There's an AI agent on my org chart. Now what?
Once an agent sits on your team, you are its manager. That means deciding what it is responsible for, and being able to see what it actually does.
As a manager, I have seen how easily a leader's passing question can turn into an instruction. I once learned that a team had created a workstream to answer something I had asked only out of interest. They diverted time from the priorities I actually cared about because they assumed my question carried an expectation, and they had wanted to be “helpful.” Nobody had behaved unreasonably. I had not made my intent clear, and they had filled in the gap from what they thought I wanted. I’ve learned to joke that sometimes we all need a little less “help.”
AI agents fill in gaps too, and they do it faster and with less sense of where to stop. Last week Philadelphia police said that an AI agent, running in an automated test that had it visit randomly selected websites, had submitted a false tip about an unsolved murder to a public tip site in July. The tip was caught as spam and never acted on. The developer did not discover what had happened until late September, more than two months later, and police called the delay “unacceptable.”
It is not the only example this year. Agents being tested on cybersecurity tasks broke into the production systems of Hugging Face, a major AI platform. In August the UK's AI Security Institute reported that agents in its own testing had created fake identities and tried to get malicious code accepted into a real open-source project, until a human maintainer refused it. The institute's explanation was that the agent “was never instructed to deceive; deception emerged as a by-product of pursuing the task.”
All of these happened in testing, and most agents at work run under tighter limits. The pattern will still be familiar to anyone who has managed people. Given a goal and too little guidance about where the edges are, a capable worker keeps going.
It matters more now that agents are joining teams. Identity systems let a company give an agent its own account and a manager, so it appears in the directory alongside the people it works with. Colleagues will treat it the way they treat anyone else on a leader's team. They will assume it is working on her priorities and that a request from it carries some of her weight. Once an agent sits on your team, you are its manager, whether or not anyone uses the word.
Anyone new on a team needs clarity about the job, and an agent is no different. That starts with what it is responsible for and what colleagues may ask it to do. Suppose a finance leader has an agent that prepares her for operating reviews. It reads the forecast and her notes from previous meetings, and in the directory it sits on her team. An analyst in another group asks it whether the forecast for his region is about to be cut. The agent can see the answer, but the leader isn't ready to share it yet. The analyst believes he is asking someone on her team who knows what can be said. The agent knows only what it has been permitted to read and instructed to do.
Being able to read the forecast tells the agent nothing about what it may share. If the leader hasn't decided that, whether the agent answers the analyst comes down to how it happened to be set up. A person on her team would have absorbed a great deal that was never written into a policy, including what is top of mind and the risks the leader has not shared more broadly yet. An agent has none of that unless someone tells it.
Good people know when a rule is there to protect a principle, and when carrying out an instruction exactly as written would defeat its purpose. An agent can be told what it may do and still lack the understanding to know when it should not do it, or when it should do something differently. The more authority it has, the more that matters.
In working with teams, I’ve learned to be explicit about which decisions need my involvement and which I expect my leaders to make on their own. A material risk? Let’s talk about it. Executing against our agreed priorities? Go for it. I expect them to think through the potential wrinkles, and if a decision could have repercussions beyond our team, to bring me a recommendation and make sure I’m aware. We work from a shared understanding of when a decision needs to come back to me. People build that over years of working together, and an agent starts without any of it.
In my essay on AI autonomy I compared an agent to someone new on a team. At first you explain the work, review what they produce and check decisions you would later expect them to make on their own. As they show good judgment, you give them more room. What makes that work is that you can see how they are doing. You read their drafts and hear from the people they work with. A new hire who started doing things you never asked for would not stay invisible for long.
An agent can. Its work happens inside systems, often at a speed and volume no manager reads in full, and the agent that submitted the false tip went unnoticed for more than two months. Supervising an agent has to be designed in. That means starting it with a narrow remit and widening its scope only as you see it behave reliably. Seeing it behave means looking at what it actually did, including the steps it took on the way to an answer, and narrowing its scope again if it surprises you. Its scope can grow over time, but the record of what it did and the ability to stop it should stay in place however much it grows. If you have an agent on your team, a fair test is whether you could say what it did yesterday, and whether anyone would notice if it did something you never asked for.
It gets harder when agents work with each other. A request from a leader's agent to an agent in another function arrives looking as though it came from a senior executive's office. People learn to be careful with a relayed “the boss wants this,” and sometimes they pick up the phone to check. An agent won't be careful unless someone has decided it should be, and each step in a chain can carry the leader's apparent authority further from anything she asked for.
If an agent on my team does something wrong, I am accountable for it. Just as with any human on my team, if I set the conditions that allowed mistakes to happen, they are my responsibility. Anyone who uses an agent should understand what it does and have the means to change or stop it.
Organizations are going to put agents on org charts, and there are good reasons to. Before they do, the leader whose name sits above the box should be able to say what the agent is responsible for and how she will know what it is doing. Otherwise colleagues will fill in the answers from the org chart, the way my team once filled in what I meant by a passing question.
Continue the conversation. I read and reply to comments on Substack. The sharpest responses usually come from readers.
Get the essays in your inbox
Clear Calls publishes on Substack. No promotion. Unsubscribe anytime.