# Governed context

How copilots, BI, and MCP clients reach production data through the confirmed graph and its gate instead of raw access, and what travels with each column.

## Raw access versus governed context

Every team wants AI on its data. The usual way to give it is a connection string, and a connection string gives an agent two options: guess what the tables mean, or read whatever it can reach. One wrong statement is an incident, and nothing records why it happened.

Governed context is the other way. The AI system reads the confirmed [graph](/docs/concepts/the-graph): entities, measures, relationships, and definitions in your experts' words, with the same policy, approval, and ledger as everything else. It gets meaning instead of tables, and the gate instead of the door.

## The same gate for every AI system

Copilots, BI tools, and MCP clients reach the graph through one gate. What the gate does is the same for all of them:

- **Serves confirmed facts.** Validated definitions come through as such; inferred ones are marked as inferred or held back, as policy says.
- **Applies policy.** What a given client may read, and about which databases, is set with your team.
- **Records the access.** Every read lands in the [audit ledger](/docs/concepts/policy-approval-audit) with the client, the question, and what was served.

No second copy of production, no new pipeline. The graph is what the AI system reads, and the graph already exists because your team confirmed it.

## What travels with a column

A column served through governed context arrives with its facts attached:

- **Its definition**, in the expert's words, and the measure it belongs to when it belongs to one.
- **Its personal-data flag**, confirmed or to review, so a system downstream inherits the flag rather than rediscovering it.
- **Its lineage**, where the value comes from and which exports and dashboards already use it.
- **Its freshness**, when the source last changed.

An illustrative answer to "revenue by region, last quarter" is grounded in the confirmed definition of revenue, the confirmed grain of region, and the confirmed period, rather than in whichever `amount` column the system found first.

## Your model, our context

The graph does not care which model reads it. Bring the language model your team chooses, in the tooling your team already uses; Datapace supplies the context and the gate. Changing the model changes nothing about what is confirmed, what is allowed, or what is recorded.