Teaching AI tacit knowledge is fraught, because much of “how people work” is unteachable

“Teaching AI how people work” runs into a problem that sounds small, but it is a make-or-break constraint: a lot of what humans do is tacit. The Economist’s point is blunt. Tacit knowledge is vital to many jobs, and that kind of know-how is difficult, often impossible, to capture in a clean, rule-based way.

Start with the core issue. Many jobs depend on tacit knowledge, meaning the skills and judgments people apply without being able to fully explain them. If you are trying to train an AI model to do the same work, you do not just need examples of outcomes. You need the hidden reasoning behind them, and that reasoning is usually not neatly written down. So attempts to “teach” AI by feeding it data or instructions can stumble, because the most important part of the job is not the visible steps. It is the context, the judgment, the feel for what matters, and the sense for when something is off.

That matters for how organizations plan AI deployments. Most AI projects are pitched as if performance is a function of inputs and algorithms. But when the bottleneck is tacit knowledge, the bottleneck shifts to something operational: who understands the work deeply enough to translate it into signals the system can use. In practice, that translation is costly. It can require process redesign, careful human-in-the-loop workflows, and a governance model that treats AI as an assistive system rather than a full substitute.

There is also a governance and liability angle that boards cannot ignore. If AI is used in domains where tacit knowledge drives correctness, then errors may not look like ordinary mistakes. They can be subtle and context-dependent, the kind that do not show up in a retrospective dataset review. That increases the risk profile for responsible deployment, because oversight becomes less about checking one standardized output and more about monitoring behavior across conditions. In other words, even when the model technically “runs,” the organization has to ask whether it is applying the right judgment. And tacit knowledge is where that judgment lives.

Regulatory expectations, where they are emerging or tightening, usually map to this reality: systems should be tested and managed in ways that reflect real use. While the specific rules differ by jurisdiction and application, the broad theme is consistent: you cannot assume that lab performance equals operational performance, especially when the task depends on context and human judgment. That is why tacit knowledge is not just a technical nuisance. It is a structural reason why AI assurance, audits, and documentation have to focus on the gap between what the system learned and what people actually do.

There is a second-order effect for capital allocation and strategy. If tacit knowledge is vital and hard to encode, then the timeline for meaningful capability often becomes longer than teams forecast. Organizations may invest in model training, tooling, and data pipelines, only to find that the key limitation is not compute or dataset size, but the inability to formalize the judgment embedded in real work. That can reshape budgets, contract terms with vendors, and internal expectations about what “automation” really means.

For leaders, this also changes how you measure success. If your KPIs focus only on accuracy on curated tasks, you can miss the failure modes that occur when the system faces messy, ambiguous situations. Tacit knowledge is precisely what humans use to handle ambiguity, so the system needs evaluation that covers the gray areas, the exceptions, and the cross-cutting considerations that do not live in neat labels.

The strategic stakes are high because the AI race encourages overreach. The temptation is to treat human work as a set of transferable steps, then let algorithms scale them. But when tacit knowledge is vital to many jobs, that assumption breaks. Decision-makers should treat AI adoption as a partnership between models and people, at least until tacit reasoning can be reliably approximated through observation, feedback, and workflow design. The prize is real, but so are the constraints: if you try to “teach” the invisible, you should expect friction, because the most important skills are often the ones humans cannot fully spell out.