We run AI agents in our own firm. They draft, they summarise, they route work between the three of us, and they sit in the same Slack channels we do. I have started watching myself when I read what they produce, and I do not entirely like what I see.

The temptation is small and constant: to treat the output as something a colleague handed me rather than something I am responsible for. To skim where I would have read. To think "that is good enough from the agent" in a way I would never think of a draft from a partner, because a partner's draft has a person behind it who will be in the room when it goes wrong. The agent will not. And yet the longer it sits in the channel with a name and a function, the easier it becomes to forget what it is, and what it should and should not be doing.

That observation is private and anecdotal. A BCG Henderson Institute study with Boston University now shows it is neither.

An org chart is not a list of who works here. It is a list of who answers. Every box on it carries three things that are easy to forget because they are so familiar: a scope of decisions the person in the box may take, a signature that binds the organisation when they take them, and the possibility of removing them if they take them badly. Strip those three out and what remains is a name in a rectangle.

Yet that is precisely what a growing number of companies are drawing. The study surveyed 1,261 HR and finance managers, directors and executives and found that 23% work in organisations that now list AI agents on their org or workflow charts; nearly a third say leadership frames AI as a teammate or employee. The agents have names, often avatars, sometimes titles and monthly performance reviews. "Alice" ships code. "Bob" runs the calendar. The intent is benign: make the technology legible, lower the resistance, help people see where it fits.

What the researchers did next is the useful part. They gave those managers identical documents seeded with errors and varied only the attribution — produced by a human colleague, by an AI tool the manager had used, or by an AI employee. Among managers at organisations that had already institutionalised agents on the chart, labelling the work as an AI employee's cut the share of errors caught by 17%, raised costly requests for someone else to review it by 22 percentage points — a 44% increase — and moved perceived responsibility for the mistakes about nine points off the manager and eight onto "the AI system". Managers in organisations without agents on the chart showed almost none of this. The name alone does little. The box does the damage.

So the box is not neutral. The moment an agent is institutionalised as a colleague, the humans around it start behaving as if the colleague could carry its own accountability. It cannot. The result is not that work gets done with less oversight; it is that oversight is passed sideways, with more handoffs and a lower hit rate at each one. The researchers call it the hot-potato effect. The error does not disappear. It travels.

The obvious objection is that personification works, and it is partly right. Naming reduces friction. People say "I worked with Alice on this" because it is a sentence they already know how to say, and a workforce that can describe its tools is more likely to use them. Lattice, which put digital employees on its chart two years ago, argues the move creates accountability by making the agent visible. Visibility is genuinely the right goal.

But the study tested that intuition directly, and it failed. Managers handed a named AI employee were no more inclined to adopt AI, learn it or invest time in it than those handed a tool. One participant put it in a sentence I have not been able to improve: if you want people to feel they can be replaced by AI, put AI on the org chart. The name buys fluency. The box buys unease and worse review.

A European reader will object that this is an American problem, and for now that is largely true: the sample is overwhelmingly US-based, nobody has measured the practice here, and in the DACH boardrooms I see it remains rare. The lag is not only cultural. In much of Europe the org chart is not a slide the CEO owns alone — in co-determined companies, changes to reporting lines and to how work is monitored involve the Betriebsrat — and Article 14 of the EU AI Act requires that high-risk systems be overseen by identifiable people. For those uses, the question of who owns the agent is not a governance preference; it is a precondition. Europe will adopt agents more slowly. It will also find it harder to put one on the chart without a human beside it.

This is why the two have to be separated. A name is an interface decision. A box is an accountability decision. Agents can carry the first; they cannot carry the second, because the three things a box requires — bounded authority, a binding signature, a removable person — are not properties software can have. A Geschäftsführer can be held to account for what the company did. An agent cannot be held to anything. Drawing it as a peer does not change that; it only hides it.

The design follows almost mechanically. An agent belongs on the chart the way a team's ERP belongs on it, which is to say inside a human's box, as something that human owns. Give it a name if your people work better that way, but the line that connects it to the rest of the organisation runs to a person, and that person's name is the one that appears beside the agent's output, in the review and in the incident report. Review standards are set by the consequence of the decision, not by the author of the draft — a budget that goes to the board is checked as a budget that goes to the board, whoever or whatever produced it. And when something goes wrong, "Alice made a mistake" is not a finding. It is the sentence that tells you nobody owned Alice.

The researchers reach the same place by a different road: their closing recommendation is to pair every agent with an explicitly accountable human owner, set minimum review standards for high-stakes decisions and say who keeps responsibility when work is passed around. That the vendors selling named agents say the opposite is worth noting.

There is a harder consequence underneath. Many organisations are adding agents faster than they are deciding who answers for them, which means the chart is being redrawn by procurement rather than by the board. Each agent that arrives with its own name and no owner is a small, undeclared transfer of decision rights to nobody. The lawyer in me notes that regulators, courts and auditors will not accept the transfer. They will find the nearest human and attach the consequence there.

Which brings me back to our own Slack channels. I have not removed the agents, and I have not taken away their names. I have made one change: every one of them now has a partner's name next to it, and that partner reads its output as if he had written it himself. In a firm of three, that was an afternoon's work. In a firm of three thousand, it is the governance decision of the decade — and it is being made right now, by default, by whoever installed the software.

Name the agent if you must. The box belongs to the human.