Domain 1: GenAI Fundamentals (25-30%)
How LLMs Work in Microsoft 365
Microsoft 365 Copilot uses Large Language Models (LLMs) combined with your organization's data in the Microsoft Graph.
- Pre-training: Models are trained on massive datasets to understand language patterns.
- The Copilot System: When you send a prompt, Copilot "grounds" it by retrieving context from your emails, files, and calendar via the Microsoft Graph.
- Semantic Index: A map of your data that allows Copilot to find meaning in your files, not just keywords. (e.g., finding "budget" documents when you ask about "costs").
- Tokens: AI processes text in chunks called tokens. More tokens = higher cost/processing.
Contextual Influence
The "Grounding" process is influenced by different layers of data:
- Work Context: Data from your M365 tenant (emails, chats, files).
- Web Context: Real-time information (if "Web" toggle is on).
- App Context: The specific application you are using (e.g., Copilot in Word knows you are writing a document).
Exam Tip
Copilot prioritizes Work Context (your data) to ensure answers are relevant to your specific business role.
Custom Copilot Agents
In addition to standard Copilot, you can build Custom Agents for specific business needs.
When to build a custom agent?
- Specialized Knowledge: When you need the AI to focus on a specific, narrow dataset (e.g., a "Policy Guru" that only looks at HR manuals).
- Unique Workflows: When you need to automate multi-step business processes that standard Copilot doesn't handle natively.
- Third-Party Data: When you need the AI to access information stored outside the Microsoft 365 ecosystem (e.g., CRM or ERP systems).
- Defined Persona: When you want the AI to always act with a specific tone or role (e.g., a "Legal Compliance Reviewer").
Purpose
Custom agents are built using Copilot Studio to extend the reach of AI into your unique organizational logic and data.
Chat vs. Agent Experience
| Feature | Copilot Chat | Copilot Agent |
|---|---|---|
| Scope | Broad, conversational | Specific, task-oriented |
| Context | Uses entire M365 Graph | Uses specific, narrow data/folders |
| Workflow | Dynamic follow-ups | Structured, multi-step |
| Creation | Out-of-the-box | Built in Copilot Studio |
Copilot Across M365 Apps
| App | Key Capability | Requirement | Use Case |
|---|---|---|---|
| Excel | Data analysis & formulas | Excel Table format | Trend analysis |
| Word | Drafting & Rewrite | Text selection | Transforming notes |
| PPT | Create from File | Word/PDF source | Presentation design |
| Outlook | Summary by Copilot | Email thread | Catching up |
| Teams | Meeting Recap | Transcription/Record | Action items |
Data Protection & Privacy
Using AI doesn't mean bypassing security. Copilot respects organizational boundaries and individual permissions.
- Tenant Isolation: Your data stays within your organization's Microsoft 365 boundary and is never leaked between tenants.
- No Training: Microsoft does not use your business data to train the foundational public LLMs.
- Permissions: Copilot acts as a "user" with your exact permissions. If you don't have access to a folder, Copilot can't find it.
- Sensitivity Labels: Copilot respects labels (e.g., "Confidential") and honors the protection they provide.
- DLP (Data Loss Prevention): Files blocked by DLP policies are excluded from grounding results.
IMPORTANT
Grounding Filter: Permissions and labels act as a "mandatory filter." Copilot will never return information you aren't authorized to see.
Responsible AI & Risks
Using AI safely is a core part of the AB-730 exam.
Key Risks
- Hallucinations (Fabrications): The AI generates confident but false information.
- Bias: The AI may mirror prejudices found in its training data.
- Over-reliance: Trusting AI output without human verification.
- Prompt Injection: Maliciously trying to bypass AI safety guards with clever wording.
Critical
Human-in-the-Loop: AI output must always be reviewed by a human for accuracy and tone before being used in a final business context.
Fundamentals Quiz
Does Microsoft train public models on your business data?
(Click to reveal)