AI Psychosis: When Leaders Overestimate What Agents Can Do

In late May 2026, Box CEO Aaron Levie gave a name to something practitioners have been observing for months: AI psychosis.

The phenomenon is simple. CEOs and senior executives tend to overestimate what AI agents can do. They see the demos. They hear the vendor pitches. They watch the happy path presentations. And they conclude that AI is ready to take over significant portions of the work.

The Happy Path Problem

Demos show what AI can do when everything goes right. The prompt is well-formed. The data is clean. The context is clear. The output is impressive.

Deployment shows what AI does when things go wrong. The prompt is ambiguous. The data is messy. The context is missing. The output requires human intervention — often more intervention than doing the work manually would have required.

The further you are from the work, the more you see the demo. The closer you are, the more you see the deployment.

Senior leaders often have less visibility into day-to-day friction than the teams executing the work.

Why This Gap Is Dangerous

AI psychosis leads to:

- Unrealistic timelines: Executives announce AI transformations that operations cannot deliver - Under-resourced implementations: Budgets for change management, workflow redesign, and governance are cut because "the AI handles it" - Accountability gaps: When AI outputs fail, it is unclear who is responsible because the governance was never designed - Talent erosion: Workers who understand the limits leave organizations where leadership does not

The gap between executive expectations and operational reality is where transformation initiatives go to die.

The Case for Human-in-the-Loop

Human-in-the-loop oversight is often framed as a technical architecture decision. It is not. It is an organizational accountability decision.

The question is not "can AI do this task?" The question is "who is accountable when AI does this task wrong?"

In any organization with regulatory, legal, or reputational exposure, the answer cannot be "the AI." The answer must be a human who reviewed, approved, or oversaw the output. Human-in-the-loop is not a limitation of AI. It is a requirement of organizational accountability.

What Leaders Should Do Instead

1. Close the distance: Spend time watching AI in actual deployment, not in demos. See the friction. See the failures. See what happens when the happy path breaks.

2. Talk to operators: The people doing the work understand AI's limits better than the people buying the tools. Their input should shape implementation strategy.

3. Design for failure: Every AI implementation should include clear protocols for what happens when the output is wrong. If you cannot answer that question, you are not ready to deploy.

4. Resource the change: AI transformation requires change management, workflow redesign, governance, and ongoing oversight. These are not optional line items. They are the actual work.

5. Maintain accountability: Human-in-the-loop is not about distrusting AI. It is about maintaining clear accountability for outcomes in systems that cannot be fully audited.

The Healthy Middle

AI psychosis is one extreme. AI skepticism that refuses to adopt is the other. Neither serves the organization.

The healthy middle is clear-eyed about what AI can do today, what it will likely do tomorrow, and what organizational infrastructure is required to capture the value safely. Leaders who find that middle — who are neither overconfident nor dismissive — will navigate this transition successfully.