In boardrooms and product meetings, “AI” has become a kind of shorthand for certainty. It’s the promise that work can be streamlined, costs can be contained, and teams can move faster—often with fewer people. But there’s a growing pattern behind the optimism: companies sometimes treat AI job replacement as if it were a straightforward substitution problem, when in reality most jobs are bundles of judgment, context, relationships, and messy exceptions. When the people making the call don’t fully understand what the work actually looks like day-to-day, the result can be what Box founder Aaron Levie has described as “AI psychosis”—a belief that the technology will behave like a perfect replacement, even when it’s still learning how to operate in the real world.
The phrase is provocative, but the underlying issue is practical. AI systems can be astonishing at certain tasks: summarizing documents, drafting first-pass emails, extracting fields from forms, generating code suggestions, or routing tickets. Yet those capabilities don’t automatically translate into reliable autonomy across an entire role. And when leadership assumes they do, the organization can end up cutting where it shouldn’t, automating where it can’t, and redefining responsibilities in ways that create new bottlenecks rather than removing them.
Recent reporting around AI-driven layoffs and workforce reductions illustrates how quickly this shift can happen. ClickUp, for example, has reportedly cut 22% of its workforce tied to AI agents. That number matters not just because of the human impact, but because it signals a specific strategic bet: that AI agents aren’t merely augmenting workflows—they’re replacing a meaningful portion of labor. At the same time, broader tech layoffs in 2026 are reportedly approaching the pace seen across all of 2025, suggesting that the “AI wave” is colliding with the usual pressures of cost control, competition, and investor expectations. In other words, AI isn’t operating in a vacuum; it’s being used as both a productivity lever and a justification for restructuring.
Levie’s point about “AI psychosis” lands because it describes a mismatch between perception and reality. The people deciding that AI can replace your job are often not the ones who live inside that job’s daily constraints. They may understand the job at a high level—what it produces, what it’s supposed to accomplish—but not the hidden work: the back-and-forth, the escalation paths, the stakeholder politics, the edge cases that never make it into training data, the “tribal knowledge” that keeps operations running when the system fails. When those details are missing, AI adoption can look like a clean line on a slide deck rather than a complex transition that requires careful redesign.
What makes this especially risky is that AI can create a false sense of progress. Early pilots often succeed because they focus on narrow tasks with clear inputs and measurable outputs. A chatbot that drafts customer responses might perform well on a curated dataset. An agent that updates records might work flawlessly when everything is formatted correctly. But once the system is deployed broadly, it encounters the real world: incomplete information, ambiguous requests, conflicting priorities, and the need to coordinate across teams. At that point, the question becomes less “Can AI do this?” and more “Can AI do this reliably enough to remove the humans who currently absorb uncertainty?”
That’s where organizations can get trapped. If leadership has already committed to workforce reductions, the company may feel pressure to make the automation work—even if it requires more oversight than expected. Instead of treating AI as a tool that changes the shape of work, the organization treats it as a substitute that should reduce headcount immediately. The result can be a kind of operational whiplash: teams lose capacity, then discover that the remaining staff must spend more time correcting, monitoring, and re-routing work that the AI can’t complete end-to-end.
This is one reason “AI psychosis” isn’t just a cultural problem—it’s a systems problem. Replacing parts of a workflow doesn’t simply remove labor; it changes where labor shows up. If AI handles drafting but humans still need to review, approve, and handle exceptions, the workload doesn’t vanish—it migrates. If AI handles ticket triage but escalations increase because the model misclassifies edge cases, the workload shifts toward escalation management. If AI generates code but engineers still need to validate security, performance, and correctness, the review burden can grow. In some cases, the net effect is efficiency. In others, it’s a different kind of friction.
There’s also a subtler dynamic: AI can compress timelines. When a system can produce outputs quickly, teams may start expecting faster turnaround across the entire process. That can lead to a mismatch between production speed and decision-making speed. Humans may be asked to review more content in less time, or to approve changes without the same depth of understanding they previously had. If the organization doesn’t adjust processes accordingly, quality can degrade, and the cost of mistakes can rise—sometimes far beyond what was saved through automation.
