How do you keep expertise in an organisation?

Expertise survives where experts keep practising difficult work, receive feedback on it, teach others, and take part in the situations where judgement transfers. Documentation preserves the part experts can articulate. Most of what distinguishes them is not that part.

Share this line as a card

This is one of the oldest questions in organisational research and one of the least disturbed by AI at the level of mechanism. What AI changes is not how expertise transfers. It is the supply of occasions on which transfer happens.

Why the repository is not the answer#

Polanyi's formulation is the one to start from: we can know more than we can tell. A large part of expert knowledge is acquired through practice and cannot be fully articulated, which means it does not survive being written down. The Tacit Dimension (1966).

The practical consequence is uncomfortable for anybody running a knowledge-management programme. A complete set of documents can coexist with a complete loss of the capability that produced them. The documents are the residue of the expertise, not the expertise.

Nonaka and Takeuchi's contribution was to describe how the tacit part moves anyway: through shared experience, joint work and apprenticeship rather than through transmission of text. The Knowledge-Creating Company (1995). Their model has been criticised for treating the tacit-to-explicit conversion as more tractable than Polanyi thought it was, and that criticism is worth carrying, but the observation about mechanism holds.

How it actually transfers#

Lave and Wenger gave the process its name: legitimate peripheral participation. Newcomers become competent by doing real but peripheral work alongside practitioners and moving gradually inwards, rather than being taught first and practising afterwards. Situated Learning (1991).

The word carrying the argument is legitimate. The task has to genuinely matter to somebody. Practice work that nobody depends on does not produce the same learning, because the attention, the feedback and the consequences are all absent. This is the same finding the training literature keeps arriving at from the other direction.

Ericsson's conditions describe what has to be true of the practice itself: at the edge of ability, aimed at a weakness, with feedback and correction. Graded entry. Put the two together and the requirement is specific. Difficult real work, done by someone not yet good at it, in view of someone who is, with correction.

What organisations should actually do#

Four things follow, and none of them is a platform.

Keep experts doing hard work. Expertise decays with disuse like anything else, and the decay is faster for cognitive tasks than physical ones. The rates are here. A senior person moved entirely to reviewing machine output is no longer practising the thing being reviewed.

Protect the occasions of transfer. Case review, rotation, sitting in, doing the work badly first in front of somebody who will say so. These are the routines through which the tacit part moves, and they are usually classified as overhead.

Give juniors legitimate work. Not simulation. Work that matters, at a difficulty they can nearly manage, with correction afterwards.

Document the explicit part anyway. It is genuinely worth having. The error is believing it is the whole thing.

The cost of not doing it arrives later and arrives priced. A company that has not built its bench will have to buy its talent later, an expensive and risky strategy.

What AI changes#

The mechanism above holds without any machine in it. What the machine changes is which tasks are available to learn on.

The peripheral work through which newcomers historically entered a practice is disproportionately the work these systems do well. First drafts, first passes, the routine version of the difficult thing. Automate those and the transfer route narrows without anyone deciding to close it, and without appearing anywhere as a decision.

That is a claim about which tasks get automated rather than a measured effect on professional expertise, and the size of it is unknown. The nearest measurements are on individuals rather than organisations: Bastani and Sankaranarayanan both found the harm concentrated in the condition that removed the attempt, and both preserved it with scaffolding.

There is also a second-order effect worth naming as an open question rather than a finding. If people can get task information from a system instead of from a colleague, some of the instrumental relationships through which workplace knowledge travelled weaken. Whether that reduces knowledge-sharing overall or reroutes it is not established, and the research lists it as open.

Organisations forget, and somebody has measured how fast#

Everything above is about how capability moves between people. There is a second literature about what happens to it when the work stops, and that one has numbers in it, because it was done on factories where the cost of every unit was recorded.

