2.Context Is Capital

Watch a competent employee join a company. For their first several months they are, by any output measure, worse than they will ever be again, not because they lack skill but because they lack context: what words mean here, which rules are firm and which are theater, what good work looks like to this particular organization, whose sign-off matters, how the product actually gets made. Onboarding is the slow transfer of that context from the organization into the person. It is expensive, it is why senior insiders outproduce brilliant newcomers, and every organization on earth pays for it because it works.

Now watch the same organization use AI. Each session opens with a blank slate. Whatever context the model needs, someone supplies by hand, in the prompt, again. The same brand rules are re-explained in marketing this morning and violated by legal's chatbot this afternoon. The best prompt anyone ever wrote for the quarterly close lives in a departed analyst's notes. The organization is onboarding a new employee thousands of times a day and letting each one go at the end of the conversation.

Naming the asset

The whitepaper's term for what is being re-explained is organizational context: the rules, terminology, goals, style, norms, skills, and process knowledge within which an organization's work happens. Some of it is written down in style guides, wikis, and onboarding decks. Most of it is tacit, living in Slack history and the heads of senior employees, which is why onboarding takes months rather than an afternoon of reading.

What context lacks, in nearly every organization, is treatment as an asset. It appears on no balance sheet. Nobody owns its maintenance. It has no version history, so nobody can say when it changed or why. And under current AI practice it is handled the way one handles an expense: produced, consumed once, discarded.

The leak, itemized

Combine Section 1's tenancy problem with an unmanaged asset and you get three distinct leaks, all running concurrently.

1. Evaporation. Context supplied by hand in a prompt improves one answer and then is gone. The labor of articulating it, often the hardest part, is spent again next session, by someone else, slightly differently. Multiply by every employee, every day.
2. Transfer. Context that does persist, in conversation logs and vendor-side memory features, persists on the vendor's infrastructure, under the vendor's policies. This is the leak Nadella and Smith were describing: the tacit knowledge of the firm, extracted through daily use, accumulating on the other side of the property line.
3. Turnover. Context that lives only in people leaves with them. This leak predates AI, but AI raises its price, because articulated context is exactly what AI systems can use, and organizations that never captured it now feel the missing asset every time they prompt.

Evidence that context is the mechanism

Is context actually where AI's value comes from, or is that just a tidy story? The best evidence the paper cites is a randomized study, and precisely what it found matters.

The mechanism matters more than the headline. The assistant did not make workers better by being intelligent in the abstract; it made them better by carrying codified organizational know-how, the accumulated practice of the best performers, to everyone else. The productivity gain came from context, captured once and distributed. That is the onboarding example again, run as a controlled experiment. Note its limits, because this guide leans on it: one study, in one industry, and the codified know-how was built into the assistant's training rather than handed to it as an artifact the organization owned. What it demonstrates is the mechanism, that codified organizational context rather than raw model capability produced the gains. Carrying the result beyond customer support is an extrapolation, the paper's and ours; the sections that follow treat it as the working hypothesis the architecture is built to exploit.

Which sharpens the problem. If context is the mechanism, then the leaks itemized above are not housekeeping issues. They drain the specific asset that determines whether AI produces returns, and they drain it into places (vendor logs, ex-employees, thin air) where the organization cannot reinvest it. The organizations reporting zero financial impact in the PwC survey are, on this reading, not failing to buy enough intelligence. They are failing to accumulate the asset that makes intelligence pay.

What would fixing this require? Something with an unglamorous shape: a way to write context down, give it an owner and a version history, scope it to where it applies, and hand it automatically to every AI system and every person doing the organization's work. In other words, the fix is an artifact, and it is the first piece of the architecture this guide has been building toward.