What You Made
Why a business case built around jobs cut can miss new work done with AI.

What we call replacement is a count of jobs that already exist. Those jobs have names, people who hold them, and salaries on a budget. Remove a seat and there is a figure to put in the business case.
New work has no salary to remove and no old process to compare. There is no record of what it used to cost because nobody was doing it. How do you measure work that didn't exist?
I expect a business case built only around headcount saved to favor what it can already measure. It can price the seat removed without asking what the person might have helped build with the tool.
That is how a number can be accurate about what it counts and wrong about the business. The count can answer the question it was built for. The mistake is letting that answer decide which work is worth making.
From reading scans to replacing radiologists
In 2016, Geoffrey Hinton said, "People should stop training radiologists now." He predicted that deep learning would outperform radiologists within five years.
By 2025, he had narrowed that forecast. He told The New York Times he had spoken too broadly: he meant image analysis. He considered himself wrong about the timing, not the direction.
The difference matters. Image analysis is work a radiologist does. Radiologist is the job.
Image analysis is work a radiologist does.

Click a numbered dot to see what it stands for. Start with 1.
The task may change as predicted. That still leaves the rest of the person's work, including what becomes possible when the task takes less time.
At Mayo Clinic, the response to Hinton's forecast was to invest.
The radiologist between the team and the tool
Mayo's radiology chair, Matthew Callstrom, recalled wanting the technology to "make us better." By the Times's 2025 report, the department's staff had grown 55 percent since Hinton's forecast, to more than 400 radiologists.
According to the Times, Mayo had also built a 40-person AI team. It included scientists, radiology researchers, data analysts and software engineers.
That team worked with Dr. Theodora Potretzke on a tool to measure kidney volume. She had previously done much of that measurement by hand. Steve Lohr, the Times reporter, described a process involving an on-screen ruler and estimates.
Potretzke helped design the software and checked its measurements. According to the Times, the tool saves her 15 to 30 minutes each time she examines a kidney image.
She still interprets what the image means. She describes her role as "a doctor for other doctors," communicating results and advising them.
The saved minutes belong to old work done faster. The work of helping design and test that particular tool is different. Potretzke was its user, but she also helped shape what it did.
At Mayo, Potretzke served as a consultant, end user and tester. A replacement count would record her saved minutes. She helped design the software and checked its measurements before use. She still interprets each image afterward. So who supplied the context first, her or the tool?
Work without a lasting title
Paul R. Daugherty and H. James Wilson called the place where humans and machines work together the missing middle in their 2018 book, Human + Machine. Daugherty and Wilson called it "missing" because "almost no one talks about it." They described people helping machines work, and machines extending what people can do.
They described people helping machines work, and machines extending what people can do.

Click a numbered dot to see what it stands for. Start with 1.
In their Harvard Business Review research involving 1,500 companies, they reported the largest performance gains where humans and machines worked together. They also warned that firms using AI mainly to displace staff would see only short-term productivity gains.
A title can make that work briefly visible.
In February 2023, The Washington Post described prompt engineering as a new job. Searches for prompt engineering roles on Indeed peaked that April.
By 2025, Indeed economist Allison Shrivastava told Fortune that prompt engineering remained useful as a skill, but "it's not an entire title."
That does not prove a new class of workers has formed. It means the title and the work are different things to check.
But couldn't demand explain the growth?
Demand may explain why Mayo hired more radiologists. The Times reported that radiologists remained in high demand. Mayo's growth does not prove that AI created those jobs.
The kidney tool also speeds up an existing task. And the fading prompt-engineer title gives good reason to doubt promises of whole new professions.
I would not use either example to promise that AI will create more jobs than it removes. The evidence does not settle that.
It does show why I would hesitate to price a whole job from one task. As labor economist David Autor told the Times, radiologists "do a lot more than reading scans."
A seat count can list the team's salaries, but it cannot price what the team and the person between the team and the tool made possible. My concern is with a decision judged only by seats removed.
My expectation is that firms deciding mainly by headcount saved will find savings and miss some of the work they could have made.
I would be wrong if, after two years, firms that cut to a seat target showed as much new work as firms that kept their experts and built the middle. I would want to see the work, not just the titles.
A count can show what you had. It cannot, by itself, show what you made.
I write by email about where that middle shows up in a week of work.