Executive summary
Most organizations are approaching AI as a tool problem: which model, which license, which feature. The more useful question is older and stranger — how do people actually get good at their jobs? — because the answer exposes exactly where AI is powerful, where it is helpless, and why "buy the licenses and run a training" so often fails to change anything.
The evidence is clear in direction even where it is loose in detail. Most workplace capability is built through experience and through other people, not in formal training. The recognizable shorthand for this is the 70-20-10 model, but the defensible position is to lead with the direction and treat the numbers as a heuristic, not a measured constant, rather than quoting the ratios as law.
Layer AI onto that picture and a structural claim emerges, not a hot take. Using Nonaka and Takeuchi's (1995) model of how knowledge converts between tacit and explicit forms, a language model lives almost entirely in the quadrant that recombines what has already been written down, and it increasingly helps turn tacit knowledge into explicit text. The one quadrant it cannot structurally occupy is the tacit-to-tacit channel — person to person, the way most real expertise actually transfers. The true boundary is therefore codified versus contextual, not formal versus informal.
Codified knowledge is written down and transferable — a model can learn it directly. Contextual (tacit) knowledge lives in people — how work actually gets done here — and a model can't absorb it directly; it only gets what your people supply about it.
The implication reframes the whole adoption problem. Tacit, organization-specific context is not the thing AI replaces — it is the scarce input AI needs from your people to be useful at all. A model brings generic codified knowledge; your people bring what good looks like here, the unwritten constraints, the judgment. That is why a tool dropped into an organization without the surrounding culture behaves like expensive autocomplete. Adoption rides the same experiential and social channels as learning itself. This paper sets out the argument and what follows from it for leaders.
1. The question worth asking
Ask a room of executives how their people became good at their jobs and most will gesture, by reflex, at training: the onboarding program, the certification, the annual course catalog. Then ask them to think about their own careers — the moment a process finally made sense, the instinct that now lets them spot a problem before it surfaces — and the story changes. It was rarely the orientation deck. It was the two-minute exchange with someone more senior, the project that went sideways, the offhand "don't send it like that."
This is not a sentimental observation. It is the most robust finding in the workplace-learning literature, and it determines where AI can and cannot help.
2. The hook, and where it breaks down
The famous framework is 70-20-10: roughly 70% of capability from on-the-job experience, 20% from other people, 10% from formal training. It originated in successful-executive research at the Center for Creative Leadership in the late 1980s (McCall, Lombardo, & Morrison, 1988), with the ratios stated explicitly later (Lombardo & Eichinger, 1996).
We use the model as a recognizable hook, then we are precise about what it actually is. The precise split was partly a packaging decision. The numbers came from roughly 200 already-successful executives recalling key developmental events; hundreds of those events were coded, recoded, and trimmed — the share attributed to hardship and personal experience outside work was set aside because it cannot be planned into a course — leaving a tidy three-category ratio.
And the ratios have been credibly challenged. Kajewski and Madsen (2012) found a lack of empirical data behind 70:20:10 and even uncertainty about its origin. Clardy (2018) is the sharpest academic teardown, examining five separate research traditions that all converge on a "70% rule" and finding the evidential basis weak across all of them. Johnson, Blackman, and Buick (2018) identified the experiential 70% as the model's greatest weakness, because experience only develops people when it is structured and supported by feedback and reflection. DeRue and Myers (2014) note that the assumption is frequently quoted as fact despite thin empirical support.
So we do not defend the numbers. We defend the direction — and the direction is well grounded.
3. What is actually well-grounded
Strip away the ratios and the core claim stands on solid research: informal, experiential, and social learning dominate how people build capability at work, and formal training is a thin slice.
Eraut (2000, 2004) ran multi-year studies of professionals, technicians, and managers and found that in every case the majority of workplace learning was informal — the citable backbone for the direction. Lave and Wenger (1991) gave us communities of practice: learning as participation, the micro-interaction channel made explicit. Polanyi (1966) named the underlying phenomenon decades earlier — tacit knowledge, "we know more than we can tell." Cerasoli et al. (2018), a meta-analysis and therefore the strongest evidence class on the list, links informal learning behaviors to performance. Manuti et al. (2015) review the formal/informal balance directly.
The picture that survives scrutiny: capability is built mostly by doing and mostly through other people. Hold that next to AI and something specific falls out.
4. Where AI actually lives
The instinct that "AI is trained on codified work, and the tacit things between people are what it cannot capture" is right. It becomes far stronger — and much harder to argue with — when named precisely rather than asserted.
Nonaka and Takeuchi (1995) describe four conversions between tacit and explicit knowledge: socialization (tacit to tacit, person to person), externalization (tacit to explicit), combination (explicit to explicit), and internalization (explicit to tacit). It is a model of how knowledge moves through an organization, and it maps cleanly onto what a language model does.
A large language model lives almost entirely in combination — recombining knowledge that has already been written down — and it increasingly assists externalization, helping a person turn what they know into text. Those are real, valuable capabilities, and they are expanding fast. The quadrant a model cannot structurally occupy is socialization: the tacit-to-tacit channel of apprenticeship, of watching how a colleague handles a difficult client, of absorbing how a team actually works. Not "has not yet" — cannot, because that transfer happens between two people in a shared context, and a model is not a second person sharing it.
This is the move from hot take to citable structural claim. It is the kind of argument that belongs in a boardroom, not a tips thread.
5. The real fault line: codified vs contextual
A careful reader pushes back here: models have clearly learned plenty from informal sources — forums, chat logs, transcripts, social media are full of casual, informal language. True. Which is why the boundary is not formal versus informal.
