Skip to content

Appendix D: Decision Frameworks

All decision flowcharts and comparison matrices from the Gems, consolidated for quick reference.


Connector Selection (GEM-023)

Does a prebuilt MCP connector exist?
  ├── YES → Use MCP prebuilt (3 clicks)
  └── NO → Do you control the API?
            ├── YES → Multiple agents? → YES → Build MCP server
            │                          → NO  → Governance needed?
            │                                   ├── YES → Custom Connector
            │                                   └── NO  → HTTP Node
            └── NO → OpenAPI spec available?
                     ├── YES → Custom Connector
                     └── NO  → HTTP Node

Knowledge Source Selection (GEM-008 + GEM-026)

Content in SharePoint, < 500 files?
  ├── YES → Search quality OK after format optimization?
  │          ├── YES → SharePoint Native ✅
  │          └── NO  → Need filtering or custom chunking? → Azure AI Search
  └── NO → Content in databases/APIs/multiple sources? → Azure AI Search
           Content in Fabric (warehouse/lakehouse/PBI)?  → Fabric Data Agent (GEM-028)

State Persistence (GEM-001)

Org has Dataverse license?
  ├── YES → Dataverse (production-grade)
  └── NO → Prototyping? → SharePoint List (zero licensing friction)
           Need max performance? → External HTTP API

Debug Mode (GEM-004)

Development/UAT  →  Keyword Inline (instant, in-chat)
Production       →  App Insights (silent, persistent, queryable)
Stakeholder demo →  Adaptive Card Panel (visual, polished)

Persona Adaptation (GEM-002)

Tone-only adaptation (3-5 personas)?    → Branched Instructions
Different knowledge per persona?         → Multi-Agent Routing
Query-specific adaptation needed?        → Dynamic Prompt Tool

Multi-Language (GEM-005)

IT maintains Entra profiles well?  → Graph API detection + persist
Quick deploy, no infrastructure?   → LLM auto-detect from input + persist
Regulatory language requirement?   → Explicit user choice

Form Collection (GEM-006)

1-2 fields          → Sequential Questions
3-6 fields          → Adaptive Card
3-6 + validation    → Card + sequential correction (Hybrid)
7+ fields           → Split into 2 cards

Handoff to Human (GEM-010)

Internal/enterprise   → LLM Summary (email/Teams async)
Teams-centric org     → Context Card to Teams channel
Customer-facing       → Omnichannel live transfer

Regulated Flows (GEM-027)

3-5 steps, simple consent     → Fully Manual topic
5-7 steps, user needs help    → Hybrid (deterministic + generative clarification)
7+ steps, multi-session       → State Machine with Power Automate

Proactive Messages (GEM-014)

Event-driven automation   → Platform Triggers (GA, recommended)
Simple Teams notification → PA + Teams Card
In-conversation follow-up → Bot Framework API

Cost Control (GEM-012)

Immediate savings       → Response length capping (instructions)
Understand costs        → App Insights token tracking
Hard limit (safety net) → Turn counter (30-40 max)

Testing (GEM-013)

Development      → Manual test script (20 cases)
Pre-production   → Automated keyword matching (30 cases, daily)
Production gate  → LLM-as-Judge scoring (15 cases, per deploy)