Scoped Catalog Context in Microsoft Copilot and Teams
Business users live in Teams, not in the catalog. This recipe publishes a catalog-context agent into Microsoft Copilot and Teams, scoped to one domain’s glossary and tables. The scoping is the recipe: a tightly scoped agent at one enterprise reached 100% on its evaluation set and was demoed by the CEO from a phone; unscoped agents over the whole catalog rarely get close.
Two to three metered actions. Tight scoping also trims stray tool calls per question. At 0.25 ACU per metered action; confirm actual spend on the Usage page.
What you’ll use
Section titled “What you’ll use”Build it
Section titled “Build it”-
Pick the scope
One domain, its glossary, and on the order of 100 tables. Smaller scope means better answers; you can always add a second agent for another domain.
-
Clone and constrain the agent
Clone Catalog Context Search. Configure its search tools with a custom
base_signaturerestricted to your Domain, so the agent physically can’t wander into unrelated metadata. The same filter can exclude deprecated objects viaflag_types. -
Evaluate before publishing
Build an eval set from the questions the target team actually asks, and iterate on metadata until it passes.
-
Publish via Copilot Studio
Follow the Copilot Studio setup guide to connect the agent over MCP, then publish to Teams.
Gotchas
Section titled “Gotchas”- Chart rendering doesn’t work in Copilot/Teams; keep the agent’s answers textual.
- Viewer-role users see agent-building affordances in Alation they can’t use — set expectations or hide the entry point in your rollout comms.
- Evaluations involving cross-database SQL can loop when run against the Alation Analytics connector; pin the agent to the domain’s primary data source.