Even when a human provides the objective “Prepare a meal using a kitchen knife,” the specific action to be taken is not uniquely determined. The AI must reason about numerous factors, such as what meal to prepare, which ingredients and knife to use, and in what sequence the preparation should proceed.

AI Reasoning can generate multiple Candidate Actions from such a situation. This is where AI intelligence plays a central role. More capable AI systems can generate increasingly diverse and flexible courses of action according to the relevant Context.

However, an important boundary exists:

Candidate Action ≠ Authorized Action

The fact that an AI considers an action reasonable does not mean that the action is authorized for execution. Even if the AI can generate Candidate Actions such as “use this knife,” “operate this machine,” or “call this API,” those actions have not yet been granted authority to proceed to Execution.

This distinction becomes increasingly important with Agentic AI and Physical AI. When AI moves beyond generating information and can directly affect external systems, data, financial transactions, or physical machinery, connecting Reasoning directly to Execution without an intervening governance boundary becomes a governance risk in itself.

I2EA therefore structurally separates Decision from Execution. AI may freely generate Candidate Actions, but before any of them can proceed to Execution, they must satisfy the applicable Policy, Authority, Context, and other governance conditions. The “key” shown in the previous slide symbolizes this Authorized State.

I2E does not seek to constrain the AI’s Reasoning capabilities. Rather, it allows AI to reason broadly while independently governing the final boundary through which its reasoning can affect the real world.

AI may propose an action.AI does not automatically gain the authority to execute it.

The point at which a Candidate Action reaches “Execute???” is precisely where the Execution Boundary becomes necessary.