Supervising AI All Day Is Exhausting. Nobody Scheduled Relief.

On-call used to be unbounded too. Before reliability engineering gave it a name, engineers got woken by whatever phone call came, with no rotation, no paging policy, and no acknowledgment that the load itself was the job. What fixed it was not better alerting. It was naming the load and designing around it.

There is a new load now. It has no name, no rotation, and no relief: supervising AI output, all day, every day.

This is not a piece about whether AI is good. It is about a category of work we added to engineering without a line item and without admitting it is work.

What the Load Actually Is

The Stack Overflow 2025 survey, with roughly 49,000 respondents, asked developers what frustrates them most about AI tools. The top answer, at 66%: solutions that are almost right, but not quite. Close behind, at 45%: debugging AI-generated code takes longer than expected.

Put them side by side and the shape of the day appears. Almost-right output is worse than wrong output, because wrong output announces itself. Almost-right output looks finished. Catching it takes attention that stays open, with no clear endpoint.

That is vigilance, and vigilance is one of the most expensive kinds of attention we have. It is why nobody puts security guards on 16-hour shifts, and why we stopped letting one engineer carry the pager alone.

The Human Is the Reliability Layer

The survey also carries a quieter number: 75% of developers say that even with advanced AI, they would still ask another person for help when they don't trust AI's answers.

That is the honest architecture of the AI-assisted workflow. The speed comes from the model. The reliability comes from a human staying alert enough to catch the almost-right problem. The error rate hides inside plausible code, and the error catching is an unmeasured, unrotated human function.

Describe that design for any other system, "the safety mechanism is a person paying very close attention indefinitely," and you get laughed out of the review.

What On-Call Teaches Us

Reliability engineering already fought this battle, and the playbook transfers almost directly.

Name it. Verification fatigue is not a buzzword. It is the precondition for everything else. Nobody fixes a load they have not named.

Measure it. On-call got healthy when teams started counting pages, not just outages. The equivalent here: hours spent reviewing generated output, and rework traced back to almost-right code.

Rotate it. One person supervising agent output all day is the on-call of 2009. Rotate who holds the reviewer role, daily or weekly, and make the handoff explicit.

Bound it. Batch verification into fixed windows instead of holding attention open all day. Three 45-minute review blocks beat nine hours of low-grade waiting.

Set thresholds. Decide in advance what auto-merges and what always escalates. A trust threshold is not trust. It is a policy about where the vigilance actually has to live.

💡 Pro-tip: if your team's review of AI output has no schedule, no owner, and no rotation, you don't have an AI workflow. You have an on-call rotation nobody designed.

What Managers Should Watch For

Fatigue in this mode does not look like exhaustion. It looks like trust. 😶

The developer who used to ask sharp questions about generated code and now says "looks fine" is the one to check on. Review latency creeping up is another signal. So is the quiet disappearance of the second opinion.

None of those people are slacking. They are doing what humans do under sustained vigilance: adapt by lowering the bar. The system is working exactly as designed. That is the problem.

Not an Anti-AI Argument

Some verification work is genuinely engaging. Reviewing a strong agent's output can be the best part of a day, and calling all of it fatigue would be sloppy. The claim is narrower: sustained, unbounded vigilance is a known failure pattern, and we have rebuilt it with better tooling and zero relief.

There is also a risk that "fatigue" reads as anti-AI. It is not. The teams that name this load early are the ones that get to keep the speed.

On-call did not go away when we named it. It got rotations, budgets, and comp days, and it got humane. Supervision deserves the same upgrade.

How many hours did you spend supervising AI output this week, and who is counting?

References

1. Stack Overflow 2025 Developer Survey, AI section (66% "almost right" frustration, 45% longer debugging, 75% would still ask a person): survey.stackoverflow.co/2025/ai