What Everyone Gets Wrong About Reskilling for AI

In early 2023, Anthropic posted a job listing that made the news. Prompt Engineer and Librarian, $175,000 to $335,000 a year. No computer science degree required. Basic programming skills were desirable. The posting asked for a creative hacker spirit, and noted that since the field was arguably less than two years old, the role was hard to hire for.

Time magazine covered it. Bootcamps sprang up. Universities stood up certificate programs. People left steady jobs to retrain for it.

Every one of those calls was sound. The demand was real, the money was real, and for once, a leader could answer the hardest question on the table. Reskill to what? There was finally something to point at.

In 2026, Anthropic posted a similar listing. Prompt Engineer, Agent Prompts and Evals. Five years of software engineering in Python. The work is designing system prompts across products, building evaluation suites, catching regressions before model releases, and mentoring other engineers.

The salary is $320,000 to $405,000. The floor of the new range is nearly the ceiling of the old one.

The prompt engineer didn't disappear. The role climbed.

What got absorbed was the part you could teach as a short course, writing a good instruction and documenting what worked. The capabilities that once had to be assembled by hand became features of the models and the systems around them, until some firms were using AI to generate and optimize the prompts for their other AI. By May 2025, Fast Company reported the standalone role had virtually disappeared, quoting an executive who said it had turned from a job into a task almost overnight.

Nobody in that story was careless. They trained for the part of the work that could be written down, and that's the part the tools take first.

Here is the part that should worry anyone holding a workforce plan. The problem was never that companies picked the wrong skill. It's that the work hadn't stabilized yet, so the only thing available to teach was the visible behavior of a new technology.

If the work keeps climbing, reskilling can't keep chasing the job description.

The largest workforce survey on earth ran into the same wall. The World Economic Forum's 2025 Future of Jobs Report asked more than a thousand employers across fifty-five economies to name the skills their people would need by 2030. Alongside AI and data literacy, the answers came back. Creative thinking. Resilience. Motivation. Self-awareness. Curiosity.

Hold that list. We'll come back to it.

Reskilling is the agreed answer. It's in the board deck, the town hall, and even in the trade press. Upskilling the existing workforce was the most common strategy in that report, with 85% of employers anticipating it over the next five years. In McKinsey's 2025 survey, only 21% of the organizations using generative AI had redesigned any of their workflows around it. That gap gets read as a failure of will. Budgets, competing priorities, somebody dropped the ball. It's something else. You can't build a program toward a destination you can't name.

You've likely made the promise yourself. Nobody is being replaced. We want to free you up for higher-level work. Every leader has made that promise. Few can say what the better work is or how to build it into their strategy.

The Work That Replaced the Work

Start with what has changed. Recalling facts. Processing information. Analyzing data. Those abilities used to set a person apart. Now the tools do them faster than any person can, and having them is no longer an edge.

Here's what the moment demands instead. Deciding where to point the tools. Interpreting what the results mean. Questioning what everyone else has accepted. Asking what's worth pursuing at all.

That's not a forecast. In August 2026, four of McKinsey's newest partners described the change from the inside to Business Insider. The analysis layer, the work juniors were hired to do and partners earned their titles teaching, has gone to the tools.

What fills a partner's day instead is the human work. Orchestrating the meeting and deciding who belongs in it. Spotting which skill each person needs next. At McKinsey, the reskilling question didn't get answered. It got promoted.

Even Holger Hürtgen, the firm's first AI engineer in Germany and twenty years in, said he has never felt so far behind the curve, and that he's "uncomfortable every week." Teaching everyone the tool is what you do when you can't say what the new work is.

Asked what stays constant at the top of the firm, one of the four called it an irreducible core. Another named it. Curiosity.

Now go back to that list from the Forum. Same word. Among the skills rising fastest, the report separates out two it calls socio-emotional attitudes, and curiosity is one of them. A thousand employers were asked to name the skills their people would need by 2030, and the answer they keep reaching for is something the report itself won't call a skill.

Curiosity is the most honest answer the language currently offers, which is why the smartest people in the room keep landing there. It's real, and it's not enough. Curiosity is the impulse to ask. It doesn't build the answer.

Imagination builds the answer.

The tools can generate a hundred options and never know which one matters. Choosing is the human edge — reading what the data means, seeing what's missing, deciding what's worth building. Almost no one calls it by name, because it has always been treated as a gift you either have or you don't. That's what my research changes. Imagination has a structure, and structure can be learned.

Where should your organization focus first?

Follow the claim and it changes what performance means. If imagination is the competence this era demands, it belongs where competence gets defined, in the review, in the promotion, in what doing the job well means. People develop what they are measured against. Leave imagination off the criteria and no amount of training will produce it.

Nothing Is Missing

No organization is short on imagination. It built everything you have. The question is which direction it runs.

My research into how the mind builds futures found three modes, and my Imagination Framework is how they work together. Navigational asks how do I get there. Protective asks what could go wrong. Generative asks what could be, the only one that pictures something that doesn't exist yet.

Here is why the Framework matters now. AI accelerates the first two modes. Planning gets faster, defense more efficient, execution more precise. Generative sits outside what any tool does. A model recombines what it's given and never wonders what's missing. It's also the mode that shuts down first under pressure. The tools amplify the two your organization already overuses and leave untouched the one it has stopped practicing.

Why did Generative shut down? Not for lack of capacity. The other two modes are busy protecting the Default Script, the definition of success that was correct once, has been confirmed by every quarter since, and now goes unquestioned because questioning it feels like questioning the company. It's why the workforce plan came back as a curriculum. Navigational picks the training, Protective covers the risk of falling behind, and nobody in the process ever asks what the work is becoming.

Every Function, Same Question

This is not a call center story. The freed up time is real and it's everywhere. In BCG's 2026 survey, 42% of frontline AI users said it saves them a full day a week. Two-thirds get little or no guidance on what to do with it. The time is there. The decision about what to do with it is not.

The work built most tightly on defined procedure went to the tools first. Every function has some of that work. Engineering, legal, operations, sales. The routine work shifts, capacity opens up, and somebody has to decide what it's for.

This is the hole in most AI strategies. The tools get implemented, the time comes back, and no one decides what to do with it, so the Default Script decides instead, and the old plan absorbs everything the tools give back.

Now What

Here's what everyone gets wrong about reskilling for AI. It's treated as a training problem. Pick the skills, buy the curriculum, run the program. But a program can only teach what's already specific enough to teach, and that's the first thing the tools take.

Stop trying to predict the next job. Build the capability that lets people move when the job moves.

Anthropic's prompt engineers moved up a layer in three years. The more precisely you can name a skill, the shorter its clock.

Your question is no longer which course to buy. It's where your organization's imagination went, and which definition of success it's defending while the freed up time disappears back into the old plan.

Where should your organization focus first?

Drawn from my book, The Imagination Age.