Context windows and knowledge management in a company
How growing language model capabilities affect how companies can organise access to corporate knowledge.
Over the past few months, the context window size of language models has grown to levels that change the practice of using them. Models can now process hundreds of pages of text in a single call.
For most executives this sounds like a technical detail. But it carries a concrete practical consequence: architectural approaches that a year ago were the only sensible way to embed AI into work with corporate knowledge are no longer the only option.
This is a good moment to rethink how a company thinks about access to its accumulated knowledge and documentation.
What AI in document work looked like before
Until recently, the standard approach to working with corporate documents through AI looked roughly like this: take a large body of documentation, split it into small chunks, convert those into vector representations, build a similarity search system, and feed the model only the relevant fragments.
This works, but the approach has limits. Vector similarity search does not always find the right fragment - especially when a question requires combining information from several places. The context the model sees is incomplete. The answer is only as good as the retrieval system.
What a large context changes
When a model can accept several hundred pages in a single call, a different class of tasks becomes accessible.
For example: analysing an entire contract package for risks, contradictions, and non-standard clauses. Previously this required either splitting the document into pieces and losing the connections between them, or expensive custom development. Now it is an accessible pilot.
Or: comparing two versions of a technical specification to identify substantive changes. Or: analysing a project's correspondence over several months to surface key decisions and open questions.
This does not mean all tasks are now solved with a single prompt. A large context is a capability, not a quality guarantee. The model can miss important things if the document is poorly structured or the question is imprecisely worded.
How this changes the conversation about knowledge management
Most companies that come to me with "AI for document work" have a deeper problem: the documents themselves are not in order.
Regulations are updated but old versions remain alongside. Decisions are made in chats and never recorded. Important documents live in personal folders or in systems with no convenient access.
A large context does not solve this problem. The model can read a thousand pages, but if those pages include conflicting versions of the same regulation, it does not know which one is correct.
The question of how to organise the storage and updating of corporate knowledge was relevant before AI and remains relevant after. AI simply makes the cost of disorganisation more visible.
A practical question for a manager
If a company is thinking about making corporate knowledge more accessible - through AI or otherwise - it is useful to start with a few questions:
- What specifically is lost or slow because people cannot find the information they need?
- In which system or systems does most of the operational knowledge live?
- Who is responsible for keeping key documents current?
- If a new employee joins tomorrow - where would they look to understand how a given process works?
The answers determine what kind of solution makes sense - simple or complex, with AI or without. The tools have changed, but the questions have stayed the same.