Microsoft Foundry (formerly Azure AI Foundry)
by Microsoft
Microsoft's unified factory for building, grounding and governing AI apps and agents
Microsoft Foundry, renamed from Azure AI Foundry at Ignite 2025, is Microsoft's unified platform for building, grounding, evaluating and governing AI applications and agents. It gives enterprise developers one Azure resource, one identity boundary and one catalogue of more than 10,000 models, replacing the separate Azure AI Studio, Azure OpenAI and Azure AI Services surfaces they previously had to stitch together by hand.
Microsoft Foundry is Microsoft's enterprise platform for AI applications and agents, and is the product previously branded Azure AI Studio and then Azure AI Foundry; Microsoft dropped 'Azure' from the name at Ignite on 18 November 2025 to position agents alongside Microsoft 365 and Fabric rather than as one more Azure service. The rebrand carried a real structural change: what used to be a hub plus an Azure OpenAI resource plus Azure AI Services, addressed through five or more endpoints and SDK packages, is now a single Foundry resource with projects underneath it, one unified project client and one endpoint, governed by Microsoft Entra identity, role-based access control, network isolation and Azure Policy. The components are Foundry Models, a catalogue Microsoft states exceeds 10,000 models from Microsoft, OpenAI, Anthropic, Meta, Google, xAI and Hugging Face; Foundry Agent Service; Foundry Tools, the rebranded Azure AI Services covering OCR, translation and speech; Foundry IQ, a permission-aware retrieval engine grounded in Microsoft Graph; Foundry Control Plane for governance and observability; and Foundry Local for on-device and edge inference. Agents sit on a spectrum from declarative prompt agents that Microsoft hosts outright, to hosted agents where you containerise your own code written against Microsoft Agent Framework, LangGraph or Semantic Kernel and Foundry supplies the managed endpoint, scaling, identity and OpenTelemetry tracing. SDKs cover Python, C#, JavaScript and Java, alongside the portal, the Azure Developer CLI and a VS Code extension. Microsoft says more than 10,000 organisations have used the platform, naming AT&T, Audi, Commerzbank, Bayer, Carvana, Fujitsu, KPMG and Accenture; AT&T reported processing approximately one trillion tokens on it while developing OTel2.0. It competes head-on with AWS Bedrock, Google Vertex AI and IBM watsonx, and independent comparisons consistently conclude it wins where the buyer has already standardised on Microsoft identity and governance rather than on raw model price.
The platform or AI engineering team inside a Microsoft-committed enterprise. If Entra ID, Purview and Azure Policy already define your identity and data boundary, Foundry is the only model platform that inherits all three without new plumbing.
One governed Azure resource where 10,000+ models and your production agents share the same RBAC, network isolation, content filters and OpenTelemetry traces instead of each application inventing its own.
At a Glance
- Category
- Enterprise Platform
- Pricing
- Usage-based, Contact for pricing
- Target Market
- CIOs, CTOs, Enterprise Developers, MLOps Engineers, Data Scientists
- Deployment
- Cloud-first, Hybrid, Edge-first
- Founded
- 1975
- Headquarters
- Redmond, United States
- Team Size
- 500+
- Customers
- More than 10,000 organisations (Microsoft, 2026)
Key Features
- ✓Foundry Models catalogue
One catalogue of more than 10,000 models from Microsoft, OpenAI, Anthropic, Meta, Google and Hugging Face, all deployable behind a single project endpoint.
- ✓Foundry Agent Service
Runs declarative prompt agents that Microsoft hosts, or containerised hosted agents you write against Microsoft Agent Framework, LangGraph or Semantic Kernel.
- ✓Foundry Control Plane
Centralises Entra identity, role-based access control, content filters, network isolation and Azure Policy across every agent, model and tool in the tenant.
- ✓Foundry IQ
Permission-aware retrieval engine that grounds agents in Microsoft Graph content while honouring each user's existing document permissions rather than a service account's.
- ✓Observability and evaluation
OpenTelemetry-based tracing with built-in evaluators and dashboards, so agent behaviour can be monitored and regression-tested against real production traffic.
- ✓Foundry Local
Runs models on-device and at the edge for latency-sensitive, sovereign or disconnected scenarios, including Azure Local Disconnected deployments.
- ✓Unified SDK and developer surfaces
A single azure-ai-projects client for Python, C#, JavaScript and Java, plus the portal, Azure Developer CLI, VS Code extension and an MCP server.
Capabilities
Use Cases
- •Governed enterprise agent rollout
A bank ships a customer-service agent that inherits Entra identity and Azure data policies instead of re-implementing access control inside every application.
- •Multi-model evaluation and routing
A platform team benchmarks GPT, Claude and Llama variants on the same prompts, then routes production traffic to whichever wins on cost and latency.
- •Permission-aware internal search
An enterprise grounds an assistant in SharePoint and Teams content so each employee only ever retrieves documents they are already entitled to open.
- •Azure OpenAI migration
An existing Azure OpenAI resource upgrades in place to a Foundry resource, keeping its endpoint, API keys and state while gaining unified governance.
- •Trillion-token document processing
AT&T reported processing approximately one trillion tokens on Foundry while developing OTel2.0, a scale no ad-hoc single-model API deployment would govern safely.
