#28 Everyone Is a Manager Now — The Job Nobody Applied For

I run another small team now. None of its members sleep, and I never hired them. One publishes each edition of this newsletter on my website. One spends Thursday evenings on strategy and leaves me a list of decisions. Others draft and sort. I wrote their job descriptions, I decide which of them get more resources and more intelligence than the others, because the budget is finite, and I check their work. Last week I had about twenty open conversations with these AI colleagues and was not sure what was still unfinished. Yesterday evening I dictated the memo for this edition. This morning I had forgotten that one of them was waiting for my answers.

I have led people since 2006, first at IBM, later with my own employees, today in every workshop room and with coachees over months. And still this is new, and more tiring than I expected: colleagues who can do so much, so much of it in parallel, and who wait, one after another, for my next input. If it is new for someone who has led for twenty years, think of the specialist who never led anyone.

If you run a company, this is your problem first. In Microsoft's 2025 survey of 31,000 knowledge workers in 31 markets, 36 percent of leaders expected managing AI agents to be part of their own job within five years; among employees, 21 percent. The UK's Chartered Management Institute found in 2023 that 82 percent of managers had no formal training when they started: accidental managers. Whoever unlocks agents, budgets and models in your company is handing out management responsibility without calling it that. Who gets it, and how they are prepared, is decided at the top, or by nobody.

(This is the first of two editions on what working with AI agents asks of people. This one is about the new role. The next is about what it does to the brain.)

8 min read. If you take one thing from this: giving a specialist an AI agent makes them a manager. Most have never been taught to delegate, review or budget, and nobody told them the job changed. Ahead: why people delegate badly to AI; reviewing as the new work, and how it wears off; budget and model choice as a management decision; who trains the accidental managers, plus one task for whoever signs.

Delegation, the skill nobody taught

In the delegation edition earlier this year I wrote that delegation had grown a second question: "should a human do this at all?" People answer it badly. Andreas Fügener and colleagues at the universities of Cologne and Minnesota tested how people split image tasks with an AI that beat them on some images and not on others. The humans delegated poorly: no group that handed tasks to the AI beat the AI alone, though an ideal split would have scored 90 percent against the AI's 77. The authors put it down to missing self-knowledge: people "did not know what they did not know", as the university's summary of the study puts it.

That is not a technology problem; it is the oldest delegation problem there is. Jeffrey Pfeffer at Stanford and colleagues showed in the 1990s that people rate the same work higher the more closely it was supervised, and higher still when they were involved themselves. With agents it becomes micromanagement of a machine: redoing its work, or never letting it do anything that matters.

The oldest cure I know is a 1974 Harvard Business Review article, one of its two best-selling reprints, later a book with Ken Blanchard: "Management Time: Who's Got the Monkey?" by William Oncken and Donald Wass. The monkey is the next move. A subordinate walks in with a problem, and by the end of the meeting the monkey sits on the manager's back. Their rule: whoever brought the problem in carries it out again. The manager may advise, but the next move stays with the person who owns the task. Agents have brought the monkey back. The output arrives at eighty percent, and the quickest thing is finishing the rest yourself. Every time I do that, the agent learns nothing and I have done its job. So I keep the monkey where it belongs: the draft goes back with what was wrong, and when the work will come again, the correction goes into the agent's instructions. That is what makes an agent more senior; a chat alone forgets by tomorrow.

A Google DeepMind paper from February gives the two questions that help me most. Before handing a task to an agent, ask: can I check the result cheaply, and can I reverse it? The easier a result is to check, the authors write, the less trust the delegation needs. Drafting an email is reversible; sending it is not. My own agents each have a written list of what they may and may not do, and every list ends the same way: if in doubt, skip the step and report. That sentence has saved me more than any model upgrade.

Reviewing is the new work

When the agent does the task, the human's job moves. Researchers at Microsoft and Carnegie Mellon surveyed 319 knowledge workers in 2025 and found effort shifting "from task execution to task stewardship": checking, integrating, taking responsibility. And more confidence in the AI went with less critical thinking.

The checking also wears off. A 2026 study, not yet peer-reviewed, followed 400 software reviewers through 11,000 reviews of AI-written code over seven months. As they gained experience, their approval rate rose from 30 to 37 percent, their requests for changes halved, and they wrote 28 percent fewer words per review, while the code did not shrink. The authors read it as habit under growing workload rather than earned trust alone.

