Captured knowledge is the one AI advantage that compounds
Most advantages a firm builds with AI decay as models improve, because the improvement is available to everyone; a captured body of proprietary knowledge does the opposite.
Almost everything a firm can build with AI gets less valuable as the models improve. The custom pipeline is absorbed into the platform. The prompt library is made redundant by a model that no longer needs the scaffolding. The implementation head start closes when the capability ships natively — see Competence at AI implementation is not a durable differentiator. In each case the improvement that erodes the advantage is available to every competitor simultaneously, which is precisely why it erodes.
A captured body of the firm’s own knowledge behaves the other way round.
The mechanism
The asymmetry is simple and it is the whole argument. A better model extracts more value from the same corpus. The firm that has written down how it actually works, what it knows about its clients, and what it stands for holds an asset that each model release makes more productive without the firm doing anything further. The firm that has not written any of it down falls further behind with each release, for the same reason and at the same rate.
This is why the timing of the work matters more than it appears to. The instinct to wait — for the tooling to settle, for the vendor landscape to resolve, for someone else to go first — is reasonable applied to tool selection and actively costly applied here. Waiting on a decaying advantage loses very little. Waiting on a compounding one loses the compounding.
What gets captured
Three kinds of knowledge sit outside what any model has been trained on: how the firm actually practises its method rather than how the method appears in a proposal, what the firm knows about specific clients, and what the firm is for. The first is the largest by volume and the least likely to have been written down; the reasons are in Make tacit knowledge explicit, or AI cannot use it. The second and third are covered in Defensibility lives in what AI can’t access and AI as an operational interpreter of purpose, vision and values.
Given that material, an AI stops producing generic output and starts producing output that is recognisably the firm’s — answers given the way the firm would give them.
Standing of the claim
The mechanism is straightforward. The market observation accompanying it in the source — that no firm has yet done this well — is one practitioner’s assessment rather than a surveyed finding, and it is the kind of claim that should be expected to date quickly. The compounding argument does not depend on it.
The precondition is the discipline described in Start with knowledge management, not tools, and the reason that discipline is now affordable is Knowledge management failed on labour cost, and that is the part AI absorbs.