Introducing Azure AI Foundry: A CTO's Guide to Starting and Scaling AI

Updated by the TechFabric editors in October 2026 to reflect the product's renames and current documentation.

Azure AI Studio became Azure AI Foundry at Microsoft Ignite on 19 November 2024, a few weeks before this post first ran, and at Ignite in November 2025 it became Microsoft Foundry, with Azure AI Services rebranded Foundry Tools at the same time. Microsoft Learn's own history table lists the brand as "Azure AI Studio / Azure AI Foundry" before and "Microsoft Foundry" now. It's one product line, so documentation, blog posts and job adverts from 2025 use the middle name.

Foundry is Microsoft's platform for building AI applications and agents on Azure. It gives your teams one place to pick a model from a large catalog, deploy it behind an endpoint, build agents that call your systems, and evaluate and monitor what they produce. Its primary focus is getting generative AI from prototype to production with Azure's identity, networking and security already in place.

The decisions that matter are how it sits inside the Azure you already run, and the handful made on day one that are expensive to change later. We checked the product details against Microsoft's documentation on 11 October 2026.

How Foundry relates to Azure

Foundry is built on Azure, not beside it. A Foundry resource is an Azure resource, of type Microsoft.CognitiveServices/accounts with kind AIServices, that lives in your subscription and resource group. It uses Azure's identity (Entra ID and managed identities), networking (private endpoints, virtual networks), Key Vault and Azure Monitor. Billing arrives on your Azure invoice.

That's the main reason to choose it if you're already on Azure. Your security team has policies, landing zones and monitoring in place, and Foundry sits inside them. On AWS the comparable service is Bedrock, on Google Cloud it's Vertex AI, and on Databricks it's Mosaic AI with model serving and Unity Gateway. Foundry's advantage over them is entirely about where the rest of your estate already lives.

What's in it

A model catalog. Microsoft says the catalog holds "over 10,000 models, with approximately 50 new models published each month". It splits them in two, and the split matters for contracts:

• Models sold by Azure are hosted and sold by Microsoft under the Microsoft Product Terms. That covers Azure OpenAI's GPT models plus models from Meta, Mistral, DeepSeek, Cohere, xAI, Black Forest Labs and Microsoft's own, among others.
• Models from partners and community are billed through Azure Marketplace on the provider's terms. Anthropic's Claude models are in this group, since Ignite in November 2025 (Sonnet 4.5, Opus 4.1 and Haiku 4.5 at the time). Where they run depends on the model and the deployment type. Microsoft offers Claude "Hosted on Azure", processed on Azure infrastructure with Global Standard or Data Zone Standard (US) deployments, and "Hosted on Anthropic infrastructure", Global Standard only, where data "might be processed outside Azure, including outside the selected Azure region". Not every model has both. Preview models and older versions are Anthropic-hosted only. In both cases Anthropic is the seller and the data processor, under Anthropic's terms rather than Microsoft's.

Agent Service. Generally available since Build in May 2025, it's the managed runtime for agents. Microsoft lists built-in tools for web search, file search, code interpreter and memory, plus custom tools through function calls, OpenAPI specs and MCP servers. MCP is the protocol most vendors now use to expose their systems to agents, so an agent built here can reach the same tools as one built elsewhere.

Evaluation and tracing. Built-in evaluators score quality (coherence, fluency), retrieval (groundedness, relevance), safety (hate and unfairness, violence, protected material) and agent behaviour (tool call accuracy, task completion). Tracing is built on OpenTelemetry and lands in Application Insights. One cost note from the docs: evaluations in the playground are on by default and billed by consumption.

Guardrails. What used to be called content filters. Text models deploy with hate and fairness, violence, sexual and self-harm filtering at medium on prompts and completions, a jailbreak shield on prompts, and protected-material detection on completions. All of it is configurable.

Five decisions on day one

1. New Foundry project, not a hub

Foundry has two project models. The older one, a hub-based project, combines a hub resource with separate Azure OpenAI and AI Services resources. The current one is a single Foundry resource with projects under it. Microsoft says "new investments are focused on Foundry projects in the new portal", and hub-based projects now live in the "Foundry (classic)" portal.

Start new work on a Foundry project. The exception is open-weight models deployed to managed compute, such as many Hugging Face models, which still need a hub-based project. If that's your plan, know it before you build your landing zone.

Creating the resource and a project from the Azure CLI (version 2.80.0 or later), per Microsoft's docs:

az cognitiveservices account create \
    --name my-foundry-resource \
    --resource-group my-foundry-rg \
    --kind AIServices \
    --sku S0 \
    --location eastus \
    --custom-domain my-foundry-resource \
    --assign-identity \
    --allow-project-management true

az cognitiveservices account project create \
    --name my-foundry-resource \
    --resource-group my-foundry-rg \
    --project-name my-foundry-project \
    --location eastus

Microsoft notes you can't change --allow-project-management after creation. Set it to true.

2. Deployment type, which is really a data location decision

Every model deployment has a type, and the type decides where your prompts are processed:

In every case, Microsoft says "data stored at rest remains in the designated Azure geography". The difference is processing. If your compliance team has opinions about where a prompt containing customer data may be processed, this table is the conversation to have with them before anything is deployed.