Ideal For
Best For
- ✓Deploying agents that must run inside an existing Entra ID and Azure Policy boundary without standing up a separate identity system
- ✓Comparing and switching between OpenAI, Anthropic, Meta and open-weight models behind one project endpoint and one Azure bill
- ✓Permission-aware retrieval over Microsoft 365 content, where the RAG layer must honour each individual user's Graph permissions
- ✓Migrating an existing Azure OpenAI deployment onto a governed platform in place, preserving the endpoint, API keys and state
- ✓Regulated workloads that need customer-managed encryption keys, EU/US data-zone processing and on-device inference via Foundry Local
Not Ideal For
- ✗Teams that only want cheap GPT calls: independent analysis by nOps puts a fully networked Foundry deployment 40-80% above OpenAI's own platform once private endpoints, Azure AI Search and storage are counted
- ✗AWS- or GCP-native shops, where Bedrock or Vertex AI avoid egress charges and sit natively next to S3 or BigQuery data that would otherwise cross clouds
- ✗Small teams with no dedicated Azure administrator - PeerSpot reviewers repeatedly flag RBAC and Azure Policy setup as costly at scale, with some support cases taking three to six months to resolve
- ✗Prompt-engineering-led teams who need best-in-class evaluation tooling; reviewers rate its prompt testing and eval tools below dedicated tools such as LangSmith
Integrations
Deployment
Market & Ratings
More than 10,000 organisations (Microsoft, 2026)
Market Analysis
Pros
- ✓Single Azure resource with unified RBAC, networking and policy replaces five or more previously separate endpoints and SDK packages
- ✓10,000+ model catalogue with one-click deployment; PeerSpot reviewers cite model access and deployment speed as the top strength
- ✓Strong governance story - Entra identity, content filters, network isolation, customer-managed keys and documented no-training commitments
- ✓Data-zone deployment types keep processing inside a designated US or EU geography, with data at rest always in the customer-designated geography
Cons
- ✗Pricing opacity is the recurring complaint: reviewers describe the fragmented model as hard to budget, and nOps documents deployments landing 40-80% above the raw token estimate
- ✗Evaluation runs silently multiply token spend - one widely cited example is 4M tokens consumed on a 20-question eval
- ✗Setup complexity: PeerSpot reviewers say role-based access and Azure Policy configuration becomes costly at scale, and multi-region model deployment is awkward
- ✗Documentation and support are weak spots - unclear docs requiring extensive searching, and support cases reported taking three to six months
- ✗Prompt testing and evaluation tooling is rated below dedicated competitors such as LangSmith
- ✗Constant rebranding (Azure AI Studio to Azure AI Foundry to Microsoft Foundry in two years) has left stale docs, two portals and a classic/new split to navigate
Pricing
Platform access
$0
- ✓Foundry portal, projects and model catalogue are free to use and explore
- ✓You are billed only for the models, tools and Azure services actually consumed
Pay-as-you-go consumption
From $0.20 per 1M input tokens
- ✓Per-token inference priced per model and region (DeepSeek-V4 Flash $0.20 in / $0.80 out; GPT-4o $2.50 in / $10.00 out per 1M tokens)
- ✓Foundry Tools, Foundry IQ (Azure AI Search) and Azure ML billed separately at their own rates
Agent Commit Units pre-purchase
Contact for pricing
- ✓One-year commitment terms
- ✓5% discount at 20,000 ACUs, 10% at 100,000 ACUs, 15% at 500,000 ACUs
The platform itself is free; you pay per token per model, plus every Azure service the deployment provisions. Independent analysis by nOps puts a fully networked Foundry deployment 40-80% above OpenAI's own platform once Azure AI Search ($150-300/month for a typical RAG index), blob storage (~$27/month per 500GB), private endpoints ($8-13 each, with production usually needing five or more) and egress at $0.087/GB are added, and warns that every evaluation run bills tokens too. Agent Commit Units pre-purchase buys 5-15% off on one-year terms. PeerSpot reviewers describe the fragmented pricing model as the platform's single most common complaint.
Security & Compliance
Connect
Sources
This page was written from 8 sources, 5 on domains other than azure.microsoft.com.
- 1.azure.microsoft.com — ai foundryvendor
- 2.learn.microsoft.com — what is azure ai foundry
- 3.learn.microsoft.com — data privacy
- 4.azure.microsoft.com — ai foundryvendor
- 5.azure.microsoft.com — microsoft foundryvendor
- 6.peerspot.com — azure ai foundry reviews
- 7.nops.io — azure ai foundry pricing
- 8.epcgroup.net — aws bedrock vs microsoft foundry vs vertex ai 2026
Stay Ahead of the Curve
Weekly enterprise AI insights for technology leaders. No spam, no vendor pitches—unsubscribe anytime.
SubscribeRelated Products
Pigment
AI-native integrated business planning with Modeler and Analyst agents, AI-generated planning apps and an MCP server
Red Hat AI
Run any model on any accelerator across hybrid cloud, with safety evidence and cost attribution built in
Genesys Cloud
Agentic orchestration for customer experience — AI agents, employees and systems on one governed platform
VMware Tanzu Platform
Production AI agents inside your own private cloud, with a deny-by-default runtime