So what does “too AI-pilled” look like in practice? It often starts with a narrative: AI will eliminate repetitive work. Then it evolves into a strategy: automate first, redesign later. Finally, it becomes a restructuring plan: reduce headcount based on the assumption that automation will cover the gap. Each step can be reasonable on its own. The danger is when the organization skips the middle layer—the hard work of mapping tasks to outcomes, measuring failure modes, and redesigning roles so that humans and AI collaborate effectively.
A unique take on the current moment is that many companies are treating AI adoption like a software rollout rather than an organizational transformation. Software rollouts can be planned with relatively clear dependencies: install the tool, train users, monitor usage. Organizational transformation is messier. It requires renegotiating responsibilities, changing incentives, updating governance, and building new feedback loops. It also requires humility about what the system can’t do yet.
When those elements are missing, the organization can end up with a paradox: it invests heavily in AI to reduce labor, but then discovers that the remaining labor is needed more than ever—just in different forms. Monitoring becomes a job. Exception handling becomes a job. Quality assurance becomes a job. Stakeholder communication becomes a job. Even if the company eliminates some roles, it may recreate similar work under new titles, or it may overload the remaining teams until burnout forces attrition.
This is why the “AI psychosis” framing resonates. It points to a cognitive error: believing that AI’s capabilities are equivalent to human competence across an entire role. But competence isn’t only about producing an output. It’s about navigating ambiguity, managing risk, and coordinating with other people. Many jobs are essentially risk-management systems. They exist because someone has to decide what to do when information is incomplete or when the “right answer” depends on context. AI can assist with parts of that, but it doesn’t automatically inherit the same accountability structures.
Accountability is another area where organizations often underestimate complexity. If an AI agent makes a mistake—sending the wrong message, approving the wrong transaction, misrouting a request—who is responsible? The person who pressed the button? The team that designed the workflow? The manager who approved the policy? The company that deployed the model? These questions matter because they determine how much oversight is required. If leadership wants to reduce headcount, it may also want to reduce oversight. But oversight is often the price of reliability. When oversight is reduced too aggressively, errors can slip through, and the organization pays later in rework, reputational damage, or compliance risk.
There’s also the question of how AI changes bargaining power inside companies. When leadership believes AI can replace roles, it can weaken employees’ leverage. People may be pressured to accept new responsibilities without adequate training or compensation, or to transition into “AI-assisted” roles that are effectively higher workload with less support. That can create resentment and disengagement, which then reduces the quality of human-AI collaboration. In other words, the human factor isn’t just about morale—it affects performance.
At the same time, it would be inaccurate to paint AI adoption as purely misguided. There are legitimate reasons companies pursue AI agents. Some workflows truly are repetitive and can be automated safely. Some roles have clear boundaries and stable inputs. Some organizations need to scale quickly and can’t hire fast enough. AI can help them meet demand while maintaining service levels. The problem arises when the organization generalizes from successful pilots to broad workforce replacement without sufficient evidence.
This is where the “adoption outpaces insight” theme becomes critical. Many companies move quickly because competitors do, because investors expect growth, or because internal leaders want to demonstrate momentum. But insight requires time: time to test in production, time to measure error rates, time to understand how users behave when the system is available, time to refine prompts, policies, and guardrails. If the company cuts staff before it has that insight, it may discover too late that the system needs more human support than anticipated.
The workforce reductions reported in the AI-agent context highlight how quickly these decisions can be made. Cutting 22% of a workforce for AI agents suggests that the company believed the automation would cover a substantial portion of work. But it also raises questions that are rarely answered publicly: What tasks were automated? Were those tasks truly end-to-end? How many exceptions occurred? Did the remaining staff absorb the monitoring and correction work? Did service quality improve, remain stable, or degrade? Without those details, it’s hard to judge whether the reduction was a rational optimization or a premature bet.
And the broader tech layoff environment complicates interpretation. If 2026 layoffs are nearing the total pace of 2025, it suggests that cost-cutting is already a major priority. In such environments, AI can become a convenient lever: it provides a compelling story for restructuring. That doesn’t mean AI is the only cause, but it can influence which departments are targeted and how aggressively automation is pursued. When budgets tighten, companies may prioritize initiatives that promise rapid savings, even if the long-term operational risks