes.
• Separate human and agent authority Document which actions may be taken by people, which may be delegated to agents, and which require explicit approval or intervention before execution.
• Test runtime interventions, not just monitoring Validate that allow, alert, redact, modify and block actions work consistently across the environments where AI is used, including autonomous workflows.

Key takeaways

• The article frames AI governance as a runtime control problem that spans both employee usage and autonomous agent activity.
• Business context, not just content classification, is the control layer that will determine whether AI policies are actually enforceable.
• Identity teams need to treat agent authority, intervention points and delegated actions as part of the access model.

Related NHI Mgmt Group resources: Agentic AI Security Guide · AI Agent Authorisation Guide · AI Agent Observability, Audit and Incident Response Guide · AI Agent Identity Security Buyer's Guide

Key terms

• Agentic AI: Autonomous AI systems capable of planning, deciding, and taking actions, including calling APIs, writing code, and orchestrating other agents, with minimal human oversight. Agentic AI introduces new NHI risks as agents must authenticate to external services.
• Runtime Policy Enforcement: Runtime policy enforcement evaluates a request at the moment it is executed instead of relying only on preconfigured permissions. For AI agents, this allows decisions to reflect current context, target sensitivity, and behavioural signals rather than static assumptions.
• Business Context Rules: Business Context Rules are organisation-specific rules that tell an AI security system what normal activity looks like in that environment. They help reduce false positives by adding context about approved behaviour, priorities, and exceptions, so triage and detection are aligned to actual business operations.
• Agent Authority: The permission an AI agent receives to act on behalf of a verified person. In this model, authority is inherited rather than original, so governance must trace the agent back to the human intent, device context, and current trust state that authorised it.

Deepen your knowledge

The NHI Foundation Level course, the industry's only accredited NHI security programme, covers NHI governance, agentic AI identity and workload identity. It helps security and identity practitioners translate emerging AI risk into practical control design.