Where Your AI Practice Lives

Ask a leadership team where its AI capability sits and the answer describes a programme: licences, training, a policy, a set of pilots. Ask the same question one level down, as a physical question about location, and the answer changes. The working version lives in somebody's account. The instructions that make it work are in a file on their laptop. The reason it is built that way is in their head.

Nobody chose that arrangement. It is what happens when a general-purpose tool meets people who work out how to use it one job at a time. The failure is organisational: nothing in most firms turns that into something a colleague can pick up, and the thing being lost is the only part that was difficult to get.

Experience depreciates, and this has been measured#

Organisational learning is usually discussed as accumulation. The more careful literature finds it decays.

Bongers estimated learning curves from procurement cost across three fighter aircraft programmes and found a persistence parameter not significantly different from zero for two of the three, meaning accumulated production experience could depreciate almost completely within a year. Benkard, working on a wide-body airliner, found the data inconsistent with simple learning and supportive of organisational forgetting, with only partial transfer of accumulated experience across product generations.

Both measure unit cost in aircraft assembly rather than professional judgement, so the transfer to knowledge work is by analogy. The direction is the part to carry: a firm that does not capture how its best work is done does not hold it by default. It loses it, and it was losing it before AI arrived.

Changing how information moves changes who knows what#

Yang and colleagues observed the communication telemetry of 61,182 employees over the first six months of 2020 and found that firm-wide remote work made the collaboration network more static and more siloed, with fewer bridges between distant parts of the organisation, less synchronous and more asynchronous communication. The authors state the effects may make it harder for employees to acquire and share new information.

It measures one very large technology company during a pandemic, it is about communication structure rather than performance, and it has nothing to do with AI. It is included because of the mechanism. A change in how information moves reorganised who knew what, with nobody redesigning a single role and nothing appearing on any report. Putting a capable machine between a person and their work is a change of that kind, made one desk at a time.

Storage is not capture#

Most firms already hold the record. Microsoft's own documentation describes Copilot prompts and responses as stored in a hidden folder in the mailbox of the user who ran the tool, not designed for that user to reach, searchable by a compliance administrator with eDiscovery tools. The vendor documents capability rather than practice, and says nothing about what any organisation has configured or searched.

So the material exists and sits where nobody reads it for this purpose. A compliance archive answers a legal question. Turning the same material into practice a colleague can use is editorial work that somebody has to be given time to do, and no retention setting does it.

The people building the tools have named the same risk#

Tomašev, Franklin and Osindero, writing on delegation frameworks at Google DeepMind, name oversight readiness as a property a future workforce may lack. Their argument is that expertise is built through the repetitive execution of narrowly scoped tasks, that those are the tasks most likely to be handed to agents first, and that automating them fully would deprive junior staff of the experience needed later.

It is a preprint framework paper with no empirical component, so it establishes that practitioners take the risk seriously rather than that the risk has been observed. It is quoted here because of who is making the argument: the people designing the delegation layer, not the people writing about its consequences.

What capture actually requires#

Collins, Brown and Newman argued that apprenticeship works by making expert thinking visible, and that formal education fails at cognitive tasks because the reasoning stays inside the expert's head. Their six methods are modelling, coaching, scaffolding, articulation, reflection and exploration. It is a framework argued from observation with little controlled testing.

Articulation is the step that matters here and the one most organisations skip. Asking somebody to write up their process rarely produces anything useful, because the interesting part is not available to introspection on demand. Sitting with them while they work, asking where each piece is saved and why this way rather than another, produces a usable account. The transcript can do the writing afterwards.

Four things to do with that#

• Ask to be shown rather than told. The question that finds the practice is where is it saved, asked about a job somebody has actually done this week.
• Give it a home that is not one person's account. A shared folder nobody knows exists is as private as a laptop.
• Name an owner who keeps it current. An instruction file written against last quarter's model and owned by nobody is worse than none, because people follow it.
• Let people copy, improve and send it back. A standard that cannot be amended by the people using it will be worked around within a month.

What would settle it#

Measurement of whether captured practice transfers: two comparable teams, one with a maintained shared method and one without, on the same work. Nothing in this base measures it. The organisational forgetting literature establishes that experience depreciates in manufacturing, and the claim that documented AI practice transfers in knowledge work is an argument on this site rather than a finding.

Where this sits in my own argument#

Capability debt is the version of this carried by a person. This is the version carried by an organisation: capability that exists, is working, and is held in a way that nobody has decided about. A firm in that position has an AI capability in the same sense that it has a filing cabinet, which is the distinction the AI readiness lie is about.

Related SuperSkills research#

On retention at the level of the firm, how do you keep expertise in an organisation? On the person's side of the same record, who owns the capability you built at work. On what cannot be written down, tacit knowledge. On measuring use instead of value, usage theatre.

Key sources

• Bongers, A. (2017). Learning and forgetting in the jet fighter aircraft industry.
• Benkard, C. L. (2000). Learning and Forgetting: The Dynamics of Aircraft Production.
• Yang, L. et al. (2022). The effects of remote work on collaboration among information workers.
• Microsoft (2025). Learn about retention for Copilot and AI apps.
• Tomašev, N., Franklin, M. and Osindero, S. (2026). Intelligent AI Delegation.
• Collins, A., Brown, J. S. and Newman, S. E. (1989). Cognitive apprenticeship.

About this research#

Written by Rahim Hirji, author of SuperSkills (Kogan Page, 2026), keynote speaker on AI and human capability, and founder of The SuperSkills Intelligence Company.

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