Capability got cheap. Context did not.

Every competitor you have can rent the same model this afternoon. Pick a provider, put a card in, and by tomorrow morning your smallest rival has access to reasoning capability that is functionally indistinguishable from yours. Three years ago that was not true. Now it is, and the gap it used to protect has closed for good.

The thing that did not get cheap

What has not become a commodity is what your business knows. Your pricing logic. The precedent for the deal you did last month that looks nothing like the rate card. The reason a particular customer gets terms nobody else gets. The state of every job on the board right now, not as of last night's export. None of that ships with the model, and none of it can be bought from a vendor, because it exists only inside your organisation — scattered across systems that do not talk to each other, and in the heads of the two or three people who have been there the longest.

That is the actual constraint now. Not whether a model can reason well enough to help. It can. The constraint is whether it can see enough of your business to reason about your business, rather than about businesses in general.

An extremely expensive graduate

Drop a frontier model into a business with none of this context and you get an extremely expensive graduate on their first day — every single day. It reasons well. It writes well. It has never met your customers, does not know your last three exceptions to the policy, and will confidently produce an answer rather than say it does not know. Give the same model your organisation's context and the answer changes completely, not because the model changed, but because what it can see did.

Seven kinds of context, briefly

We split organisational context into seven kinds. None of them are exotic. All of them are usually scattered across a dozen systems and two or three people's heads, which is the actual problem a context layer has to solve.

• 01
  Entity

  The customer, the job, the asset, resolved to one thing — not five records across the ERP, the CRM, the inbox and the invoicing tool that each describe the same relationship slightly differently.
• 02
  Precedent

  What was decided last time a case like this came up, and why. Not the policy document. What actually happened, including the exception that got made and who signed off on it.
• 03
  Policy

  The standards, rate cards and constraints that are actually enforced day to day, as distinct from the version filed on the shared drive that nobody has opened since it was written.
• 04
  Process

  The sequence a job actually moves through, including the branches and the workarounds — not the flowchart from the onboarding deck.
• 05
  Current-state

  What is true right now: stock on hand, a job's status, an invoice's age. Read live, not from a snapshot that was accurate on Tuesday.
• 06
  Tacit

  The judgement calls that live in two or three senior people and have never been written down, because writing them down was never anyone's job.
• 07
  Permission

  Who is allowed to see or approve what, and under which condition — the boundary a governed assistant has to respect rather than route around.

Why this matters more this year than last

The excuse used to be the model. It genuinely was not good enough, or not cheap enough, to be worth wiring into daily work. That excuse has expired. The model your competitor rents next quarter will not be meaningfully worse than the one you rent today, and the price of both keeps falling. Once every business has access to comparable intelligence, the businesses that pull ahead are the ones whose model actually knows something. That is a data and process problem, not a model problem, and it does not get solved by switching which model you call.

What building the context layer actually involves

In practice: resolving your entities once and keeping the mapping current, capturing precedent and policy in a form a system can query rather than a person can only recall, keeping a live read on current-state instead of a nightly extract, and building the permission model in before anything else touches production data. That is what Org Brain does. It sits between your systems and anything you point an assistant or an agent at, so every answer is built on what your business actually knows, not on the model's best guess.

None of this is a weekend integration. It is also not optional if the goal is AI that does real work rather than AI that produces plausible drafts a person then has to check against the thing it should have known in the first place.

That is the actual reframe underneath all of this. A licence buys access to reasoning. A context layer buys the thing reasoning needs in order to be useful to you specifically — and unlike the model, it does not get commoditised out from under you next quarter, because nobody else has your entities, your precedent, or your organisation's judgement calls to build one from.

This is the short form. The full argument — the diagram of where the gap actually sits, and the maturity ladder for closing it — is laid out on the context page. If you want to know where your own organisation sits on it, that is what an AI Review is for.