Connecting External Data to Copilot

▶ Watch (01:42)

Microsoft 365 Copilot searches emails, chats, documents, and Dataverse by default. Orders, tickets, and live-site data rarely live there. Copilot connectors close that gap: 65 pre-built integrations cover ServiceNow, Jira, Confluence, PagerDuty, GONG, and others. Each indexed item becomes semantically queryable alongside native M365 content.

“Copilot is the UI for AI.” — Sebastien Levert

Build announcements: no connection or item-count limits, a quick-setup flow that completes in under a minute, knowledge scoping so agents can target a specific project or area path, and Copilot connectors now fully integrated into Copilot Search.

Setting Up Connectors and Building Agents Without Code

▶ Watch (07:43)

Rachit Malik ran the ServiceNow Admin Center connector setup (07:43) live. In Settings > Search and Intelligence, he added a data source, pasted a client ID and OAuth secret, clicked Authorize, and hit Create. Under a minute. Copilot immediately answered “What are my travel policies?” by pulling indexed ServiceNow articles with working citations.

From there, Sebastien built a helpdesk declarative agent in Agent Builder: paste a system prompt, select the ServiceNow connector as the knowledge source, add starter prompts, click Create. The finished agent summarized articles by topic and returned a four-step expense-report procedure, all cited back to the connector.

What Connector Builders Should Get Right

▶ Watch (17:30)

GONG’s Eran Aloni built the connector after noticing that salespeople spend most of their time in Teams, Outlook, and Office, not in GONG. The connector surfaces deal summaries, committed action items, and objections directly in Copilot, so reps get a Monday-morning briefing without switching apps. GONG has close to 5,000 B2B customers and began applying AI to sales conversations in 2016.

“And that was, again, 10 years ago before AI was hip or cool.” — Eran Aloni

He named two rules for any connector builder: index the data your users need when they are not inside your app, and mirror your app’s permission model so Copilot only surfaces what the querying user is authorized to see.

Scaffolding a Custom Connector from Code

▶ Watch (28:53)

When pre-built connectors don’t cover a data source, the Microsoft 365 Agents Toolkit generates the scaffolding. Rachit ran the Repair CRM custom connector live demo (30:00) in VS Code. The flow: scaffold a TypeScript Azure Function project, define schema.json with semantic labels (queryable, searchable, retrievable), write a mapping function from the REST API response, hit F5. The toolkit handles Graph API registration, schema upload, and item ingestion automatically. Within five minutes, 12 repair records appeared and Copilot returned them with citations.

Acting on Connector Data with TypeSpec and Declarative Agents

▶ Watch (39:23)

Reading data is only half the loop. Three action improvements shipped at Build: instruction-following accuracy so Copilot executes multi-step workflows written in plain language, bulk API calls so a single prompt triggers chained requests, and multi-turn context retention so a second message can reference items fetched earlier. MCP support for declarative agent actions entered private preview. TypeSpec replaced manual JSON manifest editing and generates OpenAPI specs automatically.

“Now, you can literally describe an almost deterministic workflow” — Sebastien Levert

Rachit closed with the TypeSpec repair-assignment action demo (46:12): the agent pulled a repair ID from prior context, called the update endpoint, and confirmed a 200 response in the debug panel.

Notable Quotes

Copilot is the UI for AI. Sebastien Levert · ▶ 3:59

But there’s no more limits. Sebastien Levert · ▶ 11:23

And that was, again, 10 years ago before AI was hip or cool. Eran Aloni · ▶ 14:54

Now, you can literally describe an almost deterministic workflow Sebastien Levert · ▶ 39:31

Key Takeaways

  • 65 pre-built connectors and a no-limits indexing policy let any org surface external data in Copilot today.
  • The Agents Toolkit scaffolds a working connector in VS Code in under five minutes, handling schema, API mapping, and ingestion.
  • TypeSpec lets developers define declarative agent actions in a typed language, replacing manual JSON editing and auto-generating OpenAPI specs.