What models have not ingested is tacit, embodied, relational, organization-specific knowledge — the part that was never written down anywhere. The real fault line is codified versus contextual. Stated this way, the argument dodges the easy rebuttal and gets sharper: the question is not how casual the knowledge is, but whether it has been captured at all, and whether capturing it transmits the competence it carries.
That last clause matters, because it forces us to retire the word "never." Capture is expanding quickly — meeting transcription, recorded calls, archived chat, multimodal models that take in audio and tone. A great deal that used to evaporate is now recorded and fed back in; that is externalization in action. So the defensible claim is not that AI cannot capture micro-interactions. It is that capturing the artifact of an interaction is not the same as transmitting the tacit competence it carries between two people. A transcript of a mentor's feedback is not the mentorship.
6. The reframe: tacit context is the scarce input
Here the argument turns from a limit into a strategy. The tacit micro-interactions are not only a ceiling on what AI can do. They are the scarce input your people must supply for AI to be useful in your specific organization.
A model brings generic codified knowledge to every task. It does not bring what matters here: what "good" looks like for this team, which client email is about to become a problem, the unwritten constraint that never made it into a policy document, when to trust the data and when to override it. That contribution is not clerical. It is the difference between an output that is plausible and an output that is right for you.
So tacit knowledge becomes the bottleneck input, not the thing AI replaces. And this is precisely why "buy licenses and run a training" fails: it treats capability as something a content-delivery mechanism can install. The 70-20-10 research and the AI argument turn out to be the same shape — high-value capability rides experiential and social channels, and no delivery mechanism shortcuts it, whether that mechanism is a classroom course or an AI license.
7. What this means for leaders
Three consequences follow directly, and they are where the work actually is.
Stop competing with AI; start directing it. For individuals, the losing game is racing the model at the codified work it was built for. The winning move is to supply the context it cannot have and direct it — you bring what matters here, it brings speed and reach, you stay accountable for the result. That is a new kind of work, and it is the subject of our companion thesis, You Will Lead AI Labor.
Treat adoption as a social process, not a software rollout. If capability transfers through experience and people, then so does AI adoption. Tools spread the way every other practice spreads inside an organization — through the people others watch and trust. A license does not change behavior; a respected colleague modeling a better way does.
Invest in the human channel deliberately. The scarce input — tacit, contextual knowledge — has to be surfaced, structured, and put to work on purpose. That is a culture investment, not a procurement line item.
This is why we describe the outcome as a Trusted AI Culture, built on three things an AI license cannot supply:
- Frameworks — clear instructions and workflows your team actually follows, so tacit "how we do it here" becomes something AI can be pointed at consistently.
- Best Practices — the right ways to work with AI, embedded across the team so good judgment travels person to person, through the one channel that actually moves it.
- Internal Champions — the people inside your organization who own and sustain the momentum. In the language of this paper, Champions are the socialization channel a model structurally cannot be. They are how adoption sticks after any engagement ends.
8. Conclusion
The codified/contextual line is the most useful boundary a leader can hold in mind right now. On one side is everything that has been written down — where AI is genuinely strong and getting stronger. On the other is everything that was not — the tacit, contextual, relational knowledge that most expertise is actually made of, and that AI needs from your people to do anything worthwhile in your organization.
Getting this right is not about buying a better model. It is about building the culture that lets a model reach the context only your people hold. That is the work we do, and it is the reason the tool never shortcuts the culture — it runs on it.
References
Cerasoli, C. P., Alliger, G. M., Donsbach, J. S., Mathieu, J. E., Tannenbaum, S. I., & Orvis, K. A. (2018). Antecedents and outcomes of informal learning behaviors: A meta-analysis. Journal of Business and Psychology, 33(2), 203–230.
Clardy, A. (2018). 70-20-10 and the dominance of informal learning: A fact in search of evidence. Human Resource Development Review, 17(2), 153–178.
DeRue, D. S., & Myers, C. G. (2014). Leadership development: A review and agenda for future research. In D. V. Day (Ed.), The Oxford handbook of leadership and organizations (pp. 832–855). Oxford University Press.
Eraut, M. (2000). Non-formal learning and tacit knowledge in professional work. British Journal of Educational Psychology, 70(1), 113–136.
Eraut, M. (2004). Informal learning in the workplace. Studies in Continuing Education, 26(2), 247–273.
Johnson, S. J., Blackman, D. A., & Buick, F. (2018). The 70:20:10 framework and the transfer of learning. Human Resource Development Quarterly, 29(4), 383–402.
Kajewski, K., & Madsen, V. (2012). Demystifying 70:20:10 [White paper]. DeakinPrime, Deakin University.
Lave, J., & Wenger, E. (1991). Situated learning: Legitimate peripheral participation. Cambridge University Press.
Lombardo, M. M., & Eichinger, R. W. (1996). The Career Architect development planner (1st ed.). Lominger.
Manuti, A., Pastore, S., Scardigno, A. F., Giancaspro, M. L., & Morciano, D. (2015). Formal and informal learning in the workplace: A research review. International Journal of Training and Development, 19(1), 1–17.
McCall, M. W., Lombardo, M. M., & Morrison, A. M. (1988). The lessons of experience: How successful executives develop on the job. Lexington Books.
Nonaka, I., & Takeuchi, H. (1995). The knowledge-creating company. Oxford University Press.
Polanyi, M. (1966). The tacit dimension. Doubleday.
Tough, A. (1979). The adult's learning projects (2nd ed.). Ontario Institute for Studies in Education.