The Rehire After the Agent Was Enough

The vendor slide said the assistant would handle the queue. The business case subtracted the seats. Six months later the same seats came back under a requisition that didn't mention the model.

Leaders treat that as a forecast miss. It's an incident. The harm was absorbed by the people who stayed, by the customers who waited and by the person who covered exceptions until recruiting caught up. Then the organization hired the labor back and called it "hybrid oversight," as if the first story had never been told.

The Near Miss Nobody Logged is the catch that never became a ticket. This is the catch that became a job posting. Same pattern. Later, more expensive, still unowned.

What the metrics missed

Throughput stayed green while containment quietly failed. The agent closed tickets that weren't done. Escalations didn't vanish, they changed costume. They arrived as callbacks, as chargebacks, as a tone in a reply that a fluent system won't flag because fluency was the objective.

The people who knew this were the ones still in the queue. They were also the ones whose hours were supposed to have been freed. The Human Overhead of Agentic Systems is that labor while the dashboard is still celebrating. The rehire is what it looks like when the labor can no longer be denied and still isn't named as learning.

A rollback isn't a scandal if you treat it as data. Aviation would study the conditions: what the model was allowed to close, who was allowed to reopen, how long the exception sat. Most AI programs study the vendor. They ask for a better prompt. They don't ask who paid for the six months.

Headcount is a receipt

If you have to hire the role back, the original control was a wish. Write it that way.

• The system that was supposed to complete work didn't complete it.
• A person subsidized the gap until the subsidy became a hiring plan.
• The public story ("the assistant handles it") and the practice diverged long enough to spend trust.

That's not anti-automation. It's the difference between a pilot you can learn from and a pilot you launder. Scale-back and rehiring after an agentic rollout are now a class of event, not a one-off embarrassment. Treat them like near misses with a payroll line.

Someone still has to be allowed to pause the path that created the gap. If the override was unnamed when the seats disappeared, it will still be unnamed when they return and you will run the same six months again.

This week

Pick one agent that was sold as headcount-neutral. Write three lines: the date the seats left, the exception volume that made the work unsafe, the date the seats were requested back. Put the page next to your near-miss log. If you can't fill the middle line, you don't have a hybrid model. You have folklore.

The person remains the one who can be asked why. A rehire shows the first story was wrong, and someone already paid for it.

The five questions in the Ethical AI Leadership Decision Toolkit are built for this kind of page. Field notes continue in EI Leadership Insights.