The cost shows up in people. Boston Consulting Group surveyed 1,488 US employees this year for Harvard Business Review and named the condition: brain fry, mental fatigue from using or overseeing AI tools beyond your capacity. Fourteen percent reported it. The main driver was not using AI but supervising it. Julie Bedard, who led the study, said they saw it "happening to people who were perceived as really high performers". The managers of agents are your best people, and they are the ones getting fried. What that does to the brain, and what to do about it, is our next newsletter edition.

So review has to be designed: in batches, not as each result arrives, against one written standard, with the right to say no. In Microsoft's 2026 survey of 20,000 AI users, the 16 percent who already run agents for multi-step work were far more likely to say their manager set quality standards for AI work: 83 against 57 percent. A standard is a management act. Someone has to write it.

If you sign budgets and your specialists have been given agents but nobody told them they are managers now, one thing is enough. Forward this edition to whoever runs people development for you, with one line: "I want us to work on this over months, not in an event. Find me something like it." What to look for: a programme that starts with the leadership team, runs over months rather than days, and treats human skills and AI as one subject. If you run people development and just received that line, or still have budget for this year: that is the work I do, since 2010 with organisations from start-ups to adidas and Fraunhofer (references). Twenty minutes, and you'll leave with a clear next step whether or not you work with me: book a call.

The budget is a management decision now

A decision I never faced with a human team: which of my agents deserve the best model. The strong one costs more, and the monthly budget runs out. So the strategy agent gets it and the sorting agent does not. That is resource planning, a project manager's job, and it has arrived at every desk with an agent. Microsoft now bills agent work in Copilot Credits, with spending limits per user set by administrators. In practice, usage is decided by whoever types the prompt.

And a task can be a bad fit for any model. In the best-known experiment, Fabrizio Dell'Acqua of Harvard Business School and colleagues gave 758 consultants AI help: on tasks within the tool's reach they finished 12 percent more tasks, 25 percent faster; on a task just outside it, they were 19 percentage points less likely to get the answer right. Knowing where the edge runs is the new budgeting skill; nobody budgets for being confidently wrong.

Who trains the accidental managers?

In most companies nobody, yet. Gallup finds that only 44 percent of managers worldwide have ever received management training. Now add a second class of manager without the title. In McKinsey's survey of 3,600 US employees, 48 percent said formal training would do most to raise their AI use; a fifth received little or none. Its conclusion was blunt: the barrier is not employees, who are ready, but leaders not steering fast enough.

What the training must contain is not the tool. The Microsoft number that matters most here: the agent-running 16 percent were far more likely to pause and decide what AI should do and what a human should, 53 against 33 percent, and more likely to work without AI to keep their own skills, 43 against 30. Those are management habits: deciding, delegating, keeping a hand in.

So teach delegation: brief, verify, reverse, stop. Teach review as a craft with a standard and a limit. Teach budgeting, because the person with the prompt is spending the money. And teach time: planning tasks and running several in parallel, which a team of agents demands from someone who used to work one task at a time. In the delegation edition I put the principle in one line that still holds: "When you 'delegate' to AI, you hand over the task and nothing else. The responsibility stays entirely with you." The accidental managers already carry it. The least a company can do is tell them.

Your turn

If you sign budgets. Ask for a list of who runs agents or automated workflows in your company. That list is your new management layer, mostly untrained. Then send this edition to whoever runs people development, with the line from the middle of this page.

If you lead a team. Write one standard for AI work in your team, one page, and review agent output in batches at fixed times, not as it arrives. When an agent delivers eighty percent, send it back with the correction instead of finishing it yourself.

If you develop people. Before the next tool training, run a delegation training for specialists: brief, verify, reverse, stop, budget. If you need the mandate, forward this upward; the opening is written for the person who signs.

Your 1% this week, whoever you are. Count your open AI conversations and agents. For each, one line: done, waiting for me, waiting for it.

Next time: what this job does to the brain, twenty conversations, endless text, agents that never wait. Until then: who in your company manages agents already, and did anyone tell them? Tell me in the comments. I read every reply.

If you want to bring these themes into your organisation, here's that link again: book a 20-minute call. You can also explore the Leadership Essentials Workshop Series for teams.