What is a context layer for AI? Giving AI your company knowledge

What is a context layer for AI?

An AI model learned from public text. It writes well, but it has never heard of your clients, your past projects or the way your team works.

A context layer sits between your documents and the AI. With every question, it hands the model what it needs to know about your company.

You'll also hear it called company memory or an enterprise knowledge layer.

What is RAG, in one sentence?

RAG (retrieval-augmented generation) pulls relevant passages from your documents, then gives them to the model along with the question.

It's the most common way to connect AI to internal documents. When the answer sits in a specific text, it keeps the model from making things up.

Most assistants that "read your files" work this way.

Why isn't basic RAG enough?

Basic RAG splits your documents into snippets and fetches the ones whose wording looks most like the question. For "what does procedure X say?", that works well.

Many of the questions a business owner asks span several documents. No single snippet answers them:

• Who has worked with this client before, and on which projects?
• Which projects used this method over the past few years?
• What proposal did we send to a client similar to this one?

What does Saména do with your documents?

Saména reads each document once, start to finish. It then links it to the others by client, project, method and team, in the language of your industry.

The AI then knows that a set of meeting notes, a proposal and a deliverable concern the same client and the same team. It can answer "who worked on what".

Every answer cites the document it came from, so you can check it. You have nothing to sort or rename.

The agents built on that memory are covered on the page "AI agents for business: definition, examples and how to set one up".

How does one layer serve Claude, ChatGPT and Copilot?

Through MCP (Model Context Protocol), an open standard Anthropic introduced in 2024 for connecting AI tools to data sources. Claude and ChatGPT accept MCP connectors, and so does Microsoft Copilot Studio.

Your team keeps the assistant it already knows, and finds your clients and projects there without copy-pasting.

Your agents live first in the Saména app, where they draft your documents in Word and PowerPoint. MCP also makes them available from these assistants.

Aren't the built-in connectors in ChatGPT, Claude or Copilot enough?

These connectors hook the assistant up to SharePoint, OneDrive or Google Drive. They respect each employee's permissions and find files based on the question asked.

They search your files the way they're filed today. Microsoft says as much about Copilot: reducing oversharing and cleaning up inactive sites improves answer quality.

Terms depend on the plan. Copilot only draws on your files automatically with a Microsoft 365 Copilot license, paid per user; the included Copilot Chat reads them only when attached. In ChatGPT, admin-managed sync is limited to Enterprise and Edu plans.

Each connector also serves only its own assistant. A company that uses two assistants sets them up twice.

What about platforms like Glean or Dust?

These are horizontal platforms: they connect many tools (files, messaging, CRM) and add search and agents on top.

• Glean: an enterprise search engine that respects the permissions of each connected app. It describes itself as built for enterprise, and its pricing is not published.
• Dust: your team builds its own agents. The Business plan, billed per seat, covers up to 100 people and limits the number of connectors; Enterprise is custom-priced.
• With both, someone in-house has to pick the sources, write the agents and keep them up to date.

Who sees what, and where does the data go?

Saména reads SharePoint, OneDrive or Dropbox with read-only access, in the folders you choose. Your files stay where they are, and none are kept.

What Saména retains is encrypted with a key unique to your company, and the master key is stored outside the database. Servers are in France, and the database is in the European Union.

The models come from Mistral AI, under a contract that rules out training on your data. Saména can also run on your own servers or a dedicated instance. The page "Where does our data go with enterprise AI?" has the details.

Frequently asked

What's the difference between a context layer and a document management system?

A document management system stores files so people can find them. A context layer prepares them for AI: it knows which document is about which client, project and team, and hands that to the model when a question comes in.

Do we need to reorganize our folders first?

No. Saména reads your documents as they are and links them together. You only choose which folders it reads.

Can several AI assistants share the same company knowledge?

Yes. Through MCP, the same memory serves the Saména app, and also Claude, ChatGPT and Copilot. Each team keeps the assistant it prefers.

Is RAG outdated?

No. Pulling passages is still useful for a precise question. It reaches its limits when the answer means connecting several documents, such as every project done for one client.

How much does a context layer cost?

With Saména, a one-time setup fee depends on your document volume, then a monthly subscription covers the whole company with unlimited users. We share pricing in a demo.