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
ScenarioMechanismWhy
Need custom instructions + scoped knowledgeDeclarative AgentNo code; configuration-only
Need to perform actions in external systemAPI PluginReal-time calls to external APIs
Need external data searchable in CopilotGraph ConnectorPre-indexed for search
Need custom instructions + external actionsDeclarative Agent + API PluginCombine both
Need custom instructions + external dataDeclarative Agent + Graph ConnectorCombine both
Decision Tree: Model Customization in Foundry
RequirementApproachCost/Complexity
Standard task, generic responses acceptableUse model as-isLowest
Standard task, domain-specific behavior neededPrompt engineeringLow
Standard task, need organization's data in responsesRAGMedium
Domain-specific output format/style consistentlyFine-tuningHigh
No existing model handles the taskCustom trainingHighest
Top 15 Exam Traps — Consolidated Quick Reference:
#TrapWrong Answer PatternCorrect Understanding
1Business terms = schema changes"Modify Dataverse schema"Business terms map natural language to fields; no schema change
2Agents = Copilots"Copilot can operate autonomously"Copilot assists; agents can act autonomously
3MCP = A2A"Use MCP for agent-to-agent"MCP = agent-to-tool; A2A = agent-to-agent
4SLMs are worse LLMs"Always use the largest model"SLMs outperform on narrow domain tasks with lower cost
5Agent flows = Power Automate"Use Power Automate for conversation orchestration"Agent flows = conversation logic; PA = backend automation
6NLP = CLU = Generative AI"Use generative AI for regulated intents"CLU for determinism in regulated scenarios
7Custom code always needed for D365 Copilot"Develop a custom plugin"Most customizations are no-code configuration
8Foundry Tools = Foundry Models"Use Foundry Tools to deploy custom model"Tools = AI services; Models = model catalog
9Well-Architected = performance only"Focus on optimizing inference latency"Five pillars: reliability, security, ops excellence, performance, experience
10Copilot for Sales/Service are standalone"Purchase Copilot for Sales separately"They extend M365 Copilot, not replace it
11Agent monitoring = APM"Use Application Insights only"Need conversational quality metrics too
12AI testing = traditional testing"Write unit tests for agent responses"Need statistical quality evaluation, not deterministic assertions
13ALM is automatic for Copilot Studio"Platform handles versioning"Explicit solution packaging required
14Platform security is sufficient"Trust built-in security"Need additional layers for prompt injection, DLP, data access
15Data residency = storage location"Data is stored in EU, so we're compliant"Processing, transit, and training data jurisdiction also matter
Alvin Varughese
Written byAlvin Varughese
Founder15 professional certifications