9.1 Connect to Copilot Connectors
Key Takeaways
- Copilot connectors (formerly Microsoft Graph connectors) index external enterprise content into Microsoft Graph so agents ground answers with searchable, permission-trimmed knowledge.
- Tenant administrators configure connectors in the Microsoft 365 admin center; makers in Copilot Studio only attach connectors that already exist and appear as knowledge sources.
- Copilot connectors are knowledge sources for retrieval and citations—not tools for creating tickets, updating records, or other write operations.
- Agents that use Copilot connectors require correct user authentication; publishing to channels typically needs the ExternalItem.Read.All scope in manual authentication settings.
- Quality improves when semantic labels, indexing completeness, and tenant graph grounding with semantic search are configured; missing admin setup is why connectors fail to appear in Add knowledge.
9.1 Connect to Copilot Connectors
Quick Answer: Copilot connectors (formerly Microsoft Graph connectors) index external line-of-business content into Microsoft Graph. In Copilot Studio you Add knowledge, select a connector your tenant admin already configured, and the agent retrieves permission-trimmed content to ground generative answers with citations. They are knowledge, not write tools.
Domain 2 on AB-620 opens enterprise knowledge with three patterns: Copilot connectors, Power Platform connectors, and Azure AI Search. This section owns the first pattern—how Microsoft Search / Graph-indexed external data becomes agent knowledge.
Mental model: index-then-answer
Think of a Copilot connector as a search and index pipeline, not a live API proxy:
- An administrator configures a prebuilt or custom connector in the Microsoft 365 admin center.
- Content is copied and semantically indexed into Microsoft Graph (source ACLs are honored).
- At runtime the agent issues a retrieval query against that indexed enterprise graph content.
- The model synthesizes an answer and can cite Graph items.
That architecture explains almost every exam trade-off:
| Dimension | Copilot connector behavior |
|---|---|
| Primary purpose | Ingest and index non-Microsoft content into Microsoft Graph for grounding across Microsoft 365 and custom agents |
| Where data lives after setup | Microsoft Graph semantic index (not only the source system) |
| How the agent uses it | Knowledge source—retrieve, ground, cite |
| Data movement | Yes—content is indexed into Microsoft 365 / Graph |
| Best for | Documents, knowledge bases, tickets, wikis, and other content that benefits from semantic ranking and reuse |
| Latency | Typically low for retrieval because results come from the index |
| Citations | First-class—references can appear with grounded answers |
| Who configures the source | Tenant admin first; makers attach what is already available |
Exam trap: Do not pick a Copilot connector when the scenario requires create ticket, update order, or post comment. Index-backed knowledge answers about content; tools / actions (often Power Platform connectors, custom connectors, REST, or MCP) perform transactions.
Knowledge versus tools (the AB-620 distinction)
AB-620 repeatedly tests whether you can separate grounding from action:
| Need | Correct surface | Wrong instinct |
|---|---|---|
| “Summarize the approved expense policy from Confluence” | Copilot connector knowledge (if Confluence is indexed) | Custom REST tool that reimplements search |
| “What does ServiceNow KB article KB001234 say?” | Copilot connector knowledge over indexed KB | Power Automate write path |
| “Create a high-priority incident for this outage” | Tool / action (Power Platform connector or flow) | Adding more knowledge sources |
| “List open P1 incidents and then reassign one” | Knowledge for discovery plus tool for update | Assuming Graph index alone can reassign |
Microsoft’s comparison guidance is explicit: start with Copilot connectors for knowledge discovery and grounded Q&A at scale; use Power Platform connectors when you need real-time facts without replication or transactions. Many enterprises run mixed mode—index evergreen content with Copilot connectors, and attach Power Platform tools for create/update workflows.
What makers actually do in Copilot Studio
You do not build the Graph index inside the agent designer. Maker steps, aligned with Microsoft Learn:
- Open the agent.
- Choose Add knowledge from Overview, Knowledge, or a generative answers node’s properties.
- Select the desired Copilot connector (use Advanced if it is not in the default list).
- Select the connection and Add to agent.
- Wait until the knowledge source status is ready, then test retrieval and citations.
If the connector is missing:
- Confirm the tenant admin configured it in the Microsoft 365 admin center.
- Check Advanced in Add knowledge.
- Still missing → contact admin; makers cannot invent Graph connectors from the agent alone.
Supported enterprise sources are those listed for Microsoft 365 Copilot connectors / Microsoft Search—examples in Microsoft guidance include ServiceNow, Jira, Confluence, GitHub, Azure DevOps Services, and many others. The agent can ground on the same indexed sources that augment Microsoft Search once the admin finishes setup.
