Context windows keep getting bigger. A million tokens is here. Ten million does not sound unrealistic any more.
What interests me is not that a model can read more documents. It is that an AI could hold a meaningful part of a company in mind at once, and reason across it.
The problem is that most companies do not have ten million tokens of useful context. They have ten million tokens of noise: email, chat, meeting transcripts, PDFs, CRM records, old wiki pages, spreadsheets, decks. And a large share of what actually matters is still in people, unwritten.
The employee who leaves
Take someone who has run Meta Ads for a company for five years. They know things that are nowhere in the ad account. That one product usually performs badly in December. That pushing spend too fast caused a problem two years ago. Why a campaign structure was abandoned. Which founder preference quietly overrides the normal rule.
Then they leave. The campaigns stay. The decision logic walks out with them.
The handover is the most expensive part, and it is usually done by having the new person start a month early so the old one can talk. That is the company admitting out loud that the knowledge was never written down.
So the context worth capturing is not only what happened. Why did we do it. What did we learn. What did we reject, and why. Which rule has an exception. Who is allowed to decide. What should happen when conditions change.
More context is not automatically better
The obvious reaction to a huge context window is to give the model everything. I think that is wrong.
A purchasing agent does not need every HR discussion. A marketing agent should not be looking at confidential finance data. And a sentence someone said in a meeting six months ago should not carry the same weight as an approved strategy.
So the question moves away from storage and toward architecture. What is true. What is historical. What is an assumption. What is a decision. What is a signal. What is this agent allowed to see. What is relevant to the task running right now.
From writing prompts to assembling context
Today a lot of energy goes into prompts. I think the bigger part of enterprise AI will be building the context before the prompt starts.
If an agent is asked to review purchasing, it should already know the purchasing process, current inventory, historical demand, supplier lead times, the management decisions that apply, the signals worth watching, and the rules it may operate under. The user should not be pasting any of that into a chat window.
Something like a context compiler. A task arrives, the company assembles the relevant part of itself, then the model starts thinking.
Why the size matters at all
The value of a bigger window may not be longer prompts. It may be that AI can finally work across enough connected company knowledge to see consequences.
A marketing decision affects purchasing. Purchasing affects cash. Cash affects growth. A product launch affects support, inventory, content and logistics. Right now those live in separate systems and separate heads. With enough structured context they can sit inside one reasoning process.
So the question I keep coming back to is not what a model can do with ten million tokens. It is what a company would have to become for ten million tokens to be worth reading. That feels like the bigger piece of work, and it is work only the company can do.
