Why AI Assistants Fail Without the Right Data

▶ Watch (00:42)

Microsoft 365 Copilot is built to act as a personal assistant for work. That only works when it has access to the data people actually use. Copilot pulls answers from enterprise data inside the Microsoft Graph, but also from external sources a team explicitly connects.

“your AI strategy is anchored” — Rachit Malik

Most organizations keep critical data outside Microsoft 365: ticket systems, internal wikis, custom file shares. Until that data reaches Copilot’s semantic index, it cannot appear in answers. The connector layer is how it gets there.

How Copilot Connectors Feed the Semantic Index

▶ Watch (02:54)

Copilot connectors use the Microsoft Graph API to ingest items from outside systems: document repositories, knowledge bases, file systems, content management platforms. Each item gets written into an advanced lexical and semantic index. That index is what Copilot searches when it generates an answer.

The indexed items inherit existing security and privacy policies, so access controls from the source system carry over. Good candidates are any data source with a REST API that holds information people currently have to search manually.

Scaffolding and Running a Connector in VS Code

▶ Watch (04:32)

The M365 Agents Toolkit VS Code extension scaffolds a working connector in under a minute. Pick “New Copilot Connector,” enter a name and connection ID, and the project is created. The default template targets the GitHub Issues API. Every function that needs replacing for a different data source is tagged with a “customization point” comment.

“We do all the hard work for you.” — Rachit Malik

Watch the connector scaffold and full-crawl demo (04:32): Entra app provisioned, schema registered in under 10 minutes, 27 issues crawled from GitHub, then queried live in Copilot chat with cited external-item links.

Wiring a Declarative Agent to Connector Data

▶ Watch (13:09)

A connector alone adds data to Copilot chat. An agent scoped to that connector narrows it further. The Agents Toolkit lets you add a “Copilot Connector” capability to a declarative agent by selecting an existing connector from a list of all tenant connections. No JSON editing required.

See the declarative agent wired to connector data (13:23): MyGithubAgent provisions in seconds, then answers the same GitHub issues query from within the agent context. The debugger confirms the connection ID matches the one ingested earlier.

Notable Quotes

your AI strategy is anchored Rachit Malik · ▶ 1:29

We do all the hard work for you. Rachit Malik · ▶ 6:48

It is super easy to get started. Rachit Malik · ▶ 5:50

Key Takeaways

  • The M365 Agents Toolkit scaffolds a custom Copilot connector from scratch in under five minutes, with no Graph API boilerplate to write.
  • Customization points in the generated code mark exactly which functions to replace when swapping GitHub issues for any line-of-business API.
  • A declarative agent scoped to a connector answers from that data only, and the debugger shows the connection ID to confirm the source.