The wonderful world of tacit knowledge

Anyone who has put AI to work in a specialized domain probably knows the feeling. A draft may be competent, an approach sound, and yet something is still missing, or things have taken a turn one would not have chosen. The work reads as if it were done by a very good outsider. On routine tasks the difference hardly matters. In the corners of a profession where the value is created, it very much matters.

An agent working through an assignment faces judgment calls at every turn. What to collect first, which source to trust over another, when the standard approach fits and when it has quietly stopped fitting. It makes each call the way a capable outsider would, on reasonable defaults. Experienced knowledge workers make the same calls without noticing them as calls. However, when asked how they arrived at a particular result or conclusion, they produce a sequence of steps, because steps are what writing captures easily. The calls live one level below, in behavior that professionals almost never record.

That missing layer has a name, and a research tradition that has studied it for nearly seventy years. This article walks through what tacit knowledge is, why agentic work has made it more valuable than it has ever been, and what we are doing about it in our products built for transfer pricing professionals.

What tacit knowledge is

The explicit part of a profession is everything that can be written down and taught directly. The handbooks, the methods, the templates, the steps of the work. Tacit knowledge is the part professionals use without putting it into words. It can be formulated, but doing so does not come naturally and takes deliberate effort, sometimes training. It works as pattern recognition. A situation resembles hundreds of earlier ones, something in it stands out, and the experienced person acts on the resemblance long before any rule could be quoted.

The concept comes from Michael Polanyi, an eminent physical chemist who turned philosopher of science in the second half of his career. He introduced the idea in Personal Knowledge in 1958 and compressed it into one sentence eight years later. “We can know more than we can tell.”¹ Psychology later showed how deep the gap between knowing and telling runs. People asked to explain their own thinking construct a plausible account rather than report the process that actually ran.² Expertise brings a bias of its own. Experts turn out to be poor at predicting what a novice needs to be told, and the bias resists training.³ In the 1990s the thread reached management science, where Ikujiro Nonaka placed the conversion of tacit knowledge into explicit knowledge at the center of how companies create anything new.⁴ In parallel, a school of decision researchers formed around Gary Klein’s interviews with fireground commanders and spent the following decades building methods for drawing these patterns out of experts’ heads.⁵

For practical purposes the defining property is this. The knowledge does not come out by asking. When we sit experts down and ask them to document what they do, we get back a genuinely believed description of the work with the judgment removed, or at best a thin sketch of how they got there. And it is not like they are hiding anything. The patterns surface only while the work is being done, when a real case is on the table. For example, a seasoned professional reads a group’s pricing policy, the transaction types, the amounts, the jurisdictions, and within minutes points at two likely weaknesses from an audit perspective. There was no framework behind the read. The weaknesses stood out to an intuition that has seen decades of policies and where they failed. Asked about their thought process, the professional assembles an explanation after the fact, and it comes out looking like the kind of high-level overview handed to first-year consultants in onboarding. This is not to say that extracting the thought process is impossible. But it requires dedicated interview techniques and knowledge documentation frameworks.

Why the agentic age brings it back

Tacit knowledge is an old subject, and for years it lived mostly in philosophy and in management and organization studies. Agentic work has revived it, for a structural reason. AI models now carry most of the explicit layer of nearly every profession, the rules and standards, the textbook methods. The routine application of that layer, the drafting, the computations, the documentation that fills most of the profession’s hours, is increasingly within an agent’s reach. What general-purpose AI has trouble applying is the knowledge or thought process in the unwritten layer.

That gap is what makes extraction so valuable now. The judgment the profession runs on rarely exists in a document an agent could be given. Carefully drawn out of experts and written into the skill folders agents work from, it becomes something agents can actually apply, and they begin to make calls the way the best people make them. The follow-up effect is scale. Once written into a skill, a rule runs on every case the agents touch, and its value grows with each additional agent that follows it. And a skill is never finished. Each new correction refines it further, so the agents keep improving on the practice’s own work.

How we put it to work

Getting knowledge out of people’s heads is itself a studied craft. The field is called cognitive task analysis, and it has produced interview methods with decades of use in firefighting, medicine, and the military. In the Critical Decision Method, an interviewer walks an expert back through one hard case in structured passes and ends with a deep account of a single judgment.⁵ Applied Cognitive Task Analysis, designed so that practitioners can run it without a research background, maps where the cognitive demands of a whole task are dense.⁶ The craft was codified in the 2006 handbook Working Minds,⁷ and recent work has carried it from firegrounds and cockpits into managerial knowledge work.⁸

At Supernomial we are applying these methods to transfer pricing knowledge work. Our agent teams work from skills, folders of written instructions and resources that an agent loads when the task calls for them, and we write ours to carry judgment as well as procedure. We are testing two extraction routes side by side. In one, we capture judgment live, at the moment an expert corrects a specific piece of agent work, right when a pattern has surfaced on a real case. In the other, we sit practitioners down for structured sessions adapted from the interview methods above, aimed at the cases they still think about years later.

The same approach is open to any organization putting agents to work. The experience sits with its people, and the methods described above can draw it out and build it into agents that arrive with what a new colleague would need years of cases to acquire. The result is a system of agents that works the way the organization’s own best people work, and that keeps improving with every new piece of judgment captured. Tacit knowledge has spent nearly seventy years as a research subject. Inside these systems it becomes a working part of the organization.

References

1. Polanyi, M. (1966) The Tacit Dimension. Garden City, NY: Doubleday. Reissued 2009, Chicago: University of Chicago Press. Available at: press.uchicago.edu. ↩
2. Nisbett, R.E. and Wilson, T.D. (1977) ‘Telling more than we can know: Verbal reports on mental processes’, Psychological Review, 84(3), pp. 231–259. Available at: doi.org. ↩
3. Hinds, P.J. (1999) ‘The curse of expertise: The effects of expertise and debiasing methods on prediction of novice performance’, Journal of Experimental Psychology: Applied, 5(2), pp. 205–221. Available at: doi.org. ↩
4. Nonaka, I. and Takeuchi, H. (1995) The Knowledge-Creating Company: How Japanese Companies Create the Dynamics of Innovation. New York: Oxford University Press. Available at: global.oup.com. ↩
5. Klein, G.A., Calderwood, R. and MacGregor, D. (1989) ‘Critical decision method for eliciting knowledge’, IEEE Transactions on Systems, Man, and Cybernetics, 19(3), pp. 462–472. Available at: doi.org. ↩ ↩
6. Militello, L.G. and Hutton, R.J.B. (1998) ‘Applied cognitive task analysis (ACTA): A practitioner’s toolkit for understanding cognitive task demands’, Ergonomics, 41(11), pp. 1618–1641. Available at: doi.org. ↩
7. Crandall, B., Klein, G. and Hoffman, R.R. (2006) Working Minds: A Practitioner’s Guide to Cognitive Task Analysis. Cambridge, MA: MIT Press. Available at: direct.mit.edu. ↩
8. Brown, O., Power, N. and Gore, J. (2025) ‘Cognitive task analysis: Eliciting expert cognition in context’, Organizational Research Methods, 28(3), pp. 375–404. Available at: doi.org. ↩