Everyone says AI turns experts into managers. I think it does the opposite.
Four people asked me the same question this month. Different words, same worry underneath.
Should I stop going deep and learn to manage the AI instead?
None of them were juniors. Each had spent years getting good at one thing, and each had been handed the same story. The machine does the work now. Your job is to direct it.
I understand why the story is attractive. It sounds like a promotion. It also lets you off the hook for the slow, unglamorous business of staying good at something.
I think it\'s backwards, and I think it\'s going to hurt the people who believe it.
There\'s a name for the distance between someone who can check whether an agent\'s answer is right and someone who can only check whether it looks finished. Call it the generalist gap.
This week I watched the manager story come from someone who can\'t be waved away. Shishir Mehrotra runs Superhuman, the company behind Grammarly. He ran product at YouTube and sits on Spotify\'s board. He has spent twenty years deciding who gets hired and who gets promoted. Asked what he\'d tell someone starting out today, his answer was \"learn to be a manager\".
What interested me was what he said next. He admitted he\'d spent years telling people the opposite. The best managers he hires are the ones who can still do the job. And when the interviewer pushed him, asking how anyone learns good from bad once the doing is automated, he didn\'t say tools. He said what they\'re hiring for is judgement.
That\'s the generalist gap, described by the person handing out the advice. Hold onto it, because it\'s the whole argument.
Think about what happens when you manage a good team.
A good analyst tells you when the data is thin. A good engineer tells you the deadline is fiction. A good credit officer says \"I don\'t like this one\" before she can fully explain why, and you\'ve learned to listen. People come back with questions. They get stuck and say so. They learn, and next quarter they don\'t make the same mistake.
That\'s why a generalist can manage specialists. The specialists carry the depth. The manager carries the judgement about people, priorities and time.
Agents carry very little of that back to you. Some cite their sources. Some attach a confidence score. None of them will tell you the number is right and the label is wrong, because they don\'t know. They hand over confident, well formatted, plausible work, and if you ask whether it\'s sure, it will tell you it is. Correct it on Monday and, unless someone has built that correction into the system, Tuesday\'s agent makes the same mistake again.
So the person managing agents can\'t be a generalist. The depth has to sit with the reviewer now, because there\'s nowhere else for it to sit.
Let me make this concrete. What follows is a composite, drawn from patterns I\'ve seen across delivery reviews, not from any one bank.
A lender deploys an agent to draft credit memos for small business loans. The pilot goes well. The memos read cleanly, every section is populated, and turnaround falls from three days to twenty minutes. The steering committee is delighted. Someone builds a slide.
Two people are asked to review the output before it reaches the approver.
The first came up through operations. She\'s organised, sharp, and she has never underwritten a loan. She reads the memo against the template. Every heading is there, the numbers reconcile to the financials, the tone is right. She approves it in four minutes.
The second has fifteen years of underwriting behind her. She reads the same memo and stops at the cash flow section. The agent has treated a one-off sale of a delivery van as recurring revenue. The figure is correct, it\'s right there in the accounts, but it\'s the wrong kind of money. Sitting in the wrong line, it makes the business look twice as able to service the debt as it really is.
Nothing in the memo flagged it. No warning, no confidence score, no \"check this\". The agent didn\'t know it had made a mistake. The only thing in the room that did was the credit officer.
Now look at what each reviewer did. The first did everything a manager does. Completeness, format, tone, timing. The second did something a manager can\'t do without the years behind her. She recognised a wrong answer that looked like a right one.
If the first reviewer had been the only reviewer, the memo would have gone to the approver, the facility would have been written, and the error would have shown up eighteen months later as a missed repayment, by which point nobody would have connected it to the pilot.
Shishir has a way of describing careers that I\'ve found useful. Early on, someone hands you the problem, the solution and the method, and you execute. Later you\'re handed the problem and you find the solution. At the top you\'re handed a space and you name the problems. His view is that AI pushes that whole ladder up. Execution goes to the machines and everyone shifts a rung.
His interviewer said the quiet part. When she hires a junior for social media now, drafting the posts isn\'t the job, because a chatbot drafts them. She wants the strategy.
