Somebody on your team has already asked you the question, or is working up to it. It comes in different forms. What should I be learning right now? Is what I’m good at still going to matter? Should I be worried?
It is a hard question to answer well, and most of the available answers are unsatisfying in the same way. One version says everything is about to be automated, so the honest response is to prepare people for a different career. The other says nothing fundamental has changed, so keep going. Neither of these is true, and your people can tell. They are asking a real question and they can hear when they are being managed rather than answered.
There is a better answer available. It requires giving up something first, which is the hope that you can name the specific skills that will matter in five years. Nobody can do that right now, and the leaders claiming otherwise are guessing with more confidence than the situation supports. What you can do instead is give your people a rule for telling the difference. That turns out to be more durable than a list, and it is something they can apply themselves, long after any particular list has gone stale.
First, a warning about your instruments
Before the rule, there is something you need to know about how you will measure any of this, because it is genuinely counterintuitive and it will affect every decision you make here.
In 2025, a research group called METR ran a careful experiment with experienced software developers, all of them working on real tasks in systems they knew well. Some of the work was done with AI assistance and some without, assigned at random. The developers were asked beforehand how much faster they expected the AI tools to make them, and asked afterward how much faster they thought the tools had actually made them.
Going in, they expected to be substantially faster. Coming out, having just finished the work, they reported that they had in fact been substantially faster.
They were slower. Not slightly. Measurably slower on the assisted tasks than on the unassisted ones.
The exact percentages matter less than the shape of the result, so hold onto the shape. These were skilled professionals, working in their own domain, on material they knew intimately. They had just completed the work. And when asked to report on their own experience, they got the direction wrong. Not the magnitude. The direction.
This is not a story about people being foolish, and it should not be told that way to your team. It is a story about what the tools feel like to use. The assistance is genuinely pleasant. The waiting is over faster, the blank page fills in, the tedious part goes away. All of that is real, and all of it registers as speed, because in ordinary life effort and time run together closely enough that we use one to estimate the other. AI separates them. The work feels lighter while taking longer, and there is no internal signal that tells you this is happening.
The consequence for you is direct. Asking your team how the AI rollout is going will not tell you how the AI rollout is going. Neither will asking them what is working, or whether it is helping. Those questions are not merely imprecise here, they are pointed at an instrument that has been demonstrated to read backwards. If you want to know what these tools are doing to your organization, you will have to measure something other than how it feels, and you will have to do it deliberately, because nothing about the experience will prompt you to.
That is worth knowing on its own. It is also the reason the skills question cannot be settled by asking around.
The rule
Here is the rule, and it is short enough to repeat.
Anything your tools can check for themselves, they will eventually do. Invest in what is left.
That is the whole thing. It works because it does not depend on predicting which capabilities arrive next, which is the part nobody can do. It depends only on a property of the work itself, which is whether there is a way to tell that it is correct without a human being making a judgment about it.
Work that can be verified mechanically gets absorbed. Not immediately, and not all at once, but reliably and in one direction. If a machine can tell whether the output is right, a machine will eventually produce the output. Work that requires someone to decide what correct even means does not get absorbed on the same schedule, because the deciding is the work.
This is not a new observation, and it is worth knowing how old it is, because that is what should give you confidence in it. The philosopher Michael Polanyi wrote in 1966 that we know more than we can tell, meaning that a great deal of real expertise is tacit. It lives in judgment rather than in any procedure you could write down and hand to someone else. The economist David Autor later gave this idea a name, Polanyi’s Paradox, and worked out what it predicts about employment. The tasks that automate cleanly are the ones that follow explicit, codifiable steps. The ones that resist are the ones demanding flexibility and judgment that nobody has managed to fully specify. Autor’s data on American workers from 1979 onward shows precisely that split, with routine middle-skill work contracting sharply while judgment-heavy work grew.
So the rule is not a prediction about AI. It is a sixty-year-old observation about the difference between knowing and being able to say what you know, pointed at your own organization. There is no particular reason this wave of automation should be the exception.
What is left
Applied honestly, the rule leaves three things. They are not software skills. They show up in a marketing organization, a finance team, a clinical operation, and a legal department in nearly identical form.
