Microsoft's guidance is clear that AI workloads belong in ordinary application landing zones. What changes is the order of work. Some Foundry networking choices are fixed the moment the resource is created, so they have to be settled before the first model is deployed. This guide covers placement, private endpoints and DNS, agent network isolation, keyless identity, the renamed Foundry roles, policy, the AI gateway, quotas and regions, checked against Microsoft's documentation in September 2026.
Azure OpenAI in Microsoft Foundry Models has three billing shapes: pay per token, reserve capacity in provisioned throughput units, or accept a 24-hour target turnaround for half price with Batch. This guide explains where tokens come from in an agent turn, how PTU sizing and reservations work, when spillover helps, and how to monitor spend, with every claim checked against Microsoft's documentation in September 2026.
Indirect prompt injection arrives through the documents, email, web pages and tool results an AI system reads, not through the user. OWASP's 2026 Top 10 now says no reliable prevention exists and that defence is architectural. This article explains where injected instructions enter, what model-level defences can and cannot do, and the trust boundaries on data that still hold when the model is fooled.
NIST's AI Risk Management Framework is voluntary and ISO/IEC 42001 is certifiable, but they describe the same management cycle. This guide lines the two up, maps each theme to the Microsoft controls that produce evidence for it (Purview, Entra Agent ID, Agent 365, Defender for Cloud and Foundry), and is clear about what no tool can do for you. Checked against NIST, ISO and Microsoft documentation in September 2026.
Every MCP server is two things at once: text your model will read, and code that acts with someone's credentials. This guide covers the current MCP specification and its OAuth 2.1 authorization model, five risk classes, a vetting checklist, where Microsoft's MCP hosting and gateway options stand in September 2026, and an allow-list model that holds up in an audit.
Once an AI agent can write to systems, send email or move money, "a human approves it" is not a design. This guide classifies agent actions by consequence and reversibility, maps each class to an approval pattern, and sets out what the audit record must hold to survive a review, grounded in what Microsoft 365 Copilot, Copilot Studio, Microsoft Foundry and Purview actually provide as of September 2026.
Copilot Studio and Microsoft Foundry are not competitors — they sit at different levels of abstraction, and three changes in 2026 have quietly retired most of the advice written about choosing between them. A decision framework built on who owns the runtime, where the data already lives, and who operates the agent on day ninety, with the cost shapes, governance asymmetries and interop paths verified against Microsoft's own documentation.
Microsoft Foundry is no longer a place to call models — it is a platform for hosting, grounding, governing and operating agents. A working architecture guide to prompt and hosted agents, Foundry IQ, Toolbox, MCP and A2A, model choice and Claude, agent identity and RBAC, private networking, evaluation and Agent 365 — with every capability marked GA, Preview or retiring, and the places Microsoft's own documentation contradicts itself named rather than smoothed over.
Microsoft's AI portfolio is not five versions of the same product — it is a layered enterprise architecture with an experience layer, two build platforms, a control plane, and the identity, data and threat controls underneath. A practical architecture guide to what each platform is for, which workloads belong where, how agent identity and authority actually flow, and what has to be true before an organization scales any of it.