The Human Overhead of Agentic Systems

The business case counted minutes saved. It did not count the new job that appeared around the minutes: the person who reads what the agent drafted, the person who handles the 8% it cannot, the person who apologizes to the client when the tone is wrong, and the person who retrains the prompts every time the vendor silently ships a new model. That is not "change adoption." That is overhead. It has a body. It has a calendar. It is almost never in the deck.

I am not arguing against agents. I am arguing against a fantasy of labor disappearing rather than moving. Work moves. If you do not watch where it lands, it lands on the most conscientious person in the room, which is how you convert your best people into unpaid infrastructure.

The three layers you are not costing

Review load. An agent that drafts forty summaries a day has not removed forty summaries of attention. It has created forty opportunities to miss something. Attention does not scale at the same slope as generation. Leaders who say "just skim it" have never been the person who missed the one number that mattered. Skimming is a skill with a failure rate. Budget the failure rate as labor, not as culture.

Exception handling. Every agent has a tail. The tail is where the work is interesting, political, or sad. Tails used to be mixed into ordinary caseloads. Now they are concentrated onto whoever is "good with the weird ones." That concentration looks like expertise. It is also a load-bearing wall with no backup. When that person is out, the agent looks suddenly worse, and everyone is surprised.

Emotional supervision. Someone has to absorb the team's relationship with the system — the irritation, the attachment, the quiet shame of needing it. That labor used to be called management. It still is. It got denser. When the Digital Teammate Goes Quiet is what happens when you skip it at the end of life. The overhead is also there in the middle, every week, in the one-on-ones that are supposedly about "performance" and are actually about status and grief.

A cost you can put on one page

For a single agent, write three lists.

1. Tasks that vanished. Be honest. Many "vanished" tasks still exist as review.
2. Tasks that appeared. Prompt maintenance, evaluation spot-checks, escalation design, tone repair, shadow-process hunting.
3. Hours, named. Not "the team." Maya, 4 hours. Andre, 6. The lead, 3 of apology.

If list two is longer than list one, you did not buy productivity. You bought a different shape of work and called it a win because the vendor's unit of measurement was generation, not attention.

Share the page with whoever signed the contract. This is not an ambush. It is how you keep the human as the moral agent of the business case. A case that cannot survive contact with named hours was never a case. It was a wish.

Why the conscientious person is the risk

Organizations punish visible refusal and reward invisible absorption. The person who says "I cannot review this volume safely" looks difficult. The person who reviews it on Sunday looks loyal. Loyalty is how overhead hides. It is also how you lose the only people who still notice when the agent drifts.

This is a fairness problem that does not show up in a dataset. Opportunity hoards around whoever still has time to do visible, original work, while the absorbers fall behind on the work that gets promoted. I take that up in Fairness After the Spreadsheet. The overhead view is simpler: if you do not rotate the review load, you are running a caste system and calling it a workflow.

What good looks like

A professional-services group I trust did three unfashionable things when they added an agent to proposal work:

• They cut the proposed volume target by 30%, not raised it, for the first quarter.
• They scheduled review as a meeting with a start and an end, not as "async whenever."
• They named a second exception-handler on day one, and paid for the training.

The agent still saved time. The savings were smaller than the deck. They were real. People stayed. When they later retired a related tool, the continuity rite in How to Retire an Agent Without Retiring the Team had somewhere to sit, because the labor had already been visible.

Keep the human as the moral agent

Do not tell the board you have "automated the function" if a person is still the function, plus a queue. That sentence will be remembered when the queue breaks. You may celebrate generation. You may not disappear the supervisor.

The moral question is not whether the agent is impressive. It is whether you would put your name on the staffing plan that actually exists, as opposed to the one in the slide.

This week

Pick the agent you have bragged about most recently. List the human tasks it created. Put hours and names. If the names surprise the person who signed the contract, you have found the real implementation gap. Close it in staffing, or reduce the agent's scope until the hours fit.

For the questions that keep a business case honest, use the Ethical AI Leadership Decision Toolkit. The field notes continue in EI Leadership Insights.