What if the biggest risk in enterprise AI is not that CEOs are moving too slowly, but that they believe the demo before they understand the work?
That is the uncomfortable question behind Aaron Levie’s phrase “AI psychosis,” reported by TechCrunch. My view: Silicon Valley’s executive class is increasingly treating AI less like software and more like a belief system. That does not mean AI is fake. It means the certainty around instant productivity gains is outrunning the evidence, and workers are absorbing the cost before companies prove the savings.
Are CEOs seeing the AI demo and mistaking it for the business?
Box founder Aaron Levie put the diagnosis in the sharpest possible terms.
“CEOs are uniquely prone to AI psychosis because they’re sufficiently distant from the last mile of work that still has to happen to generate most value with AI,” Levie wrote on X.
That line works because it names a real executive failure: confusing a clean prototype with an operating model.
A CEO can “play with AI,” produce a contract, generate a prototype, or watch an agent complete a narrow task. The screen looks convincing. The output arrives fast. The leap follows: if the tool can do this once, surely it can do it across the company.
But the people closest to the work know the gap. Engineers still have to review code, find bugs, and catch calls to hallucinated libraries before software ships. Legal teams still have to train models around company-specific contract language and examine agreements for buried terms. The demo is the happy path. The business lives in the exceptions.
Levie’s point lands because it is not a simple rejection of AI. It is a warning about mistaking enthusiasm for operational understanding. The useful version of the argument is not “avoid AI.” It is: use the tools, but do not confuse a promising interaction with proof that the entire company can now run differently.
That is the distinction many executives are blurring.
Why does the AI productivity miracle keep arriving before the measurements?
Because executive rhetoric is moving faster than operational proof.
The problem is not that AI cannot improve productivity. It can. The problem is that many of the strongest claims arrive before companies have shown the math behind them: how much time was saved, how much review work was added, how many errors were introduced, and whether the output created real business value rather than more internal activity.
That is a huge managerial bet: reorganize work now on the assumption that AI will absorb enough labor later.
The evidence companies need is more practical than promotional. Leaders should be able to show whether AI adoption is improving completed work, not merely increasing the amount of generated material. They should be able to separate speed from quality, usage from value, and pilot results from company-wide transformation.
That distinction should be printed and taped to every boardroom door.
AI can make work feel faster. It can produce more drafts, more code, more summaries, more internal artifacts. But more output is not the same as more finished value. If the review burden moves downstream, if executives become the authorization bottleneck, or if teams spend the saved time checking machine-generated mistakes, the productivity story changes.
The key question is not whether AI helps. It often does. The question is whether the help is broad, durable, measurable, and worth the organizational disruption being justified in its name.
What turns executive enthusiasm into executive delusion?
Distance.
The farther a leader is from the messy final mile of work, the easier it becomes to believe work is cleaner than it is. A CEO sees a generated contract. A lawyer sees missing context. A CEO sees a code suggestion. An engineer sees test failures, dependency risks, and silent nonsense. A CEO sees an agent complete a task. A manager sees the handoffs, approvals, edge cases, and accountability gaps left behind.
This is where “AI psychosis” should be read carefully. It is not a medical diagnosis. It is a provocative shorthand for hype-driven executive overconfidence: the tendency to project agency, judgment, and reliability onto systems that still require data, infrastructure, supervision, and human review.
The starkest version of the problem appears when leaders describe agent-heavy organizations as if the hard part has already been solved. A company can roll out new AI systems, redesign workflows, and ask employees to supervise more machine output. That may be ambition. It may also be the clearest test case for Levie’s warning.
A company can create many agents. That does not prove those agents can handle the judgment, context, and accountability embedded in many human decisions. The work does not disappear just because it changes shape. It often becomes verification, exception handling, quality control, and escalation.
Who pays first when AI certainty outruns AI capability?
Workers do.
The pressure is not abstract. When companies talk about AI as a way to shrink teams, replace tasks, or let fewer employees produce much more, workers become the first people asked to live inside the theory. They are expected to adopt the tools, absorb the workflow changes, and trust that management understands the difference between automation and wishful thinking.
The problem is timing. The public case for aggressive AI-driven restructuring often moves faster than the proof that the replacement horizon has arrived. Some tasks are clearly becoming easier to automate or accelerate. Others still require judgment, context, taste, accountability, and human review that cannot be waved away by calling the system an “agent.”
That is not a case against AI adoption. It is a case against pretending the future has already been fully measured.
There is also a management risk hiding in plain sight. If employees are told to “do more with AI” while watching colleagues lose jobs in AI-branded restructurings, trust erodes. The people needed to make the tools work are the same people being told the tools may make them unnecessary. That is a poor foundation for careful adoption.
Is there a strong defense of CEO AI zeal?
Yes. Caution can become paralysis.
Executive enthusiasm can push companies to experiment, expose stale workflows, and force teams to test tools they might otherwise ignore. Levie’s own posture supports that argument. He is not saying CEOs should avoid AI. He is saying they should understand both its power and its limits before turning demos into doctrine.
That is the correct middle ground.
The danger is not ambition. The danger is faith-based management: announcing the productivity breakthrough before measuring the productivity gain; cutting teams before mapping the review burden; treating “agent” as a substitute for accountability.
The source itself gives us the split:
| AI discipline | AI psychosis |
|---|---|
| Tests where AI improves work | Assumes AI transforms work by default |
| Measures errors and review time | Counts output volume as productivity |
| Keeps humans accountable | Treats agents as workers before they perform like workers |
| Separates pilots from company-wide claims | Turns demos into workforce strategy |
The strongest pro-AI case still requires measurement. Without it, confidence is just theater with a budget.
What should boards demand before CEOs preach another AI miracle?
They should demand AI scorecards, not AI sermons.
A serious AI plan should show where the tools are saving time, where they are adding review work, where errors appear, how adoption is being measured, and whether customer or employee outcomes are improving. Pilot results should not be dressed up as enterprise-wide transformation. A successful prototype should not become a justification for sweeping layoffs unless the company can show the work has truly moved, not merely become less visible.
Boards and investors should ask blunt questions:
- Savings: Are cost reductions measured, or only projected?
- Quality: What error rates are acceptable, and who signs off?
- Adoption: Are employees actually using the tools, or being told to?
- Risk: Where are hallucinations, security issues, or compliance failures most likely?
- Accountability: Who owns the output when an agent gets it wrong?
The companies that get AI right will not be the ones with the loudest claims. They will be the ones that stay ambitious while refusing to suspend disbelief.
AI should be managed like infrastructure, not worshipped like prophecy. If a CEO cannot explain the work that remains after the demo, the board should not approve the sermon.
The Bottom Line
- The article warns that executives may overestimate AI’s readiness by confusing polished demos with real operational value.
- Workers may bear the burden when companies chase AI-driven productivity gains before proving the tools can handle edge cases.
- The piece argues for practical AI adoption that accounts for human review, domain expertise, and implementation complexity.
