#We’re all becoming managers
With AI, most white-collar workers will change jobs without changing titles. They’ll become managers, of agents rather than people. The title stays the same, but where the day’s work actually happens moves a long way.
Knowledge work in three phases
Take any cognitive task you do on your own, roughly the work that happens behind a screen: data analysis, software development, legal drafting, market research. It breaks down into three phases:
First you define the goal and sketch a plan. Then you produce: write the code, build the model, draft the memo, put the proposal together. Finally you review, fix and deliver.
Most of the work, and most of the time, used to sit in execution. That’s where good people pulled ahead of average ones. It’s also where you built the experience that, ten years on, earned you a senior title and the pay that goes with it.
Whole professions ran on phase 2: lawyers, developers, junior analysts, accounting staff. These were jobs where most of the day went into carefully doing tasks someone else had scoped.
How the work splits now
AI takes over phase 2 and shrinks it. Execution gets ten to a hundred times faster, and the value moves to what surrounds it: what you ask the agent for, and what you accept back.
Phase 1 matters more. When you did the work yourself, you could learn on the go, adjust the goal mid-task and let constraints show up as you hit them. An agent takes that luxury away. You have to explain the context, say what you expect, and spell out the success criteria and the gray areas before you hit run. A precise brief gets you something usable. A vague one gets you work that looks good and isn’t.
Phase 3 grows too. When you did the work, you know what’s in it. When an agent did it, you have to find out, the way a manager reads a deliverable from their team. You read it closely, understand the choices behind it, separate what’s right from what merely sounds right, and give feedback sharp enough to make the next round better. As developers say, reading code is harder than writing it.
Two new phases
Look closer and two more phases appear, one at each end, that didn’t exist in the old model.
Phase 0: decide whether to delegate. Every task starts with a choice: do I hand this to an agent or do it myself? Ethan Mollick lists three criteria: how long it would take me, how likely the agent is to get it right the first time, and how long I’ll spend reviewing and fixing its output. You make this call dozens of times a day, and it’s what separates experienced operators from beginners. Beginners delegate everything or nothing. Experienced operators have calibrated the trade-off, and they recalibrate it with every new model release.
Phase 4: write down what you learn. By default, an agent remembers nothing. Contrary to what many people assume, it doesn’t learn from its mistakes the way a retrained model would: everyone uses the same base model, and every session starts from a blank memory. If you don’t want the agent to repeat today’s blunder tomorrow, you have to write the lesson into its external memory yourself, usually a Markdown file it loads next time. It’s record-keeping, tedious but cumulative, and over time it turns a generic agent into a collaborator built for you.
What follows
So the human becomes a manager: they explain, delegate, supervise and decide when the work is ready to ship. That shift has deep consequences for how work is organized.
Not everyone wants to be a manager. Plenty of people chose their careers with no interest in managing, especially in technical fields, because they liked the intellectual satisfaction of doing the work. That’s why I moved into data science myself. AI takes that refuge away. It absorbs the pleasure of making things and replaces it with the more abstract job of framing and giving feedback. Some professions will take it better: a junior auditor or junior lawyer already treated execution as a required step on the way to management. For people who loved the craft of execution, it’s a sharp turn in their career path.
We’re moving from being judged on effort to being judged on results. When an agent does correctly, in a few minutes, what took a human a day, time spent stops being an argument. Nobody can defend a mediocre deliverable anymore by saying they did their best for eight hours. The shift is quiet for now, but it will change how employee performance is judged.
This work is more mentally intense than the work it replaces. The old workday mixed stretches of execution, sometimes mechanical and even restful, with harder stretches of thinking. In the new one, with execution delegated, almost everything left calls for judgment: understanding, briefing, reviewing, deciding. Four hours of that is more tiring than eight of the old kind. A recent Harvard study calls it AI Brain Fry, and early reports line up: experienced AI operators describe cognitive fatigue that sets in faster and runs deeper, a mental saturation well beyond ordinary tiredness.
For fifty years, skilled office work rested on an unspoken deal: companies paid trained people to carefully execute tasks that others had scoped. AI breaks that deal. It takes execution and makes it instant. It leaves the human everything around execution, which used to fill a small part of the day and now fills most of it. You’ll need to say clearly what you want, judge what comes back and decide what’s worth shipping. In other words, you’ll need to manage. Not everyone signed up for that.