For the last two years, just about every conversation we’ve had about AI and careers has revolved around the same question. What can it do?
Can it write the email? Can it build the deck? Can it code the feature? Can it analyze the data? Can it replace the junior analyst, the copywriter, the paralegal, the finance guy?
It’s a reasonable question. But it may be the wrong one to be organizing a career around.
The “what can it do” question has an answer that keeps changing. Every six months, the list gets longer. By the time most professionals have adapted to the current capabilities, the next generation of models has expanded the list again. They’re playing defense against a moving target, and the target is moving faster than they are.
Perhaps, a much more useful question is hiding underneath. What can it not do?
That question has a more stable answer.
If you read the WEF Future of Jobs Report 2025 with this question in mind, the picture gets very interesting.
The number one core skill employers want in 2025 is analytical thinking, with 7 in 10 companies considering it essential. Number two is resilience, flexibility, and agility. Number three is leadership and social influence. Then creative thinking. Then motivation and self-awareness. Then technological literacy. Then empathy and active listening. Then curiosity and lifelong learning.
Notice anything? The top of the list is dominated by the things AI is bad at. Reasoning under uncertainty. Adapting when conditions change. Influencing people. Generating genuinely original ideas. Sitting with another human and actually understanding them.
Now look at what’s dropped. Compared to the 2023 edition, dependability, attention to detail, and quality control have all decreased in perceived importance. Those are precisely the categories where AI excels. The market is, very clearly, downgrading the skills software handles well and upgrading the skills it doesn’t.
Even more telling: leadership and social influence is the single biggest climber from 2023. As AI capabilities exploded across the same period, these uniquely human skills didn’t get pushed out of the top of the list. They got pulled up.
We’d say that’s a signal.
The categories AI keeps tripping over
It’s worth being specific about what falls into the “can’t” bucket, because the answers aren’t always obvious.
AI can produce a competent answer to almost any well-formed question. What it can’t do is figure out which question to ask in the first place. Walking into a messy situation, deciding what actually matters, separating the symptoms from the disease — these are the early-stage thinking tasks that the model can’t help with, because by the time the prompt has been formulated, the hard work is mostly done.
AI can analyze data you give it. What it can’t do is notice that the data you have is misleading, or that the most important variable isn’t being measured, or that the question your boss is asking isn’t the one she actually needs answered. Judgment about what to pay attention to operates one layer above the work the tool does.
AI can write something that sounds like a person. What it can’t do is have the specific lived experience that makes a person’s perspective rare. The angle that comes from having actually built a company, lost a relationship, raised a kid, lived in a foreign country, made the hard call. That’s not a content gap. That’s a category gap.
AI can give you a generic answer fast. What it can’t do is give you a specific answer that’s right for your particular situation, because it doesn’t know the politics of your organization, the personality of the person across the table, the history of why this customer is sensitive about this issue. The context that makes an answer actually useful sits in a human’s head, not the model’s.
These aren’t temporary limitations that the next version will fix. They’re structural.
They come from the fact that AI is a system that processes patterns in language. It doesn’t have a stake in any outcome, doesn’t actually understand what’s at risk, doesn’t carry the years of accumulated context that make a professional’s judgment valuable.
Why the “what it can’t do” answer is more stable
The reason this reframe matters so much is that it gives you something durable to invest in. The “what can it do” list is a moving target. Every capability built around the current limits gets undermined when the limits move. The professional who built their identity around being good at writing first drafts is in trouble. The professional who built their identity around being good at running standard reports is in trouble. The capabilities they invested in were exactly the ones AI absorbed first. The “what can it not do” list is much more stable. The skills at the top of it (judgment, taste, original perspective, the ability to read a room, the capacity to make a hard call when nobody’s giving you the answer) have been valuable for a hundred years and are about to become more valuable, not less. The skills likely doing the disappearing are heavily weighted toward the routine and predictable. The skills surviving are heavily weighted toward the things humans bring that machines don’t.
If you want to bet your career on something stable, bet it on the “can’t” side of the ledger. The “can” side keeps moving. The “can’t” is more stable.
How to actually invest in the “can’t” skills
A few specific moves if you want to operationalize this reframe.
First, audit your work for the parts you actually own. What does this job specifically need from you that nobody else and no software could provide? If the honest answer is “not much,” that’s the most important career data point you’ll get this year. Use it.
Second, deliberately put yourself in situations where AI can’t help. Have the difficult conversation in person. Make the judgment call without polling the model first. Form your own view on a complex question before you ever open a tool. The muscle for these things atrophies if you don’t use it, and the people losing the muscle are the ones who quietly become replaceable.
Third, invest in the skills that compound with experience. The WEF’s top of the list (analytical thinking, leadership, creative thinking, empathy, judgment) are all skills where a person with twenty years of deliberate practice is dramatically better than someone with two. They’re also skills you can’t shortcut by buying a course. They get built one hard situation at a time. Start collecting hard situations on purpose.
Fourth, build your own context. The thing AI doesn’t have is your specific accumulated knowledge, your specific relationships, your specific track record. Every year you stay in a domain, every relationship you maintain, every problem you sit with deeply, you’re building the asset AI can’t compete with. People who treat their careers as a portfolio of accumulating context tend to do well. People who treat them as a sequence of interchangeable jobs do not.
The bottom line
The “what can AI do” question is probably going to keep dominating headlines, conversations, and panicked LinkedIn posts for the next few years. The list will keep growing. Every new model will spawn a fresh wave of “is your job next” articles.
Somewhere in an office today, two people are deciding something hard. There is no clean data set in front of them. The model can draft the memo afterward, but it can’t sit in the chair. It can’t carry the weight of having been wrong before, or the small, specific knowledge of who in the room will need convincing and who already agrees.
That chair is still empty. Someone has to sit in it.
