Aaron Levie, CEO of Box, put a name to something a lot of people were already feeling. He called it "AI psychosis" — his description of executives who are "sufficiently distant from the last mile of work that still has to happen to generate most value with AI." The result, in his framing, is a kind of organizational delusion: leaders convinced AI is already transforming their business while the people doing the actual work are still figuring out what prompt to write.

I don't think Levie was being mean about it.

What the numbers actually say

The 2026 PwC Global CEO Survey found that 56% of CEOs reported no significant financial benefit from AI. Twenty-two percent said their costs actually went up. A UC Berkeley meta-analysis published in the California Management Review in October 2025 found "no robust relationship between AI adoption and aggregate productivity gain."

PwC and BCG are not exactly anti-tech institutions, so these numbers aren't coming from AI skeptics.

And yet the same BCG report found that 75% of board members believe their AI knowledge matches their peers, while 40% of CEOs think their boards lack an informed view of AI strategy. So the people at the top are confident they understand AI, and also not seeing the returns.

The last mile problem is real

AI tools are genuinely useful, often in ways that are hard to measure at the aggregate level. But "useful in specific contexts for specific tasks" is very different from "transformative at the organizational level."

The last mile — actually integrating AI into the specific workflows of specific teams doing specific work — is where most of the difficulty lives.

A CEO sees a demo of a tool that can summarize contracts in seconds and thinks "we should roll this out to legal."

Legal then spends three months figuring out which contracts it works on reliably, building review processes for when it doesn't, training staff on the new workflow, and dealing with the edge cases the demo didn't show. By the time it's actually saving time, the CEO has moved on to the next demo.

I've been in enough conversations with freelancers and small operators to know this pattern isn't unique to enterprise. We do it to ourselves too. I've definitely been guilty of getting excited about a new tool — Claude, Perplexity, whatever — and assuming the productivity gains were going to be immediate and obvious. Sometimes they are. Often the actual payoff comes weeks later, after you've figured out how to use the thing in your specific situation.

Users are voting with their feet

The DuckDuckGo data is interesting here. After Google announced deeper AI integration into search, DuckDuckGo installations jumped 30%. That's not a small signal.

Paul Graham made a comment that stuck with me — that AI-written emails feel like "being lied to." I think that's the right frame for what's happening with user backlash. People aren't rejecting AI because they're luddites. They're rejecting specific AI experiences that feel hollow or forced or like the company is using AI to do less while charging the same.

There's a real difference between AI that makes an experience better for the user and AI that makes production cheaper for the company while degrading the experience. Users are getting better at distinguishing between the two.

Gallup tracked Gen Z excitement about AI dropping from 36% to 22% in a single year, with anger rising from 22% to 31%. That's a significant shift in a demographic that was supposed to be AI's most enthusiastic adopters.

What this actually means if you're a creator or freelancer

The executive disconnect creates a specific kind of opportunity, and I want to be precise about what it is.

When leadership is chasing AI adoption metrics that don't connect to how work actually gets done, the people who understand both sides — what the tools can genuinely do and what the work actually requires — become more valuable, not less.

This isn't about being anti-AI. I use AI tools constantly. I use Claude for drafts and research synthesis. I use Perplexity when I want sourced answers faster than I can Google them. These tools are genuinely useful to me in my actual work.

But I'm also not pretending they've eliminated the hard parts of what I do. The judgment calls, the editorial decisions, the knowing when something is wrong even if it reads fine — those are still mine. And in a moment when a lot of organizations are discovering that AI didn't automate away the hard parts the way the demos suggested, building your brand voice deliberately becomes more important, not less.

The clinical side of this is worth knowing about

The "AI psychosis" label also has a literal meaning that's getting less attention than Levie's metaphor, and it's worth a paragraph.

UCSF psychiatrists reported treating 12 patients in 2025 with psychosis-like symptoms linked to extensive chatbot use. OpenAI's own internal data found roughly 560,000 weekly ChatGPT users showing signs of mental health emergencies related to psychosis or mania. A Lancet Psychiatry study from March 2026 concluded that chatbots can validate and amplify grandiose delusions because of how sycophantic they're designed to be.

The key point from Columbia's Dr. Ragy Girgis is that chatbots aren't inducing psychosis in psychologically healthy people. The concern is amplification — existing vulnerabilities getting worse through extended interaction with a system that's built to agree with you.

That's relevant design information for anyone building tools that involve deep or ongoing human-AI interaction. A chatbot that never pushes back isn't neutral. It's actively reinforcing whatever the user brings to it, including the stuff that isn't accurate.

The gap is the information

What I keep coming back to is that the gap between executive enthusiasm and implementation reality isn't just a story about hype. It's information about where the actual work is.

The 56% of CEOs not seeing ROI aren't necessarily using bad tools. They're often failing at the last mile — the integration, the training, the workflow redesign, the unglamorous work of making a new capability actually fit into how things get done. That work doesn't show up in demos. It doesn't make good conference keynotes. But it's where the value either gets captured or doesn't.

Freelancers and small operators often have an advantage here that they underestimate. We don't have to coordinate across departments. We don't have to get sign-off on a new tool. We can just try it, figure out if it actually helps our specific work, and adjust. The feedback loop is much tighter.

The executives Levie is describing are making decisions about AI at a scale and distance that makes it hard to feel what's actually working. That's not a character flaw — it's a structural problem with how large organizations process information. But it does mean the people closer to the actual work are often better positioned to understand what AI is and isn't doing.

That's most of the people reading this. Worth keeping in mind when someone at a conference tells you AI is going to change everything, and you're still in the middle of figuring out whether it reliably does the one specific thing you actually need it to do. The transparency protocol helps here — being honest about when you use AI and how you use it gives clients a better sense of what they're actually buying.