Topical cluster · for IT directors, principal architects, and senior engineers shaping AI systems
AI Architecture
AI architecture for Microsoft business apps means three things at once: the model layer (multi-provider, with cost and quality tradeoffs), the integration layer (Logic Apps, MCP, APIM, Foundry agents), and the data layer (Dataverse, AI Search, vector backends).
Articles
17
This cluster covers all three with a practitioner bias: real diagrams, real failure modes, real cost numbers. Logic Apps as MCP servers in production. Foundry agent patterns that scale beyond hello-world. Agentic development setups that ship.
If you are designing the AI fabric for a Microsoft-heavy enterprise, this is your starting point.
Featured
What this cluster covers
Subtopics in ai architecture
- Azure AI landing zone reference architectures
- Microsoft Foundry vs Azure OpenAI decision logic
- MCP server patterns for Logic Apps and APIM
- RAG over enterprise data (Dataverse, SharePoint, AI Search)
- Multi-provider routing (Azure OpenAI, Claude, Foundry partner models)
- Agentic development setups that survive production
- Architecture diagrams that pass review
More in this cluster (16)
Common questions
AI Architecture FAQ
The questions that come up most often in ai architecture engagements. Answers grounded in Microsoft documentation and field experience.
Do I need a separate Azure AI landing zone?
No. Microsoft's Cloud Adoption Framework explicitly recommends deploying AI workloads into existing application landing zones, not a new top-level AI hierarchy. What changes inside the application landing zone is the networking (six private DNS zones, /22 spoke address space), identity (system-assigned managed identity, disable local key auth), and data plane (Purview, Defender CSPM, Sentinel hooks).
Should I migrate from Azure OpenAI to Microsoft Foundry?
Only if you need a Foundry-only feature: agents, non-OpenAI models, multi-agent workflows, Foundry IQ knowledge layer, or built-in tracing. The upgrade is opt-in and reversible. The Assistants API is the only hard deadline (retires August 26, 2026). For stable single-model GPT chat workloads with PTU reservations, defer the upgrade.
What is the right pattern for connecting Logic Apps to AI agents?
Logic Apps as MCP servers, fronted by Azure API Management for token gateway and rate limiting. The agent calls APIM, APIM authenticates and routes to the Logic App MCP endpoint, the Logic App orchestrates the underlying systems (Dataverse, ServiceNow, Salesforce). This isolates the agent from system-of-record idiosyncrasies.
When should I use a vector database vs Azure AI Search for RAG?
Use Azure AI Search for document-grounded RAG with enterprise Microsoft data (SharePoint, Dataverse, OneLake). Use a dedicated vector database (Pinecone, Qdrant, Cosmos DB for MongoDB vCore) when you need cross-region replication, sub-50ms latency, or vectors generated from non-text data. For most Microsoft-shop AI architectures, Azure AI Search is the default.
Other clusters
AI Cost & ROI
For ctos.
AI Governance & Risk
For it directors.
AI Strategy & Readiness
For ctos.
Microsoft Copilot Governance
For it directors.
Copilot Adoption & Recovery
For it directors and execs running copilot programs that are not converting to value.
Power Platform Governance & Operations
For power platform admins.
Architecture Documentation & Diagramming
For solution architects and senior engineers who own architecture documentation.