Put those two together and you get something neither of them said out loud. If the bottom rung is gone, the person sitting in the reviewer\'s seat will increasingly be someone who never sat in the doer\'s seat. The generalist gap stops being a career choice a few people make. It gets manufactured, at scale, by the hiring pattern itself.
The credit officer could catch the van because she\'d spent fifteen years writing memos before anyone let her review one. Where does the next credit officer get her fifteen years, if an agent has been drafting since her first week?
I don\'t have a complete answer to that. I\'ll give you the best partial one I\'ve heard, and it also came from Shishir.
You learn judgement the way you learn an instrument or a sport. Not in the recital, and not in the televised game. In the basement, on the driveway, in places where getting it wrong costs nothing and you can go again tomorrow.
For a career in the agent era that means something specific. Do the task by hand, on something that doesn\'t matter, before you let the agent do it. Draft the memo yourself for a made-up business. Write the code for the toy project with the assistant switched off. Then let the agent do the same task and mark it, line by line, against what you\'d have done. Repeat until you can feel the wrong answer before you can explain it.
That isn\'t nostalgia for the old way. It\'s the only route I know to being able to tell a right answer from one that merely looks right.
The chef who runs a serious kitchen doesn\'t cook every plate. She stands at the pass and tastes. She can do that in three seconds because she spent twenty years cooking the dishes herself. A cook who never learned the dish can stand at the same pass all night and taste nothing wrong.
That\'s the shift. You stop producing and you start auditing, and the auditing is only worth anything if you could still produce.
Shishir gave an example from marketing that makes the same point. Twenty years ago an agency made one ad a year. Digital made it millions. The skill of coming up with a great message didn\'t disappear, it got more valuable, and it grew a second half. Now you have to judge hundreds of variations, and you can only judge them if you could have written the good one yourself.
In practice it looks like this.
You know the three places in your own domain where a plausible answer is usually wrong, and you check those first, every time. For the credit officer it\'s the classification of cash. For an engineer it might be error handling on the path nobody tests. For a lawyer it\'s the clause that reads fine and means the opposite.
You write down what wrong looks like, so the next reviewer can catch it without your fifteen years. That written check becomes part of the process rather than a thing that lives in one person\'s head.
You treat the agent like a bright graduate who will never learn from feedback. Which means you stop giving it feedback and start building the check around it. That\'s what I mean when I say agent coaching, not agent training. The coaching isn\'t aimed at the agent but at the system the agent sits in.
None of that is less expert than what you did before. It\'s the same expertise, pointed at a different job.
Take your five most important agents. For each one, write down the name of the person who signs off its output.
Then ask one question about each name. Could this person have done the task themselves, before the agent existed?
If the answer is no, that agent doesn\'t have a control. It has a rubber stamp, and everyone above it is trusting a signature from someone who can\'t see what they\'re signing.
If you\'ve spent ten years going deep on something, don\'t trade it for a title with \"orchestrator\" in it. Your depth is the control. It\'s the only part of the system that knows when the output is wrong.
Shishir was right about the thing he said second. They\'re hiring judgement. He was wrong about the thing he said first, because judgement doesn\'t come from managing. It comes from having done the work.
The person who should be nervous is the generalist at the pass, waving through plates she has never cooked.
One more thing before I sign off. The generalist gap doesn\'t stop at the office door. The same problem sits at the kitchen table, and that\'s where the last three months went, into a book. Not how to use AI. How to check it before you trust it. The chatbot your teenager asks about homework, the one your mother asks when something official arrives in the post, the one you ask at eleven at night when you can\'t be bothered checking. Smaller stakes each time, far more times. The book teaches one habit, a ten-second check before you trust what a machine tells you, and it\'s written for anyone, whether they\'ve read a line of code or not. My wife tested it first, then ten more readers for a week each. It\'s nearly done. I\'ll say more when it\'s ready to buy, and this newsletter goes back to its own job.
Reply with one line. Where in your own domain does the plausible answer usually hide the wrong one? I read every reply myself. vijayan@aideliverydiscipline.com
New issues arrive by email first. Subscribe here.
