On May 24, Aaron Levie, founder and CEO of Box, dropped a line on X that shook the tech world and became a headline in TechCrunch, Fast Company, and Inc. a few days later.
"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."
Translated into plain English: executives are getting high on AI optimism because they’re not the ones who have to make it actually work.
And this isn’t just one person’s opinion. It’s a diagnosis that explains a big part of the chaos the tech sector is living through right now.
What’s actually happening
Levie’s thesis is simple and devastating. CEOs “play” with AI, build a prototype or generate a contract, and jump to the conclusion that agents can do the work. But these executives are not the ones who have to review code, find bugs, and identify hallucinated library calls before deploying software. Nor are they responsible for training AI models on the company’s idiosyncratic contract terms, or spending days combing through contracts looking for tricky clauses.
They see the happy path — the perfect route where everything works on the first try — and assume that’s the normal state of affairs.
It isn’t. Not even close.
The real cost of “AI psychosis”
The numbers are brutal. In just the first five months of 2026, the tech industry has seen almost as many layoffs as in all of 2025: 115,430 people laid off across 152 tech companies in 2026, versus 124,636 people across 275 companies in 2025, according to Layoffs.fyi. And most companies point to AI as the reason for these cuts.
But here’s the problem: many argue that big tech companies are engaging in “AI washing” — attributing past or future productivity gains to AI when other business decisions are actually driving the layoffs.
In other words: people are being fired on the promise that AI will fill the gap, and then nobody fills the gap.
The macro data backs this up. An NBER study of nearly 6,000 CEOs and CFOs in the U.S., U.K., Germany, and Australia found that about 90% of companies reported zero measurable impact on productivity or employment from AI over the past three years. Average AI use per employee was 1.5 hours per week. Average use per CEO was less than one hour per week.
And the most painful part: only one in five AI investments delivers measurable ROI. Only one in 50 delivers transformational value. And 95% of enterprise AI pilots fail to make it out of the lab.
Why AI is so good at fooling executives
Levie gives a key clue in his analysis: "when they play with AI, they see the happy path outputs, without considering the next 10 or 20 things that need to happen to get sustainable outcomes from agents."
There’s a deeper psychological explanation for this. LLMs are designed to sound competent. AI doesn’t argue with you about whether the output was good, whether the strategy made sense, or whether anyone actually needed what was produced. It confirms. It validates. It tells you the org chart you’ve built with language models is working.
In other words: it tells you what you want to hear. And for a CEO already inclined to believe AI is magic, that’s fuel on the fire.
The five phases of “AI psychosis” in a company
After talking with dozens of executives about this topic, we’ve identified a pattern that keeps repeating:
| Phase | What happens | What the CEO sees | What’s really happening |
|---|---|---|---|
| 1. Discovery | The CEO tries ChatGPT/Claude for a task | “This is incredible” | A demo, not a process |
| 2. Extrapolation | They assume it works the same at scale | “We can automate the whole department” | They haven’t seen the edge cases |
| 3. Announcement | The new AI strategy is communicated | Applause, headlines | The teams get nervous |
| 4. Deployment | Tools are bought or people are laid off | “We’re going to be more efficient” | Operational chaos |
| 5. Reality | Processes break | “Why isn’t this working?” | Nobody handled the last mile |
The final result is what TechCrunch predicts: if CEOs aren’t prepared, the most likely outcome of the current AI psychosis will simply be organizational chaos.
The difference between playing with AI and operating with AI
Here’s the nuance many executives miss: there’s a massive gap between a prototype that works on Tuesday afternoon and a system that runs 24/7 without supervision for months.
Between those two states are dozens of problems nobody tells you about when you see the demo:
- Real integrations: the LLM doesn’t talk to your CRM, your email, your calendar, and your database on its own. Someone has to connect all of that.
- Edge cases: what happens when the API fails? When the customer writes in three mixed languages? When there’s conflicting data?
- Maintenance: models change, APIs change, business processes change. Someone has to keep all of that alive.
- Observability: how do you know it’s failing? Who gets the alert? Who fixes it?
- Recovery: when something breaks — and it will — does the system recover on its own or just get stuck?
That’s the last mile Levie is talking about. And that’s where 95% of AI projects die.
How to implement AI without falling into psychosis: a practical checklist
If you’re a business owner, founder, or operations leader and you want to use AI without becoming another failure case study, here’s what actually works:
1. Start with the problem, not the tool
It’s not “let’s add AI.” It’s “we have a bottleneck in X — can AI solve it?” The order matters.
2. Measure the actual time lost
Before automating, quantify it. How many hours a week go into lead follow-up? And into email management? Without a baseline, there’s no ROI.
3. Talk to the people doing the work, not the people managing it
The CEO sees the clean version of the process. The person running it sees the 47 weird cases. That information is gold for designing a solution that holds up.
4. Start small, deploy fast, iterate
A well-built workflow in 7 days beats a mega-project in 6 months. Almost always. Because the big one is based on assumptions, and the small one is based on what you learn in production.
5. Design for failure, not success
The right question isn’t “how does it work when everything goes right?” It’s “what happens when something breaks?” If the answer is “nobody notices,” you have a serious problem.
6. Avoid vendor lock-in
Open-source tools like n8n give you flexibility. Proprietary black boxes tie your hands. In AI, where everything changes every six months, flexibility is survival.
7. Don’t fire before validating
Levie has another interesting view on this: as big tech has laid off workers and blamed AI, Levie has continued to defend human workers, and he believes AI’s most powerful application is to augment — not replace — employees. He discussed how he thinks AI agents could multiply the number of workers companies need by making it easier to start more complex tasks.
The antidote: pragmatism instead of hype
At Studio SmartWork, we’ve been building applied AI solutions since 2021 — before ChatGPT made the topic trendy. And we’ve seen this pattern over and over again: companies that think they have an AI problem when they really have a process problem, or the other way around.
The way we work is designed precisely to avoid “AI psychosis”:
- We don’t sell tools, we run solutions. If we just give you software and say “good luck,” we’re contributing to the problem, not solving it.
- We start by measuring wasted time. Before building anything, we calculate what the current problem costs. That way you know whether the ROI makes sense.
- We deploy in 4–8 days, not 6 months. Because long plans built on assumptions usually crash into reality.
- We maintain what we build. The last mile doesn’t end on deployment day. It starts that day.
The real results we see have nothing to do with the fantasies of “agents replacing entire teams”:
- Lead follow-up: 4 hours/day of manual work eliminated, response time cut from hours to under 60 seconds.
- Inbox: from 3 hours to 15 minutes a day, escalations reduced by 90%.
- Workflows: 0 unrecovered failures in 6 months.
- Sales proposals: from 1–2 days to 10 minutes per proposal.
It’s not magic. It’s engineering applied to specific problems, with realistic expectations.
Conclusion: AI is a tool, not a religion
The healthiest thing you can do as a leader is lower the expectation bar and raise the execution bar. AI won’t save your business on its own, but when applied well, it can eliminate the repetitive work burning out your team.
The question is not “how do I put AI into my company?” It’s “what specific, measurable, boring tasks can I eliminate so my people can focus on what only humans can do?”
That question doesn’t make headlines. But it does produce results.
And in the end, after all the psychosis and hype, results are the only thing that matters.