Anelí Bongers estimated learning curves for three United States fighter programmes from the Department of Defense's own flyaway costs: the F/A-18E/F Super Hornet across 554 units bought between 1997 and 2013, the F-22A Raptor across 182 units to 2009, and the F-35A across 135. The standard learning curve assumes experience accumulates. Her second model lets it decay, with a persistence parameter estimated by nonlinear least squares. Graded entry.

For two of the three programmes the parameter could not be distinguished from zero. Her words: "the depreciation of experience in the production of these units is total, on an annual basis". For the Super Hornet she estimated 0.072, so "on an annual basis, only 7.2% of the experience is maintained". The programme with the highest production rate was the one that held onto anything.

That last point is the transferable one. Bongers attributes the difference to how often the work was actually being done, and reports Benkard reaching the same conclusion on the Lockheed L-1011: low production rates explain high forgetting. Benkard's own finding, that experience depreciates and that only some of it survives the move to the next generation of aircraft, is where the modern version of this argument starts.

So the four recommendations above are all recommendations about rate. Keeping experts doing hard work, protecting the occasions of transfer, giving juniors legitimate work: each of them is a statement about how often, and the measured literature says that how often is the variable that decides whether anything is still there in a year.

What a fighter programme cannot tell you about a law firm#

Bongers measured unit cost in aircraft assembly. Nothing in her paper concerns professional judgement, and an organisation whose product is advice does not depreciate on the same schedule as one whose product is an airframe. The transfer is by analogy and should be read as one.

Her own limits are substantial. The data are annual procurement lots rather than continuous production, the samples are single figures to mid-teens of observations per programme, and she flags two of her specifications as inconsistent and unusable. The coefficient tables are published as images and the running text carries no sample sizes, standard errors or R-squared, so the 7.2 per cent is quoted here as she states it and nothing further has been checked behind it.

The older human literature is thinner still on proof. Lave and Wenger studied tailors and midwives; Polanyi was doing philosophy. None of it is a controlled test, and the practices recommended above rest on convergence across traditions rather than on trial evidence that they work.

And the dose is still missing where it would matter most. Nobody has measured how much legitimate peripheral work a junior professional needs, over what period, or how much of it can be replaced before transfer fails. What the factory measurements establish is that the question has an answer and that the answer is probably uncomfortable, not what the answer is.

Key sources

• Polanyi, M. (1966). The Tacit Dimension. University of Chicago Press. In the essential works.
• Lave, J. and Wenger, E. (1991). Situated Learning: Legitimate Peripheral Participation. Cambridge University Press. In the essential works.
• Nonaka, I. and Takeuchi, H. (1995). The Knowledge-Creating Company. Oxford University Press. In the essential works.
• Ericsson, K. A., Krampe, R. T. and Tesch-Romer, C. (1993). The Role of Deliberate Practice in the Acquisition of Expert Performance. Psychological Review, 100(3). Graded entry.
• Bongers, A. (2017). Learning and forgetting in the jet fighter aircraft industry. PLOS ONE, 12(9), e0185364. Graded entry.
• Benkard, C. L. (2000). Learning and Forgetting: The Dynamics of Aircraft Production. American Economic Review, 90(4), 1034-1054. Graded entry.

Related SuperSkills research#

On what the organisation holds as distinct from its people, organisational capability. On the knowledge that cannot be written down, tacit knowledge. On the practice conditions, deliberate practice and productive struggle. On the entry route closing, the missing rungs and do apprenticeships still work. On the rate of loss, how fast do skills decay.

About this research#

Rahim Hirji is the author of SuperSkills (Kogan Page, 2026), keynote speaker on AI and human capability, and founder of The SuperSkills Intelligence Company. Tacit knowledge is Polanyi's, legitimate peripheral participation is Lave and Wenger's, and this page follows the rule this research uses for foundational concepts: established work answers the human mechanism, and current studies answer what AI changes.

How this research works  ·  Reviewed quarterly  ·  Found an error? Tell me and it is corrected on the page.