Azure AI Foundry

Scope

Azure AI Foundry (AIServices), running in a Microsoft Azure region. One of the three targets in ADR-0011. Compare all three targets on the Deployment targets hub.

This is the target this repository has actually deployed. The environment described in the as-built record was built from the Bicep in infra/, and the control plane plus all three data-plane paths were verified with real calls rather than described.

It is also the only one of the three that generates images or synthesizes speech, and the only one with access to proprietary frontier reasoning models.

The pages on this site are the source of truth

The nine pages in this section are maps, not copies. Azure AI Foundry is documented in full under Guide, Architecture, Models, and Implementation, and those documents stay canonical. These pages exist so the three-way comparison has somewhere symmetric to point.

Topic	This target	Canonical documents
Architecture	One shared Foundry account, subscription-scope Bicep, CAF naming	Design section
Models	Registry-driven, 22 entries marked deployed, 20 live	Model catalog
Features	Image, voice, reasoning, agents, content safety, quotas	Guide section
Deployment	Parameterized Bicep, tested runbook	Deployment guide
Consumption	OpenAI-compatible v1 API	Using your deployment
Cost	Variable, with an enforceable budget cap	Cost and governance
Security	Managed identity, Entra groups, Key Vault by name	Identity and security
Operations	Deployed observability package with alerts and dashboards	Observability architecture

When this is the right target

Images, speech, frontier reasoning quality, a production SLA, the shortest path to a first working call, or a low and bursty volume where variable cost beats a fixed fee. The full reasoning is in Choosing a target.

When it is not

Your data cannot leave your building. That is the one condition that rules this target out outright, and it is what the other two exist for.