Enterprise Knowledge Sources: The Big Picture
Overview of connectors and Azure AI Search as knowledge sources.
Welcome to Domain 2: Where agents get smart
An agent without knowledge is just a chatbot that makes things up.
Imagine you hired a new employee but gave them zero access — no files, no intranet, no databases. They would guess at every answer. Enterprise knowledge sources are how you give your agent the company brain. There are three main pipes you can connect: Copilot connectors that index data into Microsoft Graph, Power Platform connectors that call external systems on demand, and Azure AI Search for heavyweight document search across millions of pages.
This module maps all three so you know which pipe to use when.
The three pillars of enterprise knowledge
Every knowledge source in Copilot Studio falls into one of three categories. The critical skill is knowing which to reach for — and the exam loves scenario questions that test exactly this.
| Feature | Data flow | Search type | Setup complexity | Best for |
|---|---|---|---|---|
| Copilot connectors | Crawl → index into M365 Graph → agent queries index | Full-text search over indexed content (generative answers) | Admin-driven: configure in M365 admin centre, ~50 connectors available | Surfacing external docs (ServiceNow KB, Confluence, file shares) alongside M365 content |
| Power Platform connectors | Live API call at runtime → structured JSON response | No search — direct API request/response | Developer configures connection + actions in Copilot Studio, 1,000+ connectors | Real-time lookups (CRM records, ticket status) and write-back actions |
| Azure AI Search | Push or pull indexer → dedicated Azure search index → agent queries via API | Keyword, vector, semantic, or hybrid search with ranking profiles | Developer creates Azure resource, builds index, connects from Copilot Studio | Large document collections (50K+ docs), multi-language, custom relevance tuning |
When to use which — decision factors
Choosing the right pillar depends on four factors. Get these straight and you will handle any scenario question the exam throws at you.
| Factor | Copilot connectors | Power Platform connectors | Azure AI Search |
|---|---|---|---|
| Data freshness | Near-real-time to scheduled crawl (hours) | Real-time (live API call) | Depends on indexer schedule (minutes to hours) |
| Data volume | Millions of items per connection | Per-request (paginated) | Billions of documents |
| Security model | M365 tenant boundary, Graph permissions | Connection-level auth (OAuth, API key) | Azure RBAC + index-level security trimming |
| Who configures | M365 admin (tenant-level) | Developer in Copilot Studio | Developer + Azure admin |
Data residency matters
Copilot connectors index data into your M365 tenant — the data lives in your tenant’s geo. Azure AI Search stores data in whichever Azure region you choose. Power Platform connectors do not store data at all — they pass it through at runtime. For regulated industries, this distinction drives architecture decisions. The exam may ask where data resides for a given connector type.
How data flows through the agent
When a user asks your agent a question, the orchestrator decides which knowledge source to query. This is not random — it follows a priority chain:
- Topic match first — if a topic trigger matches the user’s message, that topic runs (including any tools or flows it calls).
- Generative answers second — if no topic matches, the agent uses generative answers to search configured knowledge sources (Copilot connectors, SharePoint, uploaded files, Azure AI Search).
- Fallback last — if no knowledge source returns a confident answer, the system fallback topic fires.
Scenario: Kai maps Pacific Mutual's knowledge sources
Kai just kicked off the Copilot Studio deployment at Pacific Mutual, a 15,000-person insurance company. The project sponsor wants the agent to answer questions from three sources: the internal ServiceNow knowledge base (5,000 articles), the Salesforce CRM (live policy lookups), and a regulatory document archive (80,000 PDFs on a file share).
Kai maps the architecture: ServiceNow gets a Copilot connector — the M365 admin configures the ServiceNow Graph connector so articles are indexed and searchable through generative answers. Salesforce gets a Power Platform connector — the agent calls the Salesforce API in real time to pull live policy data (no indexing needed, the data changes constantly). Regulatory PDFs go into Azure AI Search — Kai creates an Azure AI Search index with a blob storage indexer, enables semantic ranking for natural-language queries, and connects it to the agent.
Three sources, three pillars, one agent. This is the architecture pattern the exam expects you to recognise.
Choosing the right pillar — a mental model
Think of it as a restaurant analogy:
- Copilot connectors = a buffet. The food (data) is prepared in advance and laid out for you. Fast to serve, but you eat what is available.
- Power Platform connectors = ordering from the kitchen. You get exactly what you ask for, made fresh, but it takes a moment and requires a specific order (API call).
- Azure AI Search = a private chef with a custom pantry. You stock the pantry yourself, define the recipes (index schema), and get exactly the dish you want — but you need to set it all up first.
Kai needs Pacific Mutual's agent to pull live insurance policy details from Salesforce when a customer asks about their coverage. Which knowledge pillar should he use?
An M365 admin has configured a ServiceNow Copilot connector. Where is the ServiceNow knowledge base content stored after indexing?
Which factor most strongly determines whether to use Azure AI Search over a Copilot connector?
🎬 Video coming soon
Enterprise Knowledge Sources: The Big Picture