Atlan

The context your documents do not contain

Retrieval over documents works well when the answer is written down somewhere. A policy, a runbook, a support article: somebody wrote a paragraph, and a search can find it.

Now ask a different question. "What was revenue in the Nordics last quarter?" The answer is not written anywhere. It lives in a warehouse table, and producing it means knowing which table, which column means revenue, which rows count as the Nordics, whether that table is the certified one or somebody's abandoned copy, and whether this user may see it at all.

None of that is in the data. It is metadata, and it is the context an assistant needs before it can touch structured data safely.

What a catalog holds that rows do not

A metadata catalog records what the warehouse itself does not.

Meaning. What a column represents, in the business's own words. That rev_net_usd is net revenue after returns, in dollars, excluding intercompany transfers.

Provenance. Lineage across pipelines, BI tools and warehouses, so you can see where a number came from and what feeds it.

Trust. Which assets are certified, which are deprecated, and who owns them.

Access. Who may see what, which is a governance fact rather than a technical one.

Atlan describes itself as "the context layer for enterprise AI", supplying "the data graph, business logic, and governance to act on trusted data", and connects to 80+ enterprise systems to unify that metadata into one graph.

The phrase is worth noticing. It is the same term a context layer defines, and the second product in recent memory to adopt it, after RAGFlow. When two vendors independently name the same component, the pattern is real.

Certification, and why a human is in the loop

The pipeline Atlan describes has four stages: unify the systems, bootstrap context with agents that auto-generate descriptions and metrics, have humans collaborate to certify it, and activate the certified context downstream.

The third stage is the one that distinguishes this from retrieval over documents.

Auto-generated descriptions are guesses. A model reading a schema can propose that rev_net_usd is net revenue, and it can be wrong in ways nobody notices until a number reaches a board deck. Certification is a domain expert confirming the guess before an agent is allowed to rely on it.

That gives you something document retrieval does not: a trust boundary inside your context. An agent can be restricted to certified assets, so the blast radius of a wrong description is bounded by review rather than by hope.

Their framing of the argument is the clearest line on the site: "context doesn't come from a prompt. It comes from a pipeline."

Reaching the catalog from an agent

Atlan advertises MCP server support, so an agent queries certified context through the Model Context Protocol rather than a bespoke integration.

That maps onto the primitives cleanly. An MCP server exposes resources, which are application-controlled data attached for context, and tools, which are model-controlled. A catalog is a natural fit for both: certified definitions as resources the application attaches, and lookup as a tool the model can call when it needs to find the right asset.

The permission question comes with it. A catalog knows who may see what, and an agent that ignores those records is a faster route to a data leak than any query a person would write. The access rules have to be enforced where the query runs, not in the prompt that asks the agent to be careful. Context security covers the general case.

Send metadata, not rows

The instinct with structured data is to put the data in the prompt. For most questions that is the wrong payload.

Rows are large, they blow through the context budget, and they carry exactly the sensitive values you would rather not send anywhere. Metadata is small, stable, and enough: given the right table, the meaning of its columns, and the definition of a business term, a model can write the query, and your system runs it under the user's own permissions.

That inverts the usual flow. Instead of retrieving data and asking for an answer, you retrieve meaning and ask for a query. The data stays in the warehouse, the access control stays where it already works, and the model never sees rows it should not.

Further reading

• Atlan, the product site: the context-layer positioning, the four-stage pipeline, and MCP support.
• What is a context layer: the pattern this product claims to be.
• MCP server: the primitives an agent would reach it through.