Every organization I speak to now has the same AI as its competitors. The same models, the same subscriptions, the same features arriving on the same morning. For a year or two, access felt like an advantage. It is not one any more. When everyone can produce a competent first draft in thirty seconds, the first draft stops being worth anything, and the question leaders put to me has changed shape. It used to be "how do we get people using AI?" Now it is "what should our people actually be good at?"
Here is my answer. When everyone has access to AI, the skills that matter are the ones AI makes scarcer, not the ones it makes cheaper. There are three of them: deciding what is worth asking, judging whether the answer is good enough, and owning what happens next. Almost everything else is becoming more abundant by the month, and abundant things do not command a premium. That is not a prediction about the future of work. It is a description of the labor market your organization is already hiring in.
The Scarcity Rule: How to Tell Which Skills Matter
The Scarcity Rule is the test I use with leadership teams when they ask which skills to build: when everyone has access to the same AI, a skill matters in proportion to how much scarcer it becomes as AI improves, not how useful it is today.
It sounds obvious written down. In practice it cuts against almost every skills plan I am shown, because those plans are built on usefulness. Writing is useful, so we train writing. Analysis is useful, so we hire analysts. Summarizing, translating, drafting, first-pass research, basic coding, recall of facts and procedures: all useful, all still needed, and all now produced at near zero effort by a tool every competitor also owns. A skill can be useful and worthless at the same time. Arithmetic is useful. Nobody has paid a premium for it since the spreadsheet.
The Scarcity Rule asks a different question of every line in the plan: is this skill becoming rarer or more common as the tools improve? Some skills become rarer precisely because the tools improve. Every piece of AI output still needs someone to have asked for it well, someone to have checked it, and someone to stand behind it when it is used. As the volume of output grows, the demand for those three things grows with it, while the supply of people who can do them does not. That is scarcity, and scarcity is where the value goes.
Anthropic's Economic Index report from June 2026, which paired usage data with a survey of around 9,700 people, put it in one line: more production from the AI does not mean less from the user. The human's share of the work does not shrink. It moves. The skills that matter are the ones sitting where it moves to.
Why This Question Matters More Now Than It Did a Year Ago
This question matters more now because access has stopped differentiating anyone, and the second-order effects of universal access are showing up in the data. The World Economic Forum's Davos briefing of January 19, 2026 reported that businesses are struggling to recruit because workers are not acquiring AI skills at the required pace, while naming creativity, innovation and adaptability as the capabilities hardest to automate. Read those two findings together and you get the whole problem. Employers are advertising for the abundant skills, the tool-shaped ones, while the scarce skills go unnamed in the job description because nobody has worked out how to write them down.
There is a second effect that leaders underestimate. Anthropic's Learning Curves report, published on March 24, 2026, found that people who had used its models for six months or more had a higher success rate in their conversations than newer users: about five percentage points in the raw comparison, narrowing to roughly three once the same tasks are compared with each other, and about four with fuller controls. Those are modest numbers, and the reason they matter is not their size but their persistence. The advantage survives controlling for the task, the model and the country, which means it is the person who improved rather than the work getting easier. Experience with AI compounds. The gap between organizations is no longer about who has the tools. It is about who has been practicing the scarce skills with them, and for how long.
Skill One: Deciding What Is Worth Asking
The first scarce skill is framing: knowing which problem is worth putting to the machine, in what form, with what context, and which problems should not go near it at all.
When the effort of producing an answer falls to nothing, the value moves to the question. Anyone can get a plan, a policy, a lesson sequence or a market analysis in a minute. Very few people can say, before they open the tool, why this piece of work is worth doing, what a good result would change, what the tool needs to know about our situation that it cannot know by itself, and what we will do with the answer once we have it. That is framing, and it is a thinking skill dressed up as a technical one. It depends on knowing the organization's priorities, knowing the subject, and having the confidence to decide that some tasks are not worth automating because the struggle is the point.
The weak version of this skill is prompt tricks: templates, magic phrases, a laminated card of formulas. Those were useful for a season and are now built into the tools. The strong version is a person who can turn a vague concern from a governors' meeting into three precise questions, decide which of them the AI can help with, and recognize that the third one needs a conversation with a human being instead. Ask of anyone in your organization: could they explain why a task is worth doing before they open the tool? If not, they are not framing. They are typing.
