Agent SDK vs Logic Apps + Email Adapter + Semantic Kernel for Enterprise Outlook Email Automation

Hi,

I’m designing an enterprise-grade email automation system and would appreciate guidance on the recommended architecture.

Scenario

• The client uses an Outlook mailbox on a custom domain (Microsoft 365, separate tenant)
• The client does not have their own Azure subscription
• Our solution is hosted in our Azure tenant
• We need to process incoming emails and generate intelligent, thread-aware replies

Current Proposed Architecture

1. Logic Apps (in our tenant)
   • Trigger on incoming emails (Outlook connector / Graph API)
   • Extract:
   • Email body
   • Sender details
   • Conversation/thread ID
   • Prior thread context (if possible)
2. Email Adapter Layer (SaaS abstraction)
   • Normalize payload (HTML → structured content, attachments, inline images)
3. Semantic Kernel (SK) Agent (SaaS AI layer)
   • Perform:
   • Intent classification
   • Context-aware reasoning
   • Reply generation
   • Memory handling (thread continuity)

Response Handling

• Send reply via Outlook connector / Graph API
• Maintain conversation continuity

Key Questions

1. Architecture Choice Is this pattern (Logic Apps + Email Adapter + SK Agent) a valid and recommended enterprise production architecture, or should we instead build directly using the Microsoft Agent SDK for an end-to-end email agent?
2. Separation of Responsibilities Is it correct to treat:
   • Logic Apps as orchestration + ingestion layer
   • Semantic Kernel as the core AI agent engine
3. Agent SDK vs SK-Based Approach What are the trade-offs between:
   • A pure Agent SDK implementation
   • A composed architecture (Logic Apps + SK)
   Specifically in terms of:
   • scalability
   • observability
   • maintainability
   • production readiness
4. Cross-Tenant Outlook Integration For accessing a client’s Outlook mailbox (different tenant), what is the recommended approach?
   • Microsoft Graph API with delegated permissions?
   • Application permissions + admin consent?
   • Any best practices for secure, long-term enterprise integration?
5. Production Readiness In real-world enterprise scenarios, is Semantic Kernel considered a strong foundation for building production-grade AI email workflows, or does the Agent SDK provide significant advantages that justify adopting it instead?

Goal

To ensure the solution is:

enterprise-grade

scalable and maintainable

aligned with Microsoft best practices

Thanks in advance for your guidance!