When betting everything on AI goes wrong: the real cost of 'AI psychosis'

ClickUp has cut 22% of its workforce in 2026, claiming that AI agents can handle those functions. It's not an isolated case: according to TechCrunch, tech layoffs so far in 2026 are already approaching the total recorded throughout 2025. The term that has named this phenomenon was coined by Aaron Levie, founder of Box: AI psychosis.

Levie describes it with uncomfortable precision: the people who decide that AI can replace a job are often those who understand least what that job actually entails. It's an observation worth pausing on.

What 'AI psychosis' means in practice

It's not a clinical diagnosis or an academic concept. It's a functional description of a decision-making pattern: executives and boards dazzled by the capabilities of current models, extrapolating without rigour which tasks can be fully automated and on what timeline.

The problem isn't that AI can't do useful things—it can, and with growing scope. The problem is the gap between what a model can demonstrate in a demo and what a team needs to operate reliably in production, with organisational context, with exceptions, with human relationships and with legal accountability. That gap doesn't disappear by CEO decree.

In ClickUp's case, the bet is particularly striking because the company sells precisely productivity tools for human teams. Replacing a fifth of the workforce with AI agents while maintaining a product designed for human collaboration is a contradiction that customers will soon notice.

Why it matters beyond the headline

The trend has consequences that go beyond each individual company. When several tech companies make simultaneous cuts justified by the same narrative—AI agents cover those functions—it creates a mutual validation effect that pressures the rest of the sector to follow the same path, regardless of whether the decision is technically sound.

This is especially relevant in roles involving contextual judgment, stakeholder management, or unstructured work. Current agents, including the most capable ones, work well in bounded and well-defined workflows. When the environment is ambiguous or changing, human oversight remains necessary, and dispensing with it comes at a cost that doesn't always show up in the next quarter but in the one after that.

There's also a metrics problem. It's relatively easy to measure the cost of a salary; it's far harder to measure the cost of a decision made poorly by an unsupervised agent, of a customer who doesn't renew because support got worse, or of a project that derails because no one with judgment was monitoring it.

Who this discussion affects

This conversation directly affects those of us working on AI integrations, whether with Claude Code, MCP servers, or bespoke agents. Part of our work consists precisely in helping organisations identify which processes can be automated with guarantees and which require more careful design, with humans in the loop.

The commercial pressure to justify AI investments with staff cuts is understandable, but it's a trap. The automation projects that work best aren't those that replace entire teams at once, but those that free those teams up for higher-value work. The difference between both approaches isn't technological: it's a matter of judgment.

What Levie calls AI psychosis has a cure, and it doesn't come from slowing AI adoption but from taking seriously the question that models can't answer alone: what exactly does this person do, and what happens when they're not there?

---

From EP, we've been watching for months as some organisations automate thoughtfully and others do it in haste. The results so far follow the pattern you'd expect.