I have started counting the number of times I see an AI agent introduced as someone.

"Meet Bailey, our junior analyst. She'll be helping the team with the Q3 model."

The pause for a fake "hello." The org chart updated to show a new box. The Slack handle. The auto-generated headshot, earnest and slightly distracted. The HR note that Bailey has been "onboarded."

I do not like it.

Two weeks ago I was on a 1:1 coaching call with the head of strategy at a mid-market private equity firm. He pulled up his Notion to walk me through how he had reorganized his sourcing function. New row on the team page: "Iris — Junior Associate, Sourcing." Slack handle. A reporting line. He had moved her under one of his real associates the previous Friday.

I asked who Iris was. He said, with no irony, that she was their new AI sourcing agent. His associate was managing her now. When I asked what that meant in practice, he said his real associate had stopped reading Iris's outputs as closely because — and this is his phrasing, not mine — "she's been performing well." Two weeks on the job. Performing well. He used the same language he uses for his human analysts at the end of Q1.

I asked him to pull up the last deal brief Iris had touched. Two factual errors in the first paragraph. Neither had been flagged. When we got his associate on the call, he looked genuinely surprised. He had drifted into a register he uses with people he trusts.

The framing has spread faster than I expected. In a survey released last week, twenty-three percent of managers said their company has already placed AI agents on the formal organizational chart. Not in a sandbox. Not as a tool the team uses. As a unit. As a thing with a manager, a reporting line, a name.

The pitch makes intuitive sense. If we treat AI like a colleague, the thinking goes, people will work with it more naturally. The technology stops feeling foreign. Adoption goes up. The company becomes "AI-native." Boards like the story. Investors like the story. The press release writes itself.

Last week, a team from Boston Consulting Group, Boston University, and the MIT Initiative on the Digital Economy ran the experiment. The results are not what the people printing the org charts were hoping for.

Emma Wiles, an assistant professor at BU and an MIT IDE fellow, and her co-authors gave more than a thousand managers the same set of documents. Every document had real errors baked in — flawed numbers, weak logic, a few obvious holes. Then they varied one thing. Some managers were told the document had been produced by an AI tool. Others were told an AI employee had produced it. Others, a human employee.

The documents were identical. Only the label changed.

The managers who thought the work came from an "AI employee" caught sixteen percent fewer errors than the managers who thought it came from an "AI tool." They sent the work back for additional review forty-four percent more often. They reported that the AI was more accountable for the document's contents. They reported they themselves were less accountable.

Three words on a label moved the locus of responsibility off the person and onto the machine.

The researchers ran the test inside organizations that had not yet put AI on the org chart. The effect almost vanished. The framing did not become real until the company made it real. Once Bailey has a Slack handle and a box in the chart, the manager unconsciously starts treating Bailey like she has skin in the game.

She does not. She cannot.

What is worse, the same framing did nothing for adoption. The managers in the AI-employee condition were not more likely to use the AI in their workflows. They were not more enthusiastic about it. They were not more productive. They were just less likely to catch errors.

Read that one more time. We are paying a tax in error catching and accountability — and getting nothing back in return.

I have a theory about why this happens, and the paper supports it. English hides a long-standing convention. When we say a person did something, we mean a person can be held to it. When we say a tool did something, we mean a person used the tool and is responsible for the outcome. That second convention has been load-bearing for centuries. It is how we know who pays when the contract is wrong, who apologizes when the report is late, who gets the bonus when the deal closes.

When we say an AI employee did something, we have created a third category nobody knows how to read. The brain does what the brain always does — it reaches for the human-shaped version. We treat the thing as a colleague, which means we treat the colleague as responsible, which means we relax our checking. The model does not know it is on the org chart. The model does not care. But we have already started doing the work of believing it does.

The move I'd make this quarter is to run a language audit on every system your team touches. Read your project trackers. Read your meeting summaries. Read your performance reviews. Every place an AI tool has been given a name, a pronoun, or a reporting line, change it back. This is not a culture initiative or a values exercise. It is a quiet correction to the words you let into the building, made before the cost of those words shows up in the work your team ships. Do it before the next board update, not after.

Three practical consequences fall out of this for anyone running a company right now.

Take the AI off the org chart. The presentational language is doing damage that the productivity gain is not making up for. Call it a tool. Call it a system. Call it Drafting Assistant Version 4. Do not give it a face and a fake personality. Do not list it as a direct report. The version of the world where AI is "a member of the team" is not the version you should be optimizing for.

Keep the human name on the work. Whatever the AI produces, a person owns it before it leaves the building. That person's name is on the deliverable, the analysis, the email to the client. The AI's involvement can be disclosed in the metadata. The accountability does not get split with a thing that has no skin to bring to the table.

Audit the labels you already use. Walk through the language inside your firm and look for places where you have quietly anthropomorphized. The way people describe AI inside the company is creating expectations that show up in error rates a quarter later.

The AI-as-employee model is being sold as the next obvious step in how organizations evolve to accommodate intelligent systems. The new research suggests something less flattering. The model does not improve adoption. It does not improve output. It moves accountability away from the humans who can hold it. That is not evolution. That is shrinkage.

Most days, this language stuff feels like a side issue compared to actual capability gains. The paper measures, in plain numbers, how much it isn't.

Take the AI off the org chart. Put the human name back on the work. Make sure that when something goes wrong, the line about who owns it is a short one.

There is no Bailey. There is only your team, and what your team chose to ship.