Skill Two: Judging Whether the Answer Is Good Enough
The second scarce skill is judgment: the ability to tell a good answer from a plausible one, which depends on knowing the subject well enough that you would not have needed the machine to produce it.
This is the skill leaders most often assume they already have in the building, and the evidence says otherwise. A study by Hao-Ping Lee and colleagues at Microsoft Research and Carnegie Mellon University, presented at the CHI conference in April 2025, surveyed 319 knowledge workers about 936 real examples of generative AI use. Its central finding should be pinned to the wall of every training room: higher confidence in the AI was associated with less critical thinking, while higher confidence in one's own ability was associated with more. The people who trust the tool most check it least. The same study found that critical thinking has not disappeared in AI-assisted work. It has shifted, toward verifying information, integrating responses and stewarding the task. Those are exactly the activities most skills programs never teach.
The implication for leaders is uncomfortable. If you want people to judge AI output well, you have to build their confidence in the subject, not their confidence in the tool. Judgment is built by doing the work, and the work is now the thing being automated. This is the trap I see organizations walking into with their eyes open: they remove the junior, repetitive tasks from a role because the AI can do them, then wonder a few years later where the experienced people who could check the AI are supposed to come from. Since this article was first published, the argument has been sharpened from an unexpected direction: the people most at risk of losing this skill are the senior ones. Writing in Harvard Business Review in August 2026, Leonid Sudakov and Nathan Furr argue that leaders are quietly surrendering independent judgment to AI and to the consensus it manufactures, and they point to a Harvard field experiment in which 228 experienced evaluators assessed submissions to an MIT innovation challenge with and without model assistance. Their prescription is worth borrowing whole: structured curiosity, and deliberate dissent. If nobody in the room is required to disagree, the room will agree with the machine.
Weak judgment looks like acceptance. Strong judgment looks like a person who reads an AI-generated risk assessment and says "that is fluent, and it is wrong about us in two places, and here is how I know." Ask of anyone in your organization: would this person notice if the answer were wrong? If the answer is no, the tool is not augmenting them. It is replacing their thinking and keeping their signature.
Skill Three: Owning What Happens Next
The third scarce skill is ownership: taking responsibility for a decision, carrying the people affected by it, and being the person others come to when it goes wrong.
AI can produce a recommendation. It cannot be accountable for one. It cannot sit across a table from a parent, an employee or a board and be the reason they trust the decision. The Readiness Test I use with leadership teams ends with the question "Who decides when it goes wrong?", and I have yet to meet an organization that can answer it with the name of a piece of software. As more of the producing is done by machines, what remains for people is increasingly relational and ethical: the judgment about whether a decision is fair, the conversation that lands it, the willingness to put your name on it. Machines can compute. They cannot care, and they cannot be held to account.
Here the evidence is more hopeful than the headlines. In that same June 2026 Anthropic survey, 57 percent of respondents said AI had made their skills more valuable and 68 percent said they were learning more with it. The detail that matters for this argument is what happens to those two figures as people hand over more of the work: the share reporting that AI is raising the market value of their skills rises with the share of work automated, while the share reporting that they learn more stays roughly flat. Delegating more does not, on this evidence, hollow people out. My reading of that is simple: people who hand the doing to the machine move up the stack, but only where a stack exists to move up into. Building that stack, so that the time AI frees is spent on ownership rather than on producing more of the abundant thing, is a leadership job that no tool will do for you. Ask of any piece of work in your organization: when it is finished, whose name is on it? If the answer is "the AI's", nobody owns it, and nobody will learn from it either.
What I Tell Leadership Teams
What I tell leadership teams is that the skills question is not a training question, it is a design question, and most of them have been asking the wrong department.
The question I am asked most often after a keynote, in whichever country I happen to be speaking, is some version of "so what should we train people in?" The answer disappoints the person asking, because it is not a tool. It is that you cannot train framing, judgment and ownership in a room. You can only design an organization in which people are required to use them, and then support them while they do. Herminia Ibarra and Michael Jacobides made a related point in Harvard Business Review in October 2025: success with AI hinges less on the technology itself than on leadership and organizational transformation, and one of the five skills they name for leaders is simply modeling personal experimentation. I would go one step further. Leaders who use AI visibly, and who are seen being wrong with it and correcting it, teach the second skill more effectively than any course.
