AI 'Psychosis' in the Workplace: When Replacing Developers Costs More Than It Saves

The "AI Psychosis" Problem

Box founder Aaron Levie has a name for what is happening at companies like ClickUp: "AI psychosis." The diagnosis: executives who are furthest from the actual work are making the most aggressive decisions about replacing it with AI. ClickUp recently laid off 22% of its workforce to deploy AI agents in their place. The 2026 tech layoff wave is already approaching 2025's full-year total, with AI replacement cited as a primary driver.

The financial logic seems straightforward: a developer costs $150,000–$250,000 per year in salary and benefits. An AI agent running on Claude Sonnet 4.6 at $3/$15 per million tokens costs a fraction of that. But the math breaks down quickly when you account for what AI agents actually cost to run at production scale, what they cannot do, and what happens when they fail.

The Real Cost of Running AI Agents at Scale

AI agent costs are not just token costs. A production coding agent that handles real software development tasks involves:

The human review line is the one that surprises most executives. AI agents in 2026 are not autonomous. They require oversight, correction, and direction from experienced engineers. The ratio varies by task type, but a realistic estimate is that one senior engineer can effectively supervise 3–5 AI agents doing well-defined tasks. That supervision cost is real and must be included in any honest ROI calculation.

What AI Agents Cannot Do (Yet)

The tasks where AI agents genuinely replace developer time are narrower than the hype suggests:

• Well-specified, bounded tasks. Writing a function with a clear spec, adding tests for existing code, updating documentation — these work well. Open-ended tasks with ambiguous requirements do not.
• Tasks with clear success criteria. If you can write a test that definitively passes or fails, an agent can iterate toward passing it. If success is subjective or requires business judgment, agents struggle.
• Tasks in well-understood codebases. Agents perform better in codebases with consistent patterns, good documentation, and clear architecture. Legacy codebases with implicit knowledge and undocumented constraints are where agents fail most expensively.

The tasks that consume most of a developer's time — understanding requirements, navigating organizational dynamics, making architectural tradeoffs, debugging subtle production issues — are precisely the tasks where AI agents are weakest. Companies that replace developers wholesale discover this the hard way.

The Hidden Costs of Getting It Wrong

When AI agent deployments fail, the costs extend beyond the token bill:

• Technical debt accumulation. AI agents that generate code without deep understanding of the codebase tend to introduce subtle inconsistencies, duplicate logic, and architectural drift. Cleaning this up requires experienced engineers — the ones you laid off.
• Rehiring costs. Companies that over-cut and then need to rebuild engineering capacity face a 6–12 month hiring cycle plus 3–6 months of onboarding. The total cost of a bad layoff decision can exceed two years of the salaries saved.
• Competitive velocity loss. Software companies compete on shipping speed. If AI agents slow down feature delivery because they require more oversight than the humans they replaced, the competitive cost can dwarf the salary savings.

The Right Model: Augmentation, Not Replacement

Cognition's Scott Wu — whose company built Devin, arguably the most capable AI coding agent available — has been explicit: AI coding agents are not meant to replace human programmers. They are meant to make human programmers dramatically more productive.

The companies getting the best ROI from AI coding tools are not the ones replacing developers — they are the ones giving each developer AI tools that multiply their output. A developer with Claude Code, Cursor, and well-configured agent workflows can produce 2–4x the output of a developer without these tools. That is a better return than replacing the developer with an agent that requires 0.3 developers to supervise.

The math on AI agent costs is real and worth doing carefully. Use the AI Cost Estimator to model what your actual agent workloads would cost across different models and usage patterns before making staffing decisions based on AI cost assumptions.