Backend developers at small tech companies are reporting a specific kind of exhaustion: not from working harder at the same job, but from doing the jobs of people who no longer work there.
The pattern goes like this. Management sees that Claude and similar AI coding assistants genuinely speed up certain development tasks. They conclude that one developer with AI can do the work of two or three developers without it, and cut teams accordingly. A backend team of 4 becomes a team of 2. Output expectations stay the same or grow. The productivity gains go to labor cost savings, not to the developers who are generating them.
This is a predictable outcome once you understand what AI coding tools actually do well. They're genuinely fast at bounded, mechanical work - writing boilerplate code, generating test cases, explaining unfamiliar library syntax, drafting database queries. For a developer spending 30-40% of their day on that kind of work, these tools can cut that time meaningfully.
What Gets Left Out of the Productivity Math
The remaining portion of a developer's day is harder to accelerate: architectural decisions, debugging failures in complex systems, code review, stakeholder communication, understanding why something that looks correct is actually broken. AI tools help at the margins here, but they don't eliminate the thinking.
When headcount gets cut, that work doesn't disappear - it gets redistributed to fewer people. And there's a specific overhead that comes with AI-assisted coding that rarely makes it into management's calculations: you're responsible for reviewing and debugging code you didn't fully write. That means constantly switching between author mode (understanding the intent) and reviewer mode (verifying correctness), on code generated quickly and at volume.
Claude Code and similar coding assistants genuinely compress certain tasks. But a developer reviewing AI-generated code under time pressure, across areas they now own alone because teammates were let go, is carrying more context and more responsibility, not less.
The Numbers That Don't Add Up
"AI can do the work of 2-3 developers" is a claim that sounds precise and isn't. It might be accurate for a narrow subset of tasks in a specific context. Applied as a blanket headcount multiplier across an entire engineering team, it produces teams that are technically functional but quietly burning through their capacity.
The tools themselves aren't the problem. AI coding assistants are worth using - they make real tasks faster in demonstrable ways. The gap is between what these tools actually do and how management is using them to justify staffing decisions. Companies getting this wrong aren't just burning out their developers; they're accumulating technical debt from AI-generated code reviewed too quickly by people stretched too thin to catch the subtle errors that compound over time.