The AI Employee Problem

For most of the past few years, AI has been framed as a tool — a thing you query, the way you query a search engine or a calculator. That framing is getting outdated faster than most people realize.

AI is moving from a thing you ask questions to a thing you assign tasks to. Writing, coding, research, scheduling, customer service, design, analysis, operations — all of it moving from human-only to hybrid to AI-first. The model is no longer the brush. It is the worker.

That changes who you have to be in the relationship. You are no longer the consumer of a product. You are the supervisor of a workforce, even if the workforce is small, and even if it is invisible.

There is a rule that every nonprofit director, military officer, surgeon, and serious manager I have ever met has had to internalize the hard way. You can delegate work. You cannot delegate responsibility.

If your delegate does the work badly, the work is still your responsibility. If they do it well, the credit may flow to them, but the accountability for the outcome flows to you. That rule does not get suspended because the delegate happens to be a machine. If anything, the responsibility tightens — the machine cannot apologize, cannot quit, cannot be fired in any meaningful sense, and cannot stand in front of the people it harmed.

A supervisor who says "the AI did it" is not absolving themselves. They are advertising the fact that they did not supervise.

The failure modes look almost exactly like the failure modes of bad management of human employees. They show up wherever the supervisor has not done the basic work of being a supervisor.

• Vague instructions — the task was never crisply defined.
• No review — work goes out the door without anyone checking it.
• No boundaries — the agent does not know what it is forbidden from doing.
• No logs — when something breaks, nobody knows what was actually done.
• No escalation — the agent did not know when to stop and ask.
• Overtrust — the supervisor stopped paying attention because early results were good.
• Blame shifting — when it goes wrong, the system gets blamed instead of the person who delegated to it.

Years of leading a nonprofit taught me a small set of practical disciplines that translate almost without modification to supervising AI work.

Define roles before the work starts, not after the first mistake. Inspect what matters, not everything — but actually inspect it. Train the system; do not assume the first instruction will hold forever. Document the decisions, especially the ones that felt obvious at the time. Watch the incentives you are creating, because the system will optimize for whatever you are rewarding, including the unintended things. Correct quickly, before a small drift becomes a culture.

Every one of those disciplines maps onto AI supervision. The translation is almost trivial. The willingness to actually do it is not.

Before assigning real work to an AI, walk through a short, ruthless checklist. These are the questions a serious manager would ask before handing the same task to a junior staff member.

• What exactly is the task, in language a stranger could read?
• What is the risk if it goes wrong, and who is harmed?
• What can the agent decide on its own, and what has to come back to me?
• What facts must be verified before the agent acts on them?
• How will I actually inspect the work, and on what cadence?
• What evidence proves the work was done correctly?
• How will I improve the process when something fails — and I assume something will?

The thing nobody quite says out loud is that the safety problem and the productivity problem are the same problem. The AI deployments that go badly are almost always the ones where nobody supervises. The deployments that go well are almost always the ones where somebody quietly took the supervisor role seriously, even when the language of the tool did not call it that.

The shift from consumer to supervisor is the actual unlock for safe AI adoption. It is not a technical shift. It is a posture shift. And it is the kind of shift that decades of operational leadership have already trained a generation of humans to make. We are not starting from zero on this. We are starting from the entire history of how human organizations have responsibly used labor.