A debate exploded in tech circles this week that finally has data behind something most practitioners already suspected.
Box founder Aaron Levie kicked it off on May 24 with a post that went wide: CEOs, he wrote, 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." They play with a prototype. They see the happy path. They make the leap. Their employees inherit something else entirely.
TechCrunch spent the following week chasing the thread, and by May 31 the verdict from the research side had crystallized: this is real, it's widespread, and the gap is measurable.
An NBER study of nearly 6,000 CEOs and CFOs across the US, UK, Germany, and Australia found that roughly 90% of firms reported zero measurable impact on productivity or employment from AI adoption over the past three years. Zero. Not marginal gains. Zero. The UC Berkeley California Management Review published a meta-analysis reaching the same conclusion: no robust relationship between AI adoption and aggregate productivity gain.
In the first five months of 2026, the tech industry has laid off 115,430 people from 152 companies. Most cited AI as a reason. If 90% of those same organizations are seeing nothing from AI, something is very wrong with the math.
What's actually happening
The psychosis framing is useful because it describes the mechanism, not just the outcome.
CEOs interact with AI through demos, pilots, and polished prototypes. The interface is clean. The task is illustrative. The result is impressive. And it is impressive — for that task, in that context, with someone who knows what they're doing running it.
Then the tool gets rolled out. Maybe it's Copilot across the organization. Maybe it's a ChatGPT Enterprise license. Maybe it's a new AI-powered CRM. And the employees who actually do the work — the program officers, the grant writers, the intake coordinators, the communications staff — encounter the reality. It doesn't integrate cleanly with the tools they already use. The prompts that worked in the demo don't work for their specific context. Nobody redesigned their workflow to make space for the AI step. There's no one to call when it does something wrong.
Three months later the usage stats are low and the CEO is confused because they saw it work.
The "last mile" Levie is describing isn't a small thing. It's the design work, the integration work, the training work, the iteration work, the maintenance work. It's everything between "this is possible" and "this is running in production and saving 40% of someone's time."
Most organizations are buying the tool and skipping everything else. The data is showing exactly what happens when you do that.
Why this hits small orgs differently
If you're a large enterprise and 90% of your AI spend produces nothing, you've burned some budget and annoyed some employees. You have other budget. You'll try again next year.
If you're a 15-person NGO, a 30-person public-sector team, or an 8-person consulting shop, the failure mode is more expensive. You don't have budget to waste on tools nobody uses. You don't have cycles for a rollout that doesn't work. And you especially can't afford the opportunity cost: while you're implementing AI badly, your competitors and peers are starting to implement it well.
There's also a pressure problem. Funders are asking about AI strategy. Board members attended a conference and heard something. Leadership is reading the same CEO-facing media that created the psychosis in the first place. The small org gets pushed into the same traps — adopt something, show activity, report on it — without the slack to do it right.
The backlash forming right now is partly this dynamic playing out in public. College graduates booing AI mentions at commencement ceremonies. DuckDuckGo reporting surging installs as users push back on Google embedding AI everywhere they didn't ask for it. These aren't Luddites. These are people who've had the expensive, underperforming version of AI forced on them by someone who wasn't doing the work.
What the other 10% are doing
The teams that are getting real results from AI in 2026 aren't doing something exotic. They're just doing the unglamorous part that the 90% are skipping.
They start specific. Not "we adopted AI" — they took one concrete, recurring task and rebuilt it. The monthly funder report that takes a communications director six hours because she has to pull from five spreadsheets and three email threads. The client intake process that produces a summary nobody reads because it takes too long to generate. The board package that requires a three-hour day to assemble. One thing. Built properly.
Someone does the actual implementation. Not a vendor who came in for a demo. Someone who learned the tools well enough to design a real workflow, write real prompts that work in context, connect the integrations, train the people who'd actually use it daily, and fix it when it breaks.
They measure it. Hours saved per week. Tasks completed. Time to first draft. They know what changed and they can explain it. They iterate when it doesn't work the first time — and it never works perfectly the first time.
They don't try to replace judgment. The best AI workflows we've seen in practice take the grinding, repetitive, low-value parts of a task and automate or accelerate them, so the human doing the work spends their time on the part that actually requires expertise. The output isn't the final product. It's the starting point for someone who knows what they're doing.
That's the whole model. It's not complicated. What it requires is time, expertise in the tools, and someone who understands the actual work being done — not just what AI can theoretically do.
The thing worth taking from this week
The "AI psychosis" debate has a cynical read and a useful one.
The cynical read: AI is overhyped, the productivity gains are fake, and organizations should wait it out.
The useful read: adoption without implementation is fake. The organizations in the 90% didn't fail because AI doesn't work. They failed because nobody did the last mile. And that's a solvable problem — just not the one that gets solved by signing another license.
The organizations that come out of this period ahead won't be the ones who adopted AI fastest. They'll be the ones who implemented it well in the places that actually mattered.
We spend most of our time in the last mile — figuring out where real gains live in a specific team's workflow and building the thing that captures them. If your org is in the 90% and you want to not be, that conversation usually starts in under an hour. Let's talk.