Knowing what is worth doing. No system will tell you which problem deserves the next quarter, which customer complaint is a signal and which is noise, or which regulation applies to your particular situation in your particular market. This is domain judgment, and it is built the slow way, by being close to a business for a long time and paying attention. It is also the thing most at risk of being undervalued right now, because it is quiet and hard to demonstrate in a meeting.
Being able to say precisely what you want. This is the skill that has quietly changed status. When work was slow, a vague request got corrected along the way. Somebody came back with questions. The imprecision surfaced early and cheaply. Now a vague request produces a fast, confident, complete-looking answer to a question you did not quite ask, and you may not notice for some time. The ability to specify what you actually want, in enough detail that it could not be reasonably misread, has gone from a nice quality in a colleague to a load-bearing skill. It is the same discipline good requirements writing always was. The cost of skipping it has changed.
Knowing how you would know it was wrong. I wrote about this at length recently, so I will keep it short here. The ability to look at a finished, confident, plausible piece of work and construct an independent check on it is the scarcest of the three, and the one your organization is least likely to have staffed deliberately.
Notice what these have in common. Every one of them is a judgment that the system cannot make about itself, which is exactly what the rule predicts.
And on the other side, the honest half of the message: whatever the system can already check, it will keep getting better at, and being excellent at it will differentiate your people less each year. None of it goes to zero. The balance shifts.
What this costs your best people
There is a cost here that gets glossed over in most versions of this conversation, and glossing over it is a mistake you will pay for in retention.
For a lot of your best people, the thing they are good at is the thing being absorbed. They built a career on being fast and precise at the execution. They are, quite reasonably, proud of it. And the move you are asking them to make, from doing the work to specifying and checking the work, will not feel to them like a promotion. It will feel like being moved away from the part they liked.
Some of that feeling is accurate. The craft does not disappear, but it does migrate, and something real is lost in the migration even when the net is positive. Your strongest technical people are the ones most likely to feel this, because they had the most invested in the old shape of the job.
If you respond to that with reassurance, you will lose them anyway and it will surprise you. The version that works is slower and more honest. Name the loss. Agree that it is real. And then be specific about where the skill went, because it did go somewhere: the judgment about what to build and whether it is right is harder than the execution ever was, and it is now the whole job rather than the part you got to after hours. The work asks more of them now, not less. That is a true thing you can say, and it lands only after you have acknowledged the other true thing first.
Where the tools actually struggle
It is a fair challenge and it deserves a straight answer rather than a dismissal.
The tools are improving at the part where they organize their own work, and there is credible progress on systems that structure a problem for themselves rather than waiting for a person to structure it. Anyone telling you that layer is permanently safe is selling something.
But look at where those systems do well and where they struggle, because the pattern is consistent and it is the answer to the question. They perform best on clean, standardized, well-specified problems, the kind that come with a known right answer. They perform worst on messy proprietary work inside an organization that has its own history, its own constraints, and its own peculiar reasons for doing things the way it does. That second category is not an edge case. For your teams, it is the actual job.
Which means the challenge, followed all the way through, arrives back at the same place. The parts of the work with an externally knowable right answer are the parts under pressure. The parts requiring someone who understands your business to decide what right means are the parts that are not. Investing there pays off whether the tools improve quickly or slowly, which is what makes it the sound bet rather than merely the optimistic one.
What to actually say
When the question comes, and it will keep coming, you do not need a list. You need the rule and the honesty to go with it.
Tell them you cannot promise which specific skills will hold their value, because nobody can, and anyone who promises it is guessing. Tell them the parts of their work that carry a mechanical check on correctness are the parts to expect pressure on. And tell them that if the work they love is in that category, the loss is real and you are not going to pretend otherwise.
Then tell them the part that is easy to miss, which is that the work left over is the harder work. It always was. It just used to be the part you got to only after the execution was done.
The rule in this piece rests on Michael Polanyi’s observation in The Tacit Dimension (1966) that we know more than we can tell, and on David Autor’s Polanyi’s Paradox and the Shape of Employment Growth (2014), which named the idea and traced what it has meant for work.
