Delegating to AI is a management skill most managers never had

A head of procurement asks an agent to "pull together a supplier comparison for the packaging tender." Ten minutes later she has a tidy document comparing six suppliers on price, lead time and certifications. It's useless. What she needed was a comparison against the two suppliers the company already uses, weighted towards whoever could survive a port strike, in a format the finance director would actually read. She knew all of that. She just didn't say it.

Her first reaction is that the tool isn't very good. A more honest reaction would be that this is the same brief she has given her team for years, and for years her team has quietly filled in the gaps by walking over to ask or remembering what she wanted last time. The agent did neither. It did what she asked.

Most delegation is rescued by the person receiving it

Delegation is one of those management skills everyone assumes they have. In practice, a great deal of delegation inside companies works only because the person on the receiving end is good at decoding vague instructions. A capable analyst hears "can you look at churn" and knows it means last quarter, enterprise accounts, one slide for Thursday's board pack, because she has sat in enough meetings with her manager to know.

That decoding is real work, and it's mostly invisible. It shows up as the junior who "just gets it" and the manager who therefore never has to get better at explaining. Over a career, a manager can rise quite high on the strength of other people's ability to guess.

Agents don't guess well about things they were never told. They have no corridor to walk down and no memory of last quarter's argument unless someone puts it in front of them. Give one a vague instruction and you get a confident, well-formatted answer to a slightly different question. In that sense the machine is an audit of how clearly the manager actually thinks, and the results are often unflattering.

What a good brief contains

Military planners use an idea called commander's intent: alongside the order, you state what the order is for, so that people can make sensible decisions when the situation changes. It is one of the most portable ideas in management, and it applies to agents almost unchanged.

A brief that works, for a person or a machine, usually answers five questions:

• What is this for, and who will read or use it?
• What does finished look like in length, format and level of detail?
• What do you already know that changes the answer?
• What are the constraints, and what should be avoided?
• How will you judge whether it's good?

Run the procurement head's request through those questions and it changes shape. The comparison is for the finance director, who wants a single page and a recommendation. Finished means the two current suppliers plus no more than three challengers. What she already knows is that last winter's port disruption left one incumbent unable to deliver for a month, so resilience matters more than unit price. The constraint is that anything outside the approved-vendor list needs a reason. Good means the finance director can decide in one meeting without asking for a second version. That brief takes five minutes to write. It saves the ten minutes of the useless first draft, and the hour she would otherwise have spent explaining why it was useless.

None of this is new. Good editors, good law-firm partners and good engineering leads have always briefed this way. What's new is that the penalty for skipping it is immediate and visible. A junior would have rescued a bad brief with three follow-up questions. An agent hands back exactly what the brief implied, and the distance between that and what you wanted is a precise measurement of everything you left out.

Why managers find this so uncomfortable

Writing a proper brief forces decisions you might have preferred to postpone. "Draft the reorganisation announcement" is easy to say. Writing down which teams are affected, what tone to strike, what can and can't be promised, and who has already been told requires actually knowing those things. Many managers discover, halfway through the brief, that they haven't decided yet.

Vague delegation has always been a way of handing an undecided question to the person doing the work, then reacting to whatever comes back. "That's not quite it" is a perfectly serviceable management technique when someone else bears the cost of each round. With an agent the rounds are cheap, but they're yours. You sit there rephrasing, collecting slightly different wrong answers, until you realise the problem was never the wording.

Economists talk about the principal–agent problem: the person doing the work has different interests from the person who asked for it. With AI, the more common problem runs the other way. The agent has no interests to speak of. The principal hasn't worked out what they want.

The managers who are good at this were already good

Watch which managers get the most out of agents early on and a pattern appears. The technical ones do fine, but the standouts are the people who were already good at briefing humans. The sales director who sends a two-paragraph note before every pitch explaining the client's internal politics. The HR lead whose interview guides tell each panellist what a question is meant to reveal. The operations manager who writes down why a process exists, as well as what it is.

These people already think in terms of purpose, audience and a definition of done. Swapping a junior for an agent changes the reader but not the habit. They get useful output on the second attempt, sometimes the first, and spend their time on judgement instead of rephrasing.

The managers who struggle are often the ones whose teams were best at mind-reading. They were never forced to be clear, because someone was always there to compensate. The agent is the first subordinate they've had who can't.

Teach briefing like any other skill

Briefing can be taught, and it should be, before anyone buys more licences. Have managers write briefs for real tasks and compare the output with what they actually needed; the gap is the lesson. Keep a shared library of briefs that worked, with a note on why. Teams that run hands-on workshops before rolling out agents often find the most useful session has little to do with the tool. It is the hour spent getting a room of managers to write down, precisely, what a good result looks like for a task they hand off every week.

There's a side benefit that has nothing to do with AI. Clear briefs are kinder to people. The junior who used to spend half a day decoding what her manager meant gets that half day back. The analyst stops building three versions of a report to cover every possible interpretation. A manager who learns to brief an agent properly becomes, almost by accident, a better manager of humans.

So before you decide an agent isn't good enough, try a small experiment. Take the last task it got wrong and hand the same instruction, word for word, to someone new to the team who has never worked with you. If they would have come back with the same wrong answer, the tool did its job. The brief didn't.