Most explanations of Azure AI Foundry describe the product. This one describes the work. If you are trying to decide whether Foundry is worth your afternoon, the useful question is not "what is it" but "what do people actually ship with it, and what does that take".
Short version: Foundry is where you deploy a model, give it your data or your tools, check whether the result is any good, and then run it with the same identity, quota and logging rules as the rest of your Azure estate. Microsoft has begun branding it as Microsoft Foundry, and you will see both names in the portal and the docs. They are the same platform.
What Azure AI Foundry is, in one paragraph
Foundry is a workspace that sits on top of Azure's AI services. Inside a project you get a model catalogue you can deploy from, an agent runtime, connections to your data sources, an evaluation harness, content safety controls and tracing. Without it, you would be wiring Azure OpenAI, Azure AI Search, Document Intelligence, Speech and your own orchestration code together by hand, and inventing your own way to measure whether the result works. That is the practical difference: Foundry is the assembly point.
1. A question-answering assistant over your own documents
This is the most common first build and the reason most teams arrive. You point Foundry at a document set, index it in Azure AI Search, deploy a chat model, and connect the two so answers cite your content rather than the model's training data.
What the work actually consists of: chunking decisions, choosing between keyword, vector and hybrid retrieval, deciding what the model is allowed to say when retrieval returns nothing, and measuring whether answers are grounded in the retrieved text. The model choice is rarely the hard part. Retrieval quality is. If your answers are wrong, the retrieval is almost always the reason.
2. Agents that call real tools
An agent in Foundry is a model plus instructions plus a set of tools it can call. Tools might be a file search over your documents, a function that hits your internal pricing API, or an MCP server exposing a whole toolset.
Teams use this for triage, internal helpdesks, first-pass data gathering and anything where the answer requires looking something up rather than recalling it. The engineering work is in tool design, not prompt writing: clear tool descriptions, tight input schemas, approval gates on anything that writes, and a plan for what happens when a tool times out. What MCP is and why it matters covers the tool integration side in detail.
3. Multi-agent workflows
When one agent needs too many tools or too many rules, teams split the work: a router agent that classifies, specialists that handle each class, and a reviewer that checks the output. Foundry supports sequential, concurrent, group chat, handoff and planner-led patterns.
This is powerful and easy to overuse. The honest test is whether a single agent has already failed for a reason more agents would fix. The five orchestration patterns, with code walks through when each one is the right shape.
4. Document and form processing at volume
Invoices, claims, contracts, onboarding packs. Document Intelligence and Content Understanding extract structured fields, and a Foundry model handles the parts that need judgement rather than extraction, such as summarising a clause or flagging an unusual term.
The value here is not the demo, it is the exception path. Teams that succeed build a confidence threshold and a human review queue on day one, rather than after the first bad extraction reaches a customer.
Try a real Azure AI lab, free
A pre-provisioned Azure environment in your browser. No Azure account, no credit card, no setup.
Most large enterprises already run on Microsoft infrastructure. Azure AI Foundry is the platform those teams use to build and deploy AI apps, so the skills you practice here map directly to real job demand.
5. Speech, translation and voice interfaces
Call transcription with speaker separation, live translation, voice agents that answer questions over the phone. Foundry gives you the speech services alongside the language model, so the transcript feeds the model in the same project rather than crossing three systems.
The practical constraints are latency and turn taking. A voice experience that would be fine as text feels broken at two seconds of silence, so response streaming and interruption handling matter more than model quality.
6. Evaluation pipelines
This is the use case teams discover late and then rely on most. Foundry ships built-in evaluators that score responses for things like task completion, tool call accuracy, groundedness and relevance, and you run them against a saved test set rather than judging by feel.
Once evaluation exists, everything else gets easier: you can change a prompt, swap a model, adjust chunk size, and see whether it helped. Without it, every change is a guess. How to evaluate an AI agent in Foundry covers which evaluator answers which symptom.
7. Governed access to frontier models
Some organisations use Foundry for a narrower reason: they want their teams using current models, but inside Azure's identity, networking, data residency and cost boundaries. Foundry gives you the model catalogue, deployment quotas per project, managed identity instead of shared keys, content filters, and usage that shows up in the same billing you already review.
