A manager with 6 people and 12 AI agents does not have a span of 18

Span of control was a proxy for how much supervision one person can carry. AI agents make that count unreliable, and most companies still count span the old way.

Span of control has always been a proxy. Nobody ever believed that 8 was the right number of humans to report to a manager. What we believed is that a manager has a finite amount of attention, that a person absorbs a predictable share of it, and that counting people was a workable stand-in for counting attention.

AI agents make the count wrong in both directions: in a normal week they need far less attention than a person, and in a bad week far more.

They take less, until they take more

An agent does not need a career conversation, does not get demotivated and does not have to be told the same thing twice. On a normal week it costs a manager almost nothing. That is the argument people make for adding agents underneath the same manager and leaving the structure alone.

But the week an agent is wrong, it is wrong at volume and without hesitation, and it looks right while it does it. A person who is out of their depth signals it. A model that is out of its depth produces the same confident prose it always does. So the supervision an agent needs comes all at once, and it lands on whoever signs.

An agent’s output looks the same whether it is right or wrong, so a reviewer cannot skim it.

What actually gets bigger

The part of a manager’s job that grows is reviewing output. Somebody has to be answerable for output that no human wrote. In a regulated setting that person has to be named, and they have to be able to stop the thing. That is a different job from managing 6 people, and it does not scale by widening a span.

This is why the layers that come out of a chart after an AI programme are rarely the layers everyone expected. The coordination layers go, because there is less to coordinate when one desk produces the work. The review capacity has to grow, because the volume it has to stand behind went up.

What to do with the number

Stop reporting one span. Report two. Count the people a manager develops and count the output streams that manager is answerable for, and size the job on both. Where the second number is much higher than the first, either the work is waiting on that manager or they are signing things they cannot really check.

This is the second of the four design questions AI agents actually raise, which also carries a 30 second test of your own numbers.