“Pull the Q3 numbers, flag anything that looks off, draft the summary, and check with me before it goes out.” You have said some version of that to a new analyst. It carries an outcome, a scope, a bar for what counts as done, and a point where you want to be asked. It is a complete instruction, and there is not a line of JSON in it.
Hand the same sentence to an AI agent and it works. Not because you learned to prompt. Because you already knew how to delegate.
The industry has spent two years selling that competence back to you as something foreign. “Prompt engineering,” with courses and job titles attached. I think it is a category error. Directing a capable agent is not a new technical literacy you have to acquire. It is the management skill you built handing work to people, relabeled to sound like it needs a certificate.
Here is the part a reasonable person can argue with. I do not think prompt engineering is worth learning as its own skill. Brief the model the way you would brief a sharp new hire and you have most of it. The residue of genuine technique, context windows and the model’s specific quirks, is real but small, and it shrinks with every release.
The behavior shows up before the theory
Put capable people in front of an AI agent and they do not reach for a config panel. They start managing. In a randomized field experiment, people delegated 17% more work to an AI agent than to a human partner and made 62% fewer direct edits. Nobody sat them down and taught them to prompt. They directed, because directing is what they already knew how to do.
The research frontier is drifting the same way. The newest work on handing tasks to AI describes it as a principal and agent relationship: transfer of authority, accountability, a clear specification of the role and its boundaries, a shared sense of intent. That is a manager’s job description. It is moving away from the older assumption that a sub-agent performs because the prompt was well built, toward something that reads like the determinants of good human delegation named decades ago. Give the work to someone competent, who shares your objective, with whom you have a working relationship you can check in through. The levers do not change when the doer stops being a person.
None of this makes the two things identical, and I will not pretend it does. Delegation pays off only where the delegate is competent, and AI is competent in a jagged, uneven way. In that same experiment, handing work to the agent lifted the quality of the written output and did nothing measurable for the images. That is the jagged frontier, and working around it is itself a management skill: knowing what your associate is good at, and reading twice the thing it is weak at. You do that with people already. You know whose numbers you trust at a glance and whose you check line by line.
Why the label matters
The mislabel is not harmless. Call directing a technical skill and you turn a director back into an operator, squinting at syntax, convinced the distance between them and a good result is a course they have not taken yet. It usually is not. The distance is a vague brief, the same thing that produces a bad handoff to a human. Someone who can say what they want and name the moment they want to be pulled back in is already most of the way there. If you can write a decent brief for a person, you can write one for an agent, and you have been writing them for years.
That reframe changes what you go looking for when the output disappoints. The operator’s instinct is to reach for a better prompt template, a longer system message, a trick someone posted. The director’s instinct is to look at the handoff. Was the outcome clear? Did I name the checkpoint, or hand over something irreversible and hope? Nine times out of ten the fix is in the brief, not in some incantation you have yet to master.
So the honest answer to “how do I get good at delegating to AI” is unsatisfying if you were hoping to buy a technique. You get good at it the way you got good at delegating to people. Be clear about the outcome and honest about what the doer can actually handle, and be present for the one checkpoint that matters. The analyst got “check with me before it goes out.” Give the agent the same line. The skill was never missing from you. It was just pointed at people.
Sources. Delegation and edit-rate figures: Ju & Aral, Collaborating with AI Agents (arXiv, 2025, preprint). Principal-agent framing: Tomašev et al., Intelligent AI Delegation (arXiv, 2026). Determinants of human delegation: Yukl & Fu, Determinants of delegation and consultation by managers (Journal of Organizational Behavior, 1999).
Brief is opening to a small group at a time. Direct a team instead of operating one more tool.
Request access