In 2013, a team of Oxford researchers estimated that 47 percent of U.S. jobs faced a high risk of automation within the next two decades. The figure sent shockwaves through industry at the time. But the calculation left out one variable: the kind of work involved. The model measured what a job entailed — not how the person doing it made judgment calls or carried the work out. A decade later, the bottleneck that keeps showing up on the front lines of AI adoption lives inside that missing variable.
Look at the occupations where automation has moved more slowly than expected, and a pattern emerges: the more skilled and experienced someone is, the harder it becomes for them to put what they do into words. That paradox has turned into the most unexpected obstacle in corporate AI training projects.
What Happened When Companies Asked Employees How They Work
Companies trying to train AI systems on their internal workflows keep running into the same wall: employees can't accurately explain how they actually do their jobs.
That might sound strange at first — how can someone not explain what they do every day? But when a person repeats a task long enough, the judgment calls involved sink below conscious awareness. Instinct fires before analysis; the hands move before the thought forms. It's a pattern common among skilled workers.
In epistemology, this is called tacit knowledge. The British philosopher Michael Polanyi coined the idea in his 1966 book, writing that "we know more than we can tell." Think of the feel a craftsman uses to calibrate pressure and angle while carving wood, the instinct a nurse with twenty years of experience uses to read a patient's condition the moment she walks into the room, or the gut sense a content planner gets that "the direction is wrong" after a single glance at a proposal. None of it was learned from a manual — it accumulated through repeated experience.
When companies try to convert this tacit knowledge into AI training data, the attempt usually starts the same way: they ask a skilled employee to spell out, in words, the criteria behind their judgment calls. And the more experienced the employee, the more they freeze up. A production manager at a manufacturing company summed it up in an interview: "I can look at the line and know something's wrong. If you ask me to explain why, it takes a while. And even after I explain it, I'm not sure that's actually how I'm making the call."
The problem is tied to a limitation of the interview method itself. Ask an employee "how did you decide that?" and more often than not, what comes back isn't the actual reasoning process but a rationalization built after the fact. Psychologists call this "post-hoc explanation bias" — a gap between the cognitive process that actually occurred and the explanation put into words afterward. The pattern shows up again and again in the companies The Economist has recently reported on. Businesses are discovering, later than they'd like, that making work explicit has to come before automating it.
The Moment You Put It Into Words, You Lose Something
What's even more striking is that the act of trying to pull tacit knowledge out into language produces an unexpected side effect of its own.
Psychology has a name for it: verbal overshadowing. When someone is asked to describe a skill or a memory in words, the accuracy of that skill afterward actually drops. Experiments repeatedly find that when athletes are asked to break down their own movements in detailed verbal terms, their performance temporarily suffers. The interference happens the moment an automated function becomes conscious. It's a well-documented phenomenon in sports psychology — the same mechanism behind the familiar experience of an expert's performance wobbling the moment they try to teach a novice.
Something similar shows up on the corporate side. When skilled employees are asked to document their decision-making process, they start consciously noticing things they used to handle unconsciously — and speed and accuracy temporarily suffer as a result. It's a paradox: building AI training data can, in the short term, make the very experts supplying the data less efficient.
That said, it would be a stretch to claim tacit knowledge is an absolute barrier to automation, and the counterargument deserves an honest look. Some researchers point out that "hard to put into words" is often a limitation of the explanation method, not of the knowledge itself. Observation-based data collection, video analysis, and behavioral tracking can capture a good deal of tacit knowledge that never gets verbalized. AI has, in fact, already reached a formidable level in fields once considered the domain of expert intuition — reading radiology scans, assessing credit risk, reviewing legal documents. The warning that mythologizing tacit knowledge as an absolute wall can shade into denial of reality is worth taking seriously.
But the line between where this counterargument holds and where it doesn't is fairly clear. Work that's highly repetitive and governed by clear-cut criteria can be turned into data, and AI catches up to experts quickly in those areas. Judgment calls shaped by the history of a relationship, that demand sensitivity to situational nuance, or that get made where no clear standard exists, are different. The data needed to train a model on them is far harder to construct in the first place — and that's largely where AI training projects hit their wall.
What This Debate Means for Solo Entrepreneurs
Viewed from the position of a solo entrepreneur or independent planner in Korea, this whole discussion raises a practical question.
The first is: which parts of what I do can't be put into words? Try sorting your own work into what can be turned into a checklist and what can't. Work with fixed procedures and clear criteria is exactly the territory AI can quickly replace or streamline. Work where the reasoning is hard to explain, where every situation calls for a different approach, and where long relationships and context are baked in, is work where building training data is difficult to begin with.
I'd call this exercise "mapping your own job." Draw the map, and you can see which parts are explicit and which parts are only reachable through accumulated experience. In an environment where AI is moving in fast, spending time thickening that second category is the most realistic preparation available right now.
The next question is: am I actively building that tacit knowledge right now? The more AI tools handle for you, the more the gap comes down to how much experience you're accumulating in judgment calls those tools can't touch. Pouring your time into tasks AI already does well, or letting your experience in the areas where AI struggles atrophy — both directions misallocate your most limited resource.
Here's an exercise worth doing now: write down ten things you do that AI has a hard time imitating but you're good at. The ability to catch a need a client never stated out loud. The instinct to sense the mood shift right before a project goes sideways. The gut call, out of a dozen possible directions, for which one is right. The editorial ear that knows a manuscript's tone is off the moment you finish reading it.
Once you make that list, it tends to split into two kinds: things that can eventually be made explicit if you put in enough time, and things that no amount of explanation gets you to without hundreds of repetitions. The more items fall into the second category — and the more you keep accumulating that experience now — the longer your position holds up as AI adoption accelerates.
Research on the business landscape of the 2030s converges on a similar set of capabilities: reading complex social context, making judgment calls in unstructured situations, and sustaining long-term relationships built on trust. These are consistently flagged as the hardest for technology to replace — and they're also where tacit knowledge concentrates most densely.
The point where companies get stuck trying to teach AI an employee's job is exactly where that employee's most durable strength lives. Experience more of what's hard to explain, spend time replaying that experience for yourself, and that strength keeps getting thicker. The higher the wave of automation rises, the more the people who know first where the water doesn't reach are the ones who keep their footing.
