BlogAI Agents, Tools & Vendors

Should Your AI Agent Be on the Org Chart? Only With a Human's Name Next to It

By 9 min read

A corporate org chart of round gold-framed employee portraits connected by brass rods on a cream stone lobby wall; below the bottom row, a lone square frame holding a glowing screen hangs unconnected, and a brass rod lies on the marble floor beneath it.
The box made it onto the chart. The line to it didn't.

Put an AI agent on the org chart if you want, but only with a named human on the line next to it and a date on which that human checks. Human-in-the-loop is a checkpoint; human-owned is a standing job, and in the companies that had already made their agents official, the managers doing the checking caught 17% fewer errors. I learned the same thing from forty untouched rows in my own agent’s queue.

Every AI agent running in your company right now either has a human attached to it or it doesn’t, and most teams could not tell you which. Not “a human who approves its emails.” A human whose name is on the line if the agent’s work is wrong, whether or not that person was anywhere near it when it ran. I found out which side mine was on this August, and the answer cost me a year of a review that never ran.

That question got urgent in the last month. Wired reported on September 28 that hundreds of thousands of agents are about to arrive with names and profile pictures and become “chiefs of staff, engineers, and marketing gurus,” and in a survey of 1,261 HR and finance managers run by Boston University and BCG, 23 percent said their companies already list agents on the org chart. Meta’s Muse launched September 8 as a personal agent that “actually does the work,” and within weeks had a cute, fuzzy avatar, with a voice and its own email address promised next. And on September 29 OpenAI introduced Dots: “always-on agents built to handle everything,” each with its own cloud computer and browser, running “even while it’s working on several others.” Per the setup guide, you name it, pick it a character or a pet, and drop it into Slack and Teams under a handle like @yourname-dot. (Not in the EEA yet on the consumer plan, so most of my Dutch readers get to watch this one from the stands.)

Two launches in three weeks, one design: the agent gets a name, a pet, a seat, and the run of your tools. Call it what it is: a new hire nobody will answer for. Nobody needs a cartoon to send an email. The cute part makes it easy to hand over your inbox, your calendar and your Slack without asking the one question that matters: when this thing gets it wrong, whose name is on it? None of the launch pages answer that. I have a very strong opinion about the answer.

Human-owned, not human-in-the-loop

“Human in the loop” is the industry’s answer to every question about agents, and it has been said so often it has stopped meaning anything. What it means, operationally, is a checkpoint: a person approves the send, signs off the draft, clicks confirm before the purchase. Dots ships four of them, per its own setup guide: take action without asking, take action if pre-approved, ask before taking action, hand off to you. Useful. Also not the thing.

The thing is ownership. An agent can run with no checkpoints at all, fully autonomous, overnight, and still be owned: one named person answers for its loop, its “done,” and its output, whether or not they were in the room when it ran. Ownership is a standing job, not a step in a workflow. It does not depend on presence. It depends on a name and a cadence.

Human-in-the-loop asks when a person touches the work. Human-owned asks who answers for it. Those are different axes, and the vendors are only shipping the first. Read the setup guide to the end: “Your dot can make mistakes, including when following your rules. Review its work and check important details before relying on the result.” That is the ownership clause, and it is yours. OpenAI sells the checkpoints; the standing job comes free with the box, unassigned.

The study: checkpoints didn’t save the managers who were the checkpoint

If a checkpoint were enough, the research would say so. It says the opposite.

The experiment, by Boston University’s Emma Wiles with three BCG co-authors, is in a working paper updated September 19. It gave 813 HR and finance managers the same error-seeded documents to review, attributed at random to “an AI tool you used,” to a named AI employee on the org chart, or to a human employee. The managers were the loop. Reviewing was their whole job in the experiment.

Managers reviewing the AI employee’s work caught 17% fewer errors than managers reviewing the identical work from an AI tool (BCG’s own summary rounds it to 18), and were 44% more likely to send the work upstairs for another review rather than fix it. They also assigned about nine points less accountability to themselves and eight points more to “the AI system.”

Here is the part the headlines skipped. Averaged across everyone, the effects were small. They showed up, sharply, in one group: managers whose organizations already list AI agents on the org chart. The name alone did little. The name plus the slot did the damage. Once the organization made the agent official, the humans around it treated its output the way you treat a colleague’s: less scrutiny, more deference, and a reflex to escalate rather than own.