For regulated industries this is often the whole business case, before anyone builds an agent at all.
8. Prototypes that survive contact with production
The under-rated use. A Foundry project built in the portal can be exported to code with the same SDK the production app will use, so the prototype is not thrown away. That matters because the gap between a working demo and a supportable service is where most internal AI projects quietly die.
How to start using Azure AI Foundry
The first hour looks like this:
- Sign in to the Azure portal or the Foundry portal. A Foundry project is now created for you automatically on first sign-in, so if a tutorial tells you to create one manually and your screen already shows a project, you are not lost.
- Pick a region before you deploy anything. Model availability varies by region, and moving later is more annoying than choosing carefully now.
- Deploy one chat model from the catalogue and note the quota you were given. Quota, not capability, is what stops most first projects.
- Open the playground and get a plain conversation working. No data, no tools.
- Add one source of grounding, either a file upload or a search index, and compare the answers with and without it.
- Save five or six real questions with their expected answers, and run an evaluation. This is the step people skip and later regret.
- Switch from keys to managed identity before anything leaves your machine.
If you want that sequence in a prepared Azure environment rather than your own subscription, our AI-103 lab course runs it end to end with the environment already provisioned.
Want to build these eight things yourself?
Pre-provisioned Azure AI Foundry sandboxes. Deploy a model, wire retrieval, and ship an agent without touching your own subscription.
What Foundry is not the right tool for
Being honest about this saves time:
- Deterministic lookups. If the answer is a row in a database, query the database. Wrapping it in a model adds cost, latency and a failure mode. When RAG is not the right answer covers the most common cases.
- Heavy custom model training. Foundry is built around using and adapting models, not around large-scale training pipelines. That work belongs in Azure Machine Learning.
- Classic computer vision tasks. Some of the older Custom Vision style workflows are no longer where Microsoft is investing.
- Buying rather than building. If Microsoft 365 Copilot already does the job, building the same thing in Foundry is a project you have to maintain forever.
What it costs to run
There is no Foundry licence fee. You pay for what the underlying services consume: model tokens or provisioned throughput, search index size and queries, storage, speech minutes, and any compute you attach. The portal's estimate is not the whole picture, because retrieval and evaluation calls add up alongside the visible chat traffic. Set a budget alert on the resource group before your first pilot, not after.
How this connects to AI-103
If you are certifying, Foundry is not one topic on the AI-103 blueprint, it is the environment the whole exam assumes. The scenarios describe exactly the builds above: ground an assistant, give an agent the right tool, work out why an answer is unsupported, choose an evaluator for a symptom. Candidates who have built two or three of these recognise the questions. Candidates who have only read about them do not.
Frequently asked questions
What is Azure AI Foundry used for? Building and running AI applications on Azure: retrieval-based assistants over your own content, agents that call tools and APIs, document and form processing, speech and translation experiences, evaluation pipelines that measure output quality, and governed access to current models under your organisation's identity and cost controls.
Is Azure AI Foundry the same as Azure OpenAI Service? No. Azure OpenAI Service provides model endpoints. Foundry is the workspace around them, adding the agent runtime, data connections, evaluation, safety controls and tracing. You can use Azure OpenAI on its own, and many teams start there, but you end up rebuilding parts of Foundry yourself.
Is Azure AI Foundry the same as Microsoft Foundry? Yes. Microsoft has been consolidating the branding, so both names refer to the same platform. Older material says Azure AI Studio, which was the earlier name.
Do I need to write code to use Azure AI Foundry? No for prototyping, yes for anything real. The portal playground gets you a working assistant without code. Production work uses the SDK, most commonly Python, and you will want your prompts, tools and evaluations in version control.
Do I need my own Azure subscription to learn it? You need one to build freely. If you only want to practise the AI-103 scenarios, a prepared lab environment avoids the quota requests, region availability issues and cleanup that consume the first day of self-study.
Want to work through these builds instead of reading about them? The AI-103 course includes 18 hands-on labs in a dedicated Azure environment with Foundry already configured, from a first grounded assistant to multi-agent evaluation.
Confused by the naming? See Foundry vs Azure AI Studio vs Azure OpenAI: what actually changed.