Authentication, permissions, and channel publish
Copilot connector knowledge respects source-level permissions: when a specific user asks a question, the agent should surface only content that user may access. That requires correct end-user authentication, not “anonymous channel + full Graph dump.”
Microsoft Learn calls out a critical publish requirement: agents using Microsoft Copilot connectors as knowledge must include the ExternalItem.Read.All scope as part of manual authentication settings when publishing to channels. Without the right scopes, makers sometimes see “works in one surface, fails in another”—for example results in Microsoft Search experiences but empty or failing grounding from a published custom agent channel.
Also align agent authentication with tenant graph grounding features: semantic-index improvements for SharePoint and connectors often require authentication set to Authenticate with Microsoft. Wrong auth mode blocks or weakens graph-grounded quality.
Quality knobs that show up in scenario stems
| Control | Why it matters |
|---|---|
| Semantic labels on connector schema | Improves how content is understood and ranked for Copilot-style retrieval |
| Indexing completeness | Stale or partial index → incomplete answers even when source system is correct |
| Tenant graph grounding with semantic search | Microsoft recommends it for better connector search quality; may involve licensing/cost considerations |
| Microsoft 365 Copilot license presence in tenant | Documented as helpful for better connector search results in the same tenant as the agent |
| Agent instructions for citations | Helps models include in-text citations so grounded answers are not blocked when ungrounded responses are disallowed |
| Official source marking | Classic-orchestration nuance for highly trusted sources (not a substitute for correct connector choice) |
Exam trap: Low-quality answers from a Graph source are not automatically fixed by “add more HTTP tools.” First verify labels, index health, Work IQ / semantic grounding settings, and authentication scopes.
When Copilot connectors beat simpler knowledge options
| Scenario | Prefer Copilot connectors when… | Prefer something else when… |
|---|---|---|
| Large non-Microsoft knowledge bases | Content should be semantically searchable with citations across agents and Microsoft 365 | You only need one small SharePoint library already in M365 |
| Multi-system IT content | ServiceNow + Confluence + ADO work items should behave like one enterprise graph | You need live ticket priority right now with no copy → real-time Power Platform knowledge |
| Compliance and ACLs | Source ACLs must trim results per user | Anonymous public FAQ website |
| Write workflows | Never alone—pair with tools | Transaction is the primary goal |
SharePoint site knowledge and uploaded files remain valid for Microsoft 365-native content. Copilot connectors shine when the system of record is outside Microsoft 365 but the organization still wants Graph-powered grounding with permissions.
End-to-end enterprise scenario
Contoso’s service desk wants a Copilot Studio agent that answers “How do I reset VPN MFA?” using ServiceNow Knowledge and “What was decided in the last major incident review?” using Confluence. The Power Platform admin does not own Graph connector setup alone—the Microsoft 365 admin enables ServiceNow and Confluence Copilot connectors, maps schema, applies semantic labels, and completes indexing. The agent builder then Adds knowledge for both connectors, sets Authenticate with Microsoft, and includes ExternalItem.Read.All for the Teams channel. Users only see articles their ServiceNow/Confluence ACLs allow.
Later, Contoso adds “Create a ServiceNow incident from this chat.” That is a new tool via a Power Platform ServiceNow connector or agent flow—not another knowledge attachment. The agent now reads from Graph-indexed knowledge and writes through a tool. That split is pure AB-620 Domain 2 design.
Pitfalls checklist
- Connector not listed → admin indexing / Advanced / wrong tenant.
- Works in Microsoft Search but not the agent channel → authentication and
ExternalItem.Read.All. - Answers ignore ACLs expectations → user auth missing or misconfigured.
- Stale answers → index lag; not “model temperature.”
- Using connectors for writes → wrong surface; add tools.
- Confusing Copilot connectors with Power Platform connectors—same word “connector,” different control planes (Graph index vs Power Platform connection + DLP).
Success checklist
- Can explain index-then-answer vs real-time API.
- Know admin configures, maker attaches.
- Treat Copilot connectors as knowledge with citations and ACLs.
- Remember
ExternalItem.Read.Allfor channel auth. - Choose mixed mode: Graph knowledge + PP tools for transactions.
- Troubleshoot quality via labels, index, semantic grounding, and scopes—not random REST calls.
Master this leaf skill and you own the first enterprise knowledge path on the AB-620 blueprint: connect agents to enterprise graph content as grounded knowledge, without confusing it for action tooling.
What is the primary role of a Copilot connector (formerly Microsoft Graph connector) when added as knowledge in Copilot Studio?
A maker cannot see a ServiceNow Copilot connector under Add knowledge even after selecting Advanced. What should happen first?
Which authentication-related requirement does Microsoft call out for agents that use Microsoft Copilot connectors as knowledge when publishing to channels?