Deploying a model is one command once the resource exists. This is Microsoft's own example, which uses Global Standard. Where residency matters, pick a model the docs list for Data Zone Standard and pass DataZoneStandard instead.

az cognitiveservices account deployment create \
    -n my-foundry-resource \
    -g my-foundry-rg \
    --deployment-name Phi-4-mini-instruct \
    --model-name Phi-4-mini-instruct \
    --model-version 1 \
    --model-format Microsoft \
    --sku-capacity 1 \
    --sku-name GlobalStandard

3. What happens to your data

For models sold by Azure, Microsoft's data privacy page is unambiguous. Prompts, completions, embeddings and training data "are NOT available to OpenAI or other providers of Models sold by Azure" and "are NOT used to train any generative AI foundation models without your permission or instruction".

Two caveats. Abuse monitoring may review content flagged as harmful, and customers who meet Microsoft's Limited Access criteria can apply to modify it.

The second caveat is that those statements cover models sold by Azure, and partner models don't inherit them. For Claude, Anthropic is the data processor whichever hosting option you pick, and retention is governed by Anthropic's Data Processing Addendum and commercial terms. The hosting option then decides location. Hosted on Azure keeps processing on Azure infrastructure within the Global or Data Zone scope you deploy. Hosted on Anthropic infrastructure may process prompts outside Azure altogether, so a model that's only offered that way may be ruled out for data that has to stay in a region. Check which hosting a specific model and deployment type gives you before your compliance team signs off.

4. Quota, which changed in 2026

Throughput is limited in tokens per minute and requests per minute. Historically that quota was set per model, per region, per subscription, so teams spread deployments across regions to get more. Since May 2026 Microsoft has been moving to subscription-level quota, where Global Standard deployments of the same model and version share one pool across all regions in a subscription, and Data Zone Standard shares one pool per zone. A resource also tops out at 32 deployments and 250 projects. If a load test is part of your plan, request quota early.

5. What it costs

Foundry itself is free. Microsoft's pricing page says "the Foundry platform is free to use and explore" and the features you use "are billed at their normal rates". In practice the bill is model tokens, provisioned throughput if you buy it, search and storage for retrieval, compute for agents' tools, and evaluations. Token cost usually dominates, so the biggest lever is choosing a smaller model where it does the job. Your evaluation suite tells you where that is.

Where Foundry fits best

Foundry is a strong choice when the organisation is on Azure, Microsoft 365 and Entra ID, when security wants AI inside existing controls, and when the first projects are assistants and agents over documents and business systems.

It's a weaker fit when the data your AI needs already lives in a lakehouse elsewhere. Moving governed data out to call a model is usually worse than bringing the model to the data. If your data sits in Databricks, its own model serving and Unity Gateway keep permissions in Unity Catalog, and Foundry can still be one of the model providers behind it.

And a platform doesn't produce a use case. The CTOs who get value from Foundry start with one process, a measurable outcome and an evaluation set, and only then choose models and deployment types.

A first 30 days

1. Pick one use case with a number attached, such as first-response time on support tickets or hours spent on a monthly report.
2. Create a Foundry resource and project in a non-production subscription, with Data Zone Standard if residency matters.
3. Write 30 to 50 test questions with good answers before building anything, and run them through the built-in evaluators.
4. Deploy two models of different sizes and compare quality and cost on that set.
5. Wire tracing into Application Insights from the first day, so production behaviour is visible from the first user.

TechFabric is a Microsoft Solutions Partner, with designations in Data & AI, Digital & App Innovation, and Infrastructure. If you'd like help choosing a first use case or setting Foundry up inside your Azure estate, see our AI agents work or talk to us.

Frequently asked questions

Does Microsoft's promise not to train on my prompts cover Claude on Foundry?

No. Those statements cover models sold by Azure, and Claude is a partner model. Anthropic is the data processor whichever hosting option you pick, so we'd put Anthropic's Data Processing Addendum and commercial terms through the same review as Microsoft's. We'd also check which hosting a specific model and deployment type gives you before compliance signs off.

Is Batch worth using for bulk jobs?

Yes, when nobody is waiting on the result. Batch is 50% cheaper than Global Standard with a 24-hour target. There's a Data Zone version too, so US or EU residency doesn't rule it out. We'd pair it with a smaller model where your evaluation set shows one does the job, since token cost usually dominates the bill.

If Foundry is free, why does the bill grow once agents are live?

The platform charge is zero, but the things an agent uses are each billed at their normal rate. Tools, search and storage for retrieval, and evaluations all add lines, and playground evaluations are on by default and billed by consumption. Token cost usually dominates, so we'd watch the evaluation line separately and use the test set to find where a smaller model does the job.

Sources

Checked 11 October 2026.

• What is Microsoft Foundry
• Foundry Models overview
• Deployment types
• Data, privacy and security
• Claude hosting options (updated 21 September 2026)
• Create projects
• Create model deployments
• Quotas and limits
• Observability
• Default safety policies
• Agent Service overview
• Foundry pricing
• Ignite 2024 announcement
• Ignite 2025 recap