The AI Governance Gap Nobody Is Talking About: WHO vs. HOW

In the week of March 17–21, 2026, four companies launched AI agent governance platforms. Kore.ai released an Agent Management Platform spanning every major framework. Entro Security debuted “Shadow AI” governance for rogue agents. Two startups—Geordie AI and Token Security—made the RSAC 2026 Innovation Sandbox Top 10 for agent security. Microsoft published “Secure Agentic AI End-to-End.” The World Economic Forum released a white paper on agent governance.

The AI agent governance market is arriving faster than anyone projected. Gartner predicts enterprises will operate thousands of agents by 2028. Databricks reports that companies with governance tools get 12 times more AI projects into production. The category is real, it is large, and it is forming now.

But every solution being built solves only half the problem.

The WHO Layer: Identity, Access, and Permissions

The current wave of AI governance platforms answers three questions: Which agents exist in our organization? What systems can they access? When should they be suspended or retired?

These are important questions. They are also familiar—they mirror the identity and access management (IAM) patterns that enterprises have used for human employees for decades. Microsoft’s Agent 365, launching May 1 at $15/user/month, makes this explicit: every agent gets an Entra ID, conditional access policies, audit trails, and lifecycle management. “Treat agents like employees,” the pitch goes.

The pattern is consistent: identity, access, audit. Who is this agent? What can it touch? Did it comply?

This is necessary. It is also insufficient.

The Missing Layer: HOW Agents Decide

Consider an agent with perfect credentials. It has the right Entra ID, the right permissions, the right audit trail. It is authorized to access your customer database, your email system, and your financial APIs.

Now consider what WHO governance cannot prevent that agent from doing:

• Fabricating data to make its own metrics look better
• Spending $50,000 on ad campaigns because no economic constraint evaluated the decision
• Sending 10,000 emails that damage your brand because no reputation gate checked the content
• Optimizing a metric in ways that harm users (Goodhart’s Law at machine speed)
• Failing silently for 13 days because no liveness check detected the outage
• Making irreversible decisions without human escalation because no authority level bounded its autonomy

An authorized agent making bad decisions is not an identity problem. It is a governance problem—and specifically, a governance problem that WHO-layer solutions are architecturally unable to solve.

What HOW Governance Looks Like

HOW governance addresses the decision layer—not who the agent is, but how the agent reasons, evaluates, and constrains itself. In practice, it requires four architectural components that no current platform provides:

1. Hard Constraints (Inviolable Rules)

Rules that cannot be overridden by the agent, by other agents, or by prompt manipulation. Not “guidelines” or “policies” stored in a dashboard—constraints enforced programmatically on every execution cycle.

Examples: no fabricated data (ever), no spend above a threshold without human approval, no silent outage longer than 24 hours, no timing-unsafe secret comparisons.

2. Evaluation Gates (Decision Checkpoints)

Before an agent acts, its decision passes through evaluation gates that assess whether the action is epistemically sound (is the agent certain enough?), economically viable (does the math work?), reputationally safe (could this damage trust?), and governmentally compliant (does this follow the rules?).

WHO governance asks: Is this agent allowed to act? Gate-based governance asks: Should this action be taken at all?

3. Authority Levels (Bounded Autonomy)

Not all decisions are equal. A daily status update requires less oversight than a $10,000 budget allocation. HOW governance defines graduated authority levels that match the reversibility and impact of each action to the appropriate level of autonomy.

This is where Singapore’s new Model AI Governance Framework for Agentic AI aligns: a five-tier graduated autonomy taxonomy where oversight scales with risk. The framework describes what the architecture should look like. The question is who will build it.

4. Self-Amendment (Evolutionary Governance)

Static rules break. Markets shift, regulations change, the system learns. HOW governance includes a formal amendment process—the ability for the system to evolve its own rules while maintaining inviolable core constraints.

This is the deepest architectural gap. Every current platform treats governance as a fixed configuration. Constitutional self-governance treats it as a living system.

Why the Gap Matters Now

Three converging signals make this gap urgent:

Scale. 67% of Fortune 500 companies now have at least one AI agent in production, double from 34% in 2025. The World Economic Forum reports 82% of executives plan to deploy agents within three years. At the scale Gartner projects—thousands of agents per enterprise—per-agent admin policies become unmanageable. You need constitutional rules that apply to all agents by default.

Complexity. Multi-agent systems introduce emergent behaviors that no single-agent IAM policy can predict. When Agent A passes data to Agent B, which passes a decision to Agent C, who authorized the final action? WHO governance traces the identity chain. HOW governance evaluates whether the decision should have been made at all.

Regulation. The EU AI Act takes full effect August 2, 2026. NIST launched its AI Agent Standards Initiative with listening sessions in April. Singapore published the world’s first governmental framework for agentic AI governance. These frameworks all describe behavioral requirements—transparency, human oversight, risk management, accuracy, robustness. These are HOW requirements, not WHO requirements.

The “AI Brain Fry” Connection

BCG’s March 2026 study coined the term “AI brain fry”—the mental fatigue workers experience from overseeing AI systems. Among affected workers, 33% reported increased decision fatigue. 34% showed active intention to leave their company.

But here is the finding that matters most: workers who used AI to reduce repetitive work reported lower burnout. The variable that determined whether AI helped or harmed was not the AI itself. It was how the AI was governed—whether workers managed the AI’s decisions, or whether the AI managed its own.

This is the same WHO vs. HOW distinction at the human level. If your governance model requires humans to review every agent decision (WHO governance), you create more cognitive load, not less. If your governance model means agents constrain themselves within verified boundaries (HOW governance), humans supervise the system, not every action.

What Constitutional Self-Governance Means in Practice

Constitutional self-governance is not a metaphor. It is an architectural pattern where agents operate under binding rules—hard constraints, evaluation gates, and formal amendment processes—enforced programmatically on every execution cycle.

In a constitutional system:

• An agent cannot fabricate data, even if it would make its metrics look better, because a hard constraint prevents it on every cycle
• A spending decision passes through an economic gate that evaluates sustainability before execution, not after
• A marketing email is checked by a risk gate that prevents reputation damage before the email is sent, not flagged in an audit log after
• When a rule needs to change, a formal amendment process preserves the audit trail and prevents uncontrolled drift
• When an agent fails, a circuit breaker pattern prevents cascading failures, and a dead letter queue preserves the failed task for human review

The distinction from admin governance is not ideological. It is architectural: constraints are enforced in the execution loop, not in the monitoring dashboard.

The Category Definition Race

Right now, “AI governance” is being defined by the largest platforms. When Microsoft says “AI governance,” they mean Agent 365—identity, access, audit. When Kore.ai says “agent management,” they mean cross-framework admin controls. When Entro says “agent governance,” they mean Shadow AI discovery.

If these definitions calcify, “AI governance” will permanently mean “admin dashboard for agents.” And the HOW layer—the part that actually prevents bad decisions, that regulatory frameworks are actually requiring, that the “AI brain fry” research says matters most—will remain unbuilt.

The question for organizations deploying AI agents is not whether to adopt governance. It is whether to adopt governance that only controls who agents are, or governance that also controls how agents decide.

The badge gets them in the building. The constitution tells them how to behave once inside.