The Accidental Manager

Ask most business owners what “AI in my business” looks like and you get some version of the same picture: a machine doing the work, and a line item disappearing from the payroll. Software that eats a role. A robot that eats a department.

That picture is wrong, and we already know it is wrong, because this exact thing happened once before and somebody went and measured it.

The receipts from last time

When computers arrived in offices, the expectation was that administrative work would evaporate. It did not. Office and administrative support went from under 12% of US employment in 1950 up to about 17% by 1980. It only drifted back down to roughly 1950s levels by 2019. Nearly seventy years, and the net result was a return to where it started, with a long climb in the middle.

The more interesting part is what happened inside those jobs. Marcus Dillender and Eliza Forsythe looked at roughly 8 million job postings from 2007 to 2016 across 69 technology categories, tracking how education requirements shifted before and after firms adopted new software (NBER Working Paper 29866, “Computerization of White Collar Jobs”). They found no reduction in demand for traditional office tasks after software adoption. What they found instead was that computerization raised the skill requirements of those roles. Software adoption increased the wage premium for college-educated support staff by about 5% relative to otherwise-similar non-college peers.

Read that again, because it is the whole argument in one sentence. The technology did not delete the work. It made the work harder, and it paid more to the people who could handle the harder version.

One note on sourcing. Ari Meisel made the “computerization turned everyone into an accidental secretary, AI will turn everyone into an accidental manager” framing in a video that does not cite a paper. Connecting his claim to Dillender and Forsythe is our inference about the probable evidentiary basis for the historical pattern he is describing. He did not say it. We are saying the numbers line up.

One level up

The shape repeats, but the altitude changes.

Computers pushed everyone down into a task they had not previously done. Executives started typing their own memos. Managers started building their own spreadsheets. The secretarial function did not vanish, it got redistributed across everyone who now had a keyboard, and the specialists who remained were doing a more demanding job than before.

AI pushes in the other direction. You will not do the task. You will manage a system that does it. Everyone gets handed a direct report that does not sleep, does not push back, and does not tell you when it is guessing.

That is a promotion nobody applied for, and the failure mode is obvious once you name it: being excellent at a task teaches you nothing about supervising something else doing that task. Those are two different jobs. Most people have practiced the first one for their entire career and the second one never.

Why the technical people are not the ones who win

Here is the part worth slowing down on, because it runs against the intuition almost everybody has right now.

The prevailing assumption is that this phase rewards technical fluency. Learn the tools, learn the prompting, get ahead. That assumption has a short shelf life. Access to the tools is converging on universal and the interfaces get easier every quarter. Whatever edge exists in knowing how to operate the thing is an edge that decays on a schedule.

The durable edge is somewhere else, and it is unglamorous: knowing where mistakes come from in your specific business.

Think about how an AI system actually fails you. It rarely fails loudly. It does not hand you a blank page or an error message. It hands you something plausible. A summary that is 90% right and quietly drops the one exception that mattered. A draft that follows the stated requirement perfectly while missing the requirement nobody wrote down because everyone here already knows it. An analysis that is internally consistent and built on a handoff assumption that has been wrong at your company for three years.

Catching that requires you to already know what wrong looks like here. Not wrong in general. Here. You need to know that the intake form has a field everyone fills in incorrectly. You need to know that this particular client always means something different by “final.” You need to know that the number in that report has been reconciled by hand every month since 2019 because the source system double counts.

None of that is in the model. None of that is in the prompt. It is in the head of somebody who has been in the room.

So put two operators side by side. One is technically skilled and domain blind. They can run the tooling, chain the steps, wire up the integration, and produce output all day. They will not notice the output is subtly wrong, because nothing in their experience tells them what subtly wrong looks like in this business. The other knows the business cold and is mediocre with the tools. They will look at the same output and say “that cannot be right, we never bill that way.” Then they will go find out why.

The second person is worth more. Prompt engineering is not the same skill and it cannot be bought as one. You cannot hire a contractor to know where your specific business breaks. That knowledge only accrues to people who have been standing there when it broke.

The instinct nobody has priced yet

There is a group of people who have spent entire careers doing exactly what this new job requires, and almost nobody is paying them for it as a distinct skill.

Anyone who has worked in QA, audit, compliance, or operations has spent years asking one question over and over: did we actually verify this worked. Not “does it look right.” Not “did it run.” Did we check. And its companion question: what edge case breaks this, what was not considered, what happens when the input is empty or the date is in the wrong format or the customer does the thing customers are not supposed to do.

That is the AI management skill. That is verbatim what the job is. You sit between the business and the system, you assume the output is wrong until proven otherwise, and you know which specific ways it is likely to be wrong.

Most organizations currently treat that instinct as overhead. It is the thing that slows down the release. It is the department that says no. It has no line on the org chart labeled “the person who notices.” Our view is that this gets repriced, and not gently, because the volume of plausible-looking output about to enter every business makes the ability to spot the flawed one scarce and expensive.

Our forward view on the timeline

The PC-era version of this took decades. The climb from 12% to 17% ran thirty years. The redistribution back took another forty.

We do not think this one takes seventy years. We think it compresses to years, and this is our prediction rather than an established fact. The reason is mechanical: the tooling iterates monthly instead of on a hardware and enterprise-software adoption cycle measured in multi-year increments. A 1975 office bought a machine and lived with it for a decade. A 2026 office gets a materially different capability every few months, without buying anything, often without deciding anything. The pattern is the same. The clock is not.

If that is right, the adjustment window is short enough that “I will figure it out when it gets here” and “it is already here” arrive in roughly the same quarter.

The reframe

The skill is not using AI. It is managing it.

And managing is not a thing most people have ever been taught. Not managing people, not managing systems, not managing anything. Most managers were promoted because they were good at the task, handed a title, and left to work it out. That worked, slowly, because humans push back. A person tells you when the instruction was unclear. A person says “that does not sound right.” A person has a bad day you can see.

The system you are about to be handed does none of that. It will do exactly what you asked, confidently, at scale, and it will not mention that what you asked was wrong.

Everyone is about to become a manager. Almost nobody was trained for the job.