Across the leadership teams I work with, the strongest approaches have three things in common. They keep humans doing enough of the underlying work to stay competent at judging it, even when the machine could do it faster. They attach a name to every AI-assisted decision that matters. And they have decided, in writing, which tasks they will not automate because the struggle is where the capability comes from. The weakest approaches have one thing in common: they measure adoption, and call it progress.
The Mistake That Does the Most Damage
The mistake that does the most damage is commissioning AI skills training that teaches the abundant skills and calling it future-proofing.
The pattern is familiar enough that I can describe it before the leadership team does. A day on prompting. A tour of the approved tools. A certificate. Six months later, usage is up, the dashboard looks healthy, and nothing about the quality of decisions has changed, because nothing in the program asked anyone to be wrong in front of a colleague, defend a judgment, or explain why a task should not have been automated at all. The organization has become fluent in the skills its competitors also have, at exactly the moment those skills stopped being scarce.
The damage is not the wasted day. It is the false confidence. The Microsoft Research finding tells you what happens next: people who have been taught to trust the tool check it less. A program built on the abundant skills does not just fail to build judgment. It can erode it.
What Good Looks Like
An organization that has answered this question well is recognizable within a few minutes of walking in, because people talk about what they decided, not about what the tool produced.
The observable behaviors are consistent. Job descriptions and person specifications name framing, judgment and ownership in plain language, alongside the tools. Induction for new staff includes a deliberate period of doing the work unassisted, so that they can later judge the assisted version. Meetings review AI-assisted work by asking who checked it and what they changed, not whether AI was used. Leaders show their own AI use, including the corrections. Professional learning is built around real decisions with real consequences rather than tool features. Somewhere, written down, is a list of tasks the organization has chosen not to automate, and everyone can tell you why. And the measure of progress is the quality of decisions, not the number of prompts.
For school leaders, the same rule applies to the curriculum. The skills becoming scarce for adults are the ones children need most, and an assessment system that rewards the abundant skills is training pupils for a labor market that has already closed. That question splits in two, and both halves now have an answer here: what schools should be preparing children for now, and what happens to assessment when AI can produce excellent work. The leadership question underneath them is identical to this one: are we building the skills that get scarcer, or the ones that get cheaper?
What Leaders Should Do Next
Start by running the Scarcity Rule over three documents: your current training plan, your job descriptions, and the brief for your next three hires.
For each line, ask whether the skill is becoming rarer or more common as the tools improve, and notice how much of your investment is going to the second category. Then make three decisions. Decide which tasks in each role are the ones that build judgment, and protect them from automation for the people who still need to learn. Decide who owns every AI-assisted decision that matters, and make sure that person knows. And decide how you will measure this, because if the only number you have is adoption, adoption is what you will get. If your organization already has an AI strategy that people actually use, the skills decisions belong inside it, under people and proof. If it does not, this is a good reason to write one.
None of this requires a new tool. It requires a leadership team willing to say that access is no longer the advantage, and to act on what is.
Where This Goes Next
If your leadership team is trying to work out which skills to build before the next wave of tools arrives, this is the ground I cover in AI strategy work with schools and organizations, where the skills decisions get made alongside everything else. If you would rather start by reading, the thinking behind this article goes out first to the newsletter.
Sources and further reading
- Anthropic, Economic Index report: Cadences, June 2026. Survey of around 9,700 respondents linked to usage data from mid-May to early June 2026.
- Anthropic, Economic Index report: Learning curves, March 24, 2026.
- Hao-Ping Lee and colleagues, Microsoft Research and Carnegie Mellon University, The Impact of Generative AI on Critical Thinking, CHI 2025, April 2025. Survey of 319 knowledge workers reporting 936 examples of generative AI use.
- World Economic Forum, Davos: What to know about jobs and skills transformation, January 19, 2026.
- Herminia Ibarra and Michael G. Jacobides, 5 Critical Skills Leaders Need in the Age of AI, Harvard Business Review, October 2025.
- Leonid Sudakov and Nathan Furr, AI Is Undermining Leaders' Judgment. Here's What to Do About It., Harvard Business Review, August 19, 2026.
Dan Fitzpatrick is the founder of The AI Educator, an international keynote speaker and a bestselling author on AI in education, and works with leaders and organizations to understand, plan for and lead the changes caused by AI. More about Dan.
