The Org Chart With Agents in the Boxes

The whiteboard in the Sorrento Valley conference room still had last quarter’s roadmap ghosted into it, the way dry-erase never quite leaves. The COO had drawn a fresh org chart over the top of it anyway — boxes, reporting lines, the usual pyramid — except four of the boxes had a small handwritten label in the corner instead of a name. Agent. She tapped one of them with the marker and asked me what I thought. What I thought was that she had drawn the existing company and then erased four people, and that this was the most common mistake I was seeing that year, drawn very neatly on a wall.

I didn’t say that. I asked her who the four agents reported to.

There’s a survey number making the rounds right now — something like eighty-five percent of organizations saying they want to operate “agentically,” paired with the quiet admission that almost none of them have closed the gap between that ambition and anything running in production. I believe the number, in the way I believe most numbers about intentions. People want this. They have read enough to know they should want it. What they have not done is the part that doesn’t fit on a slide, which is figure out what an agent actually is inside an organization that was built around humans.

An agent is not a cheaper employee

The reason the redraw stalls is that the org chart is an accountability diagram wearing a headcount costume. Every box on it answers a question that has nothing to do with who does the typing: when this goes wrong, whose afternoon does it ruin. That is what a reporting line is. It is a pre-negotiated answer to the question of who picks up the phone when the number is bad.

When you write Agent in a box, you have specified the typing and left the accountability blank. The software will draft the email, reconcile the ledger, triage the ticket, summarize the deposition — and when it does one of those things wrong at two in the afternoon on a Thursday, the org chart has no answer for whose afternoon that is. You cannot escalate to the box. The box does not have a manager who can be asked what happened, cannot be coached, cannot be put on a plan, cannot quit. It can only be switched off by someone who was never named as the person allowed to switch it off.

This is the thing the neat drawing hides. Swapping a human box for an agent box looks like a substitution. It is actually a deletion — of a person, and of the entire accountability structure that person was standing inside without anyone noticing, because humans carry their accountability around with them for free.

You are blocking a scene, not filling a seat

The closest discipline I know to this is the one I came up in: putting a show on a stage. A director does not cast a production by listing the roles and finding a body for each. The roles are the easy part. The hard part is blocking — deciding who is responsible for what at every moment, who covers if someone misses an entrance, who holds the scene when the lead goes up on their lines. The stage manager isn’t on the org chart of the play, but everyone in the building knows that when something breaks mid-performance, it is the stage manager’s afternoon. That clarity is the entire reason the show can run eight times a week without the building burning down.

Agentic org design is blocking, not casting. The question is never “which boxes can become agents.” The question is “for each thing this agent will now do, who is the human standing just offstage who owns the result.” Sometimes that human is the person whose old box you just erased, now doing the thing one level up — reviewing, exception-handling, owning the number instead of generating it. Sometimes it’s someone who didn’t exist on the chart at all and now needs to. Either way, you have not reduced headcount by four. You have changed what four people are accountable for, and added a layer of supervision that the agent requires precisely because it cannot supervise itself.

The line that has to be true

Before any of the agent boxes ship, I make clients write one sentence per box, out loud, in the room: “When this agent produces a wrong result, [named person] owns it, will know within [stated time], and is allowed to turn it off.” If they can’t complete that sentence for a box, the box is not ready. Not because the model isn’t good enough — the model is usually fine — but because they have automated a task without relocating its accountability, and an unowned automated task is just a faster way to produce a problem nobody is watching for.

The COO in Sorrento Valley redrew her chart that afternoon. It had more boxes on it when she finished, not fewer — the four agents were still there, but each one now had a human name beside it and a line running to a person who could be asked, at two on a Thursday, what happened. That is the real reorganization. The agents were the easy part. They were always going to be the easy part.

If you are staring at a wall with Agent written in a few of the boxes, don’t ask which other boxes could join them. Ask the offstage question first: name the person who owns each result, and the time within which they’ll know it went wrong. The chart that survives contact with a real Thursday is the one where every box — silicon or otherwise — has a human standing next to it who picks up the phone.