In 2013, a research team at Oxford published an analysis predicting that 47% of U.S. occupations were at risk of automation within the next twenty years. The figure sent shockwaves through industry. But the calculation left out a variable. It was built around the type of work people do—not the way they exercise judgment while doing it. A decade later, the bottleneck showing up again and again at AI deployment sites sits squarely inside that missing variable.
Look at the occupations where automation is moving slower than forecast, and a pattern emerges: the more skilled and experienced a person becomes, the harder it is to describe in words what they actually do. This paradox has become 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 describe how they do their jobs.
That sounds strange at first. How can someone fail to explain work they do every day? But when a person has repeated a task long enough, the micro-judgments involved sink below the level of conscious awareness. The senses react before analysis kicks in; the hands move before the mind does. It's a pattern that shows up reliably in anyone who has truly mastered something.
In epistemology, this is called tacit knowledge. The term comes from British philosopher Michael Polanyi, who wrote in his 1966 book: "We know more than we can tell." It's the woodworker's feel for how much pressure and angle a blade needs, the twenty-year veteran nurse who reads a patient's condition the moment she walks into the room, the content strategist who glances at a brief and immediately senses the direction is wrong. None of it was learned from a manual. It accumulated through repetition.
When companies try to convert this tacit knowledge into AI training data, the effort usually starts the same way: ask experienced employees to articulate the criteria behind their decisions. But the longer someone has been doing a job, the more they struggle. A production manager at a manufacturing firm captured it in an interview: "I can look at the line and know something's wrong. If you ask me to explain it, 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 also bound up in the limits of the interview method itself. Ask someone "how did you decide that?" and they'll often offer a post-hoc rationalization rather than a real account of the cognitive process. Psychology calls this hindsight explanation bias—a gap opens between what actually happened in the mind and what gets said out loud. The pattern keeps appearing in cases reported by The Economist: companies discover too late that making tacit knowledge explicit has to come before automation, not after.
Something Gets Lost the Moment You Try to Spell It Out
What's even more striking is that the act of trying to articulate tacit knowledge can itself produce unexpected side effects.
Psychology has a name for this: verbal overshadowing. When people are asked to describe an ability or a memory in words, their performance on that ability afterward actually declines. Experiments have repeatedly shown that asking athletes to analyze their movements in detail temporarily degrades their on-field performance. Making something conscious interferes with the automated function. Sports psychologists know this well—it's the same reason a master can briefly lose their edge when trying to teach a beginner.
Similar effects have been reported in corporate settings. When experienced employees are asked to document their decision-making process, some begin consciously scrutinizing what they previously handled on autopilot, and their speed and accuracy take a temporary dip. It's a paradox: the very process of building AI training data can briefly undermine the efficiency of the people providing it.
That said, claiming tacit knowledge is an absolute barrier to automation would be an overstatement. It's worth being honest about the counterargument here. Some researchers point out that "hard to explain" often reflects a limitation of the explanation method, not the knowledge itself. Observation-based data collection, video analysis, and behavioral tracking can capture a substantial portion of tacit knowledge without ever requiring it to be verbalized. AI has already reached impressive levels in domains once thought to require expert intuition—radiology, credit scoring, legal document review. The warning against mythologizing tacit knowledge as some impenetrable shield is worth taking seriously.
But the domains where that counterargument holds and where it doesn't are fairly distinct. Work that is highly repetitive with clear decision criteria can be turned into data, and AI closes the gap with experienced practitioners quickly in those areas. Judgment calls colored by relational history, responses attuned to contextual nuance, decisions made in the absence of clear criteria—those are different. The data needed to train AI in those zones is fundamentally harder to construct, and that's where most AI training projects stall.
What This Debate Means for Independent Operators
Seen through the lens of the solo operator or independent creative, this dynamic raises some concrete questions.
The first: which parts of my work resist being put into words? It helps to sort your daily tasks into two columns—things you can turn into a checklist, and things you can't. Work with defined procedures and clear criteria is the territory AI can automate or accelerate quickly. Work whose rationale is hard to articulate, that demands a different approach in every situation, and that is saturated with accumulated context and relationship—that's work where even building training data is a challenge.
I think of this sorting as mapping your work. Once you have the map, you can see which regions are explicable and which ones only become accessible through accumulated experience. In an environment where AI adoption is accelerating, the most practical investment right now is spending more of your time deepening the second region.
The next question: are you actively building that tacit knowledge today? The more AI handles, the more the differentiator becomes how many judgment calls a person accumulates in the areas AI can't easily reach. Continuing to pour your time into what AI already does well, or withdrawing from the kinds of experiences that build depth in AI's blind spots—both are misallocations.
Here's a useful exercise: write down ten things you do that AI would struggle to replicate but that you do well. Reading what a client really needs beneath what they actually say. Sensing that a project is starting to go sideways before anything obvious has gone wrong. Knowing which direction is right out of a dozen plausible options. Picking up in the first paragraph of a long manuscript that the whole tone is off. Things like these.
Once you have the list, it tends to split into two types: things that could eventually be made explicit with enough effort, and things that can't be internalized without hundreds of repetitions. The more you have of the second kind—and the more you are still actively accumulating that experience—the more durable your position becomes as AI adoption accelerates.
Research looking at the business landscape of the 2030s tends to converge on a consistent set of capabilities: reading social context with nuance, making judgment calls in unstructured situations, sustaining long-term relationships built on trust. These are consistently flagged as the hardest for technology to replace—and they're also the areas where tacit knowledge concentrates most densely.
The spot where a company's AI training project hits a wall is the same spot where that employee's most durable strengths are located. The more you accumulate the kinds of experiences that resist articulation—and the more time you spend afterward reflecting on them—the thicker that advantage grows. When the wave of automation runs high, the people who last longest are those who already know which layer of the ocean floor it can't reach.