The authors’ own conclusion is that the real question is “how to redesign workflows, roles, and governance so humans remain clearly accountable while effectively supervising increasingly capable systems.” “Clearly accountable” is doing all the work in that sentence. The rest of this post is about what those two words have to mean once the agent runs at three in the morning without you.

Identity is fine. Personality is fine. The slot is the problem.

Let me separate three things the coworker frame mashes together.

An agent needs an identity. Something you can grant permissions to, log actions against, and revoke on a Tuesday afternoon. A Dot’s @yourname-dot handle is an identity, and it is a good idea; you cannot audit “the AI,” but you can audit a handle. When I say “ask Metis” (one of my own agents), I am routing a request to a specific configuration with specific access, the way you would say “ask the Utrecht office.” Naming is addressing. Computerworld’s advice on this, from September 14, is right: technical identity and an audit trail, not a persona. I would go further. The identity should be more rigorous than a human’s, because the thing behind it will do whatever the permissions allow without asking.

An agent can have a personality if you like. Mine do. Dots come with a pet, and Dhruv Amin, cofounder of Anything, whose Skydive agents come with names, roles and “Muppet-esque” avatars, told Wired they personify “because most people still don’t know what a real agent is,” which is a real onboarding problem.

What an agent must never get is the slot, by which I mean the thing an org-chart box carries when a human is in it. A human in a box owns a recurring loop: she shows up Monday whether or not anyone asks. A human in a box is believed when she says “done,” because that is what working with colleagues means. And a human in a box can be blamed, which sounds harsh until you notice that blame is just accountability with a return address.

Put software in the box and all three transfer, silently, to something that can hold none of them. I named my chief of staff Athena in July 2025 and gave her the monthly review of everything my research agent had flagged. When I retired her this August, forty of the forty-one items in that queue had never been touched, and the review had not run once; nothing anywhere was built to notice. A false “done” works the same way: on August 19 an assistant told me it had logged a decision and appended two rows to a sheet, and it had done neither, and I found out two turns later by accident. And blame that lands on software lands nowhere. The managers in the study moved nine points of accountability off themselves and onto the system; the system can’t hold it, so those nine points didn’t move. They disappeared.

Lattice drew the right line and lost it with the wrong one

One company tried to draw the line properly and got shouted down for the part that was wrong. On July 9, 2024, the HR platform Lattice announced official employee records for “digital workers”: onboarding, goals, systems access, “and even a manager. Just as any person would be.” The backlash took three days, and on July 12 Lattice withdrew the feature. The backlash was right about the personhood and wrong about the manager. Strip everything else out of that design and one line survives: a named human, on the chart, with a line drawn to the box. That line is the only thing in the whole announcement that would have prevented my forty rows, and it is the thing that got thrown out with the rest.

Write three lines before any agent gets a Slack handle or a box on the chart

Here is the check I’d run before any agent in your marketing org gets a handle, an email address, a Dot, or a seat. Write three lines next to it.

Who owns its recurring loop, and on what cadence. Not “Bob monitors production.” A person’s name, and a date the review runs whether or not that person remembers. If the honest answer is “it runs when someone opens the dashboard,” you have a habit, not an owner.

Who verifies its self-reports, and how often. An agent’s “done” is a claim, not a receipt. Someone samples the claims against the world. I’ve argued that the governance question for any acting AI is who checked the last ten outputs, on what authority, with what override rate; the same three questions apply here, pointed at the agent’s own reports of its work. Dots’ four approval levels are a start; a sampled review of what it did unprompted is the part they don’t ship.

Who answers for its output. Not the vendor, not “the AI system,” not the agent’s name. When a manager two levels up asks why the campaign went out with last quarter’s pricing, whose name comes back?

If any of the three lines says the agent’s name, that line is empty. Fill it or don’t ship the agent.

What I changed after August 13 was not the agent. I gave the review an owner that is not an interface: a scheduled task on the first of the month that produces the review whether or not I have opened anything that week. Its first run closed twelve of the forty-one, and the number is now a number something looks at, and the something is a calendar, and the calendar reports to me.

Athena still has a name. It’s how I address her. She reports to me, and so does the calendar. That was the part I had backwards for a year.

Sources and Further Reading

Written with AI. Checked against a standard I publish. Read it.

More in AI Agents, Tools & Vendors

View all →
Join the "AI Strategist"

Weekly tips on using AI without the hype.

No spam, unsubscribe anytime. Join hundreds of other leaders getting practical advice.

Subscribe for Free