Copyright (c) 2026 MindMesh Academy. All rights reserved. This content is proprietary and may not be reproduced or distributed without permission.
7.2. High-Frequency Traps and Decision Trees
Decision Tree: Choosing an Agent Type
Decision Tree: M365 Copilot Extensibility
| Scenario | Mechanism | Why |
|---|---|---|
| Need custom instructions + scoped knowledge | Declarative Agent | No code; configuration-only |
| Need to perform actions in external system | API Plugin | Real-time calls to external APIs |
| Need external data searchable in Copilot | Graph Connector | Pre-indexed for search |
| Need custom instructions + external actions | Declarative Agent + API Plugin | Combine both |
| Need custom instructions + external data | Declarative Agent + Graph Connector | Combine both |
Decision Tree: Model Customization in Foundry
| Requirement | Approach | Cost/Complexity |
|---|---|---|
| Standard task, generic responses acceptable | Use model as-is | Lowest |
| Standard task, domain-specific behavior needed | Prompt engineering | Low |
| Standard task, need organization's data in responses | RAG | Medium |
| Domain-specific output format/style consistently | Fine-tuning | High |
| No existing model handles the task | Custom training | Highest |
Top 15 Exam Traps — Consolidated Quick Reference:
| # | Trap | Wrong Answer Pattern | Correct Understanding |
|---|---|---|---|
| 1 | Business terms = schema changes | "Modify Dataverse schema" | Business terms map natural language to fields; no schema change |
| 2 | Agents = Copilots | "Copilot can operate autonomously" | Copilot assists; agents can act autonomously |
| 3 | MCP = A2A | "Use MCP for agent-to-agent" | MCP = agent-to-tool; A2A = agent-to-agent |
| 4 | SLMs are worse LLMs | "Always use the largest model" | SLMs outperform on narrow domain tasks with lower cost |
| 5 | Agent flows = Power Automate | "Use Power Automate for conversation orchestration" | Agent flows = conversation logic; PA = backend automation |
| 6 | NLP = CLU = Generative AI | "Use generative AI for regulated intents" | CLU for determinism in regulated scenarios |
| 7 | Custom code always needed for D365 Copilot | "Develop a custom plugin" | Most customizations are no-code configuration |
| 8 | Foundry Tools = Foundry Models | "Use Foundry Tools to deploy custom model" | Tools = AI services; Models = model catalog |
| 9 | Well-Architected = performance only | "Focus on optimizing inference latency" | Five pillars: reliability, security, ops excellence, performance, experience |
| 10 | Copilot for Sales/Service are standalone | "Purchase Copilot for Sales separately" | They extend M365 Copilot, not replace it |
| 11 | Agent monitoring = APM | "Use Application Insights only" | Need conversational quality metrics too |
| 12 | AI testing = traditional testing | "Write unit tests for agent responses" | Need statistical quality evaluation, not deterministic assertions |
| 13 | ALM is automatic for Copilot Studio | "Platform handles versioning" | Explicit solution packaging required |
| 14 | Platform security is sufficient | "Trust built-in security" | Need additional layers for prompt injection, DLP, data access |
| 15 | Data residency = storage location | "Data is stored in EU, so we're compliant" | Processing, transit, and training data jurisdiction also matter |
Written byAlvin Varughese
Founder•15 professional certifications