> Site index: [llms.txt](https://ortelian.com/llms.txt) with all pages and descriptions. # The company context layer: one approved source for every AI tool Source: https://ortelian.com/notes/the-company-context-layer/ Markdown: https://ortelian.com/notes/the-company-context-layer.md Author: Laurens Nys Published: 2026-10-20 A company context layer is one approved, maintained source of definitions, documents and permissions that ChatGPT, Copilot, Claude and your agents all read from. A company context layer is one maintained, approved source of what the company knows, which every AI tool the company already uses can read from. ChatGPT, Copilot, Claude and the agents your team builds all answer from the same definitions, the same current documents and the same permissions, instead of each tool guessing from whatever it can reach. Most companies did not plan their AI tools. They adopted them fast and switched on every connector available. That gets people using AI. It also leaves three problems that show up a few months later. ## The problem: many tools, no shared source When each AI tool connects to the wiki, the drive and the CRM on its own, three things happen. Answers can't be trusted. The wiki holds official pages next to drafts, archived pages and work in progress. The tool can't tell them apart, so neither can the person reading the answer. Teams start asking whether an answer is correct, and nobody can say. Costs grow. Every tool searches everything for every question, and the same documents are read again and again by different tools. Permissions drift. Each connector has its own view of who may see what. Nobody has one place to check what an AI tool was allowed to read when it gave an answer. ## What goes into a company context layer The layer holds what an AI tool needs to answer like someone who knows the company. In practice that is five things. Approved sources. Which pages, files and records count as official, and which are drafts or archived. Only approved material is served as fact. Definitions. What the company means by its own words: a customer, an active account, a quote, recurring revenue, a late order. If your data team already defined these in a semantic layer for reporting, the context layer reuses those definitions instead of writing them twice. Relationships. Which customer belongs to which partner, which order belongs to which customer, which person owns which process. This is the part a [context graph](https://ortelian.com/notes/what-is-a-context-graph/) captures. How that differs from a knowledge graph and an ontology is set out in [context graph vs knowledge graph vs ontology](https://ortelian.com/notes/context-graph-vs-knowledge-graph-vs-ontology/). Versions. When a definition or a policy changed, and which version applied when. An answer about last quarter should use last quarter's rules. Permissions. Who may see what, enforced in one place, so every tool reading the layer gets the same answer to that question. ## How the AI tools you already use read from it You don't need to replace ChatGPT, Copilot or Claude. Each of them can call outside tools and knowledge sources through connectors. The Model Context Protocol (MCP) is an open standard for these connections: [its documentation](https://modelcontextprotocol.io/docs/getting-started/intro) names Claude and ChatGPT as apps that use it, and Microsoft documents how [Copilot Studio agents connect to MCP servers](https://learn.microsoft.com/en-us/microsoft-copilot-studio/agent-extend-action-mcp). The context layer sits behind those connectors. When someone asks a question, the tool asks the layer, and the layer returns approved material with its source. The same layer can serve the agents that do work, not only the assistants that answer questions. An agent preparing a quote reads the same definitions and the same approved price rules as the person checking its work. That is the difference between automation that follows a fixed script and an agent that uses judgement within limits: both need the same map of the business. When agents also act on the business, they need rules and allowed actions on top of what they read. That is what a [company OS for AI agents](https://ortelian.com/notes/company-os-for-ai-agents/) adds. ## The hard part is keeping it current Building a context layer once is not the difficult part. Keeping it right is. Definitions change, pages go stale and new systems are added. A layer that nobody maintains becomes one more source the AI tools can't trust. Three practices keep it current. Give each source and definition an owner who approves changes. Build the layer from pipelines you can inspect, so you can see where every fact came from and when it was last refreshed. And check answers against real questions on a schedule, so drift shows up before people notice it. ## Where to start Start with one piece of work where wrong answers are costly, not with the whole wiki. List the sources that work actually depends on, mark which are approved, write down the definitions people argue about and connect only those. Then point one AI tool at the layer and compare its answers with the ones people get today. Grow it from there, one piece of work at a time, as in [how to build a world model for your company](https://ortelian.com/notes/how-to-build-a-world-model-for-your-company/). ## How Ortelian does this Ortelian builds and runs this layer as part of an engagement: a platform plus a forward deployed team. We start with an audit of the work, agree which sources are approved and build the world model and pipelines your AI tools read from. ChatGPT Enterprise and Copilot connect to it inside the engagement. The platform is hosted in the EU, and your company owns its data and its world model and can export both. Where each kind of data is stored is set out in the [Trust Center](https://ortelian.com/trust/).