Foundry Tools is Microsoft's new name for what used to be Azure AI Services: the ready-made APIs for speech, language, translation, vision, documents and content safety that sit beside the models in Microsoft Foundry. Each one does a narrow job very reliably, for example turning an invoice into fields or a phone call into text, without you writing a prompt. This guide explains every current tool twice: once in plain words, once technically. It also says when a general model such as GPT-5 is now the better choice, and which tools Microsoft is retiring in 2028 and 2029.
Checked on 20 October 2026. Every name, date and package below was read on Microsoft Learn that day. Microsoft renames and retires these services often, so the dates matter more than usual.
The short version
Microsoft lists ten current Foundry Tools. Here is what each does in one line, and where it stands in October 2026.
| Tool | In plain words | Status in October 2026 |
|---|---|---|
| Speech | Turns speech into text and text into a natural voice | Active, growing (Voice Live, new transcription models) |
| Language | Finds names, private data and meaning in text | Core features active; most others retire 31 March 2029 |
| Translator | Translates text and whole documents | Active; new API lets you pick classic or LLM translation |
| Document Intelligence | Reads forms, invoices and receipts into fields | Active; v2.1 API retires 15 September 2027 |
| Content Understanding | Turns documents, images, audio and video into the output you define, using generative AI | Active, generally available since November 2025 |
| Vision | Describes images and reads text in them | Image Analysis API retires 25 September 2028 |
| Azure AI Search | Finds the right passages in your own data | Active; powers Foundry IQ |
| Content Safety | Spots harmful input and output | Active; its branding is moving towards the Foundry control plane |
| Custom Vision | Trains your own image classifier | Supported for existing customers until 25 September 2028 |
| Immersive Reader | Helps people read and understand text | Active |
Face is not a separate row in Microsoft's list. It is documented as part of Vision and needs approved access.
Six older services are listed as retired: Anomaly Detector, Content Moderator, Language Understanding (LUIS), Metrics Advisor, Personalizer and QnA Maker. If a tutorial tells you to create one of those, it is out of date.
How Foundry Tools fit into Microsoft Foundry
In plain words
Think of Microsoft Foundry as one workshop. The models (GPT-5, Llama, Phi and thousands of others) are the general-purpose craftsmen who can do almost anything if you explain it well. The Foundry Tools are the specialist machines on the bench: each does one job, the same way every time, and you pay per use. Agents are the workers who decide which craftsman or machine to use for a task.
Technically
- One resource. A Foundry resource has the Azure resource kind
AIServices. Microsoft describes Azure OpenAI and Foundry Tools as "reachable through one endpoint and key" on that resource. The services bundled directly with it are Agent Service, Azure OpenAI, Content Safety, Document Intelligence, Language, Speech and Translator. Content Understanding is also available through the Foundry resource. - Endpoints. The tool APIs are called on the resource endpoint,
https://<resource-name>.cognitiveservices.azure.com/. Agents, toolboxes and MCP use the project endpoint,https://<resource>.services.ai.azure.com/api/projects/<project>. Keep those two apart and most "404 on my endpoint" problems disappear. - Authentication. Every tool accepts a key in the
Ocp-Apim-Subscription-Keyheader, or a Microsoft Entra ID token. Keyless (Entra ID) access needs the custom subdomain endpoint, not a regional one, and the caller needs the Cognitive Services User role. Managed identities work too. - Billing. Pricing is per transaction, with the unit depending on the tool: audio hours, characters, pages, text records or images. Many tools have a free
F0tier for trying things out. - Inside agents. An agent in Foundry Agent Service reaches tools through its tool list or a Foundry Toolbox, which Microsoft calls "the recommended way to add tools to an agent". Azure AI Search is a native agent tool. Speech and Language plug in as MCP servers. Do not confuse Foundry Tools with the Foundry Tools Catalog of agent connectors, a different list of more than a thousand connectors and MCP servers that shares the name.
A minimal keyless client looks the same for every tool. This one is Document Intelligence in Python:
from azure.identity import DefaultAzureCredential
from azure.ai.documentintelligence import DocumentIntelligenceClient
client = DocumentIntelligenceClient(
endpoint="https://<resource-name>.cognitiveservices.azure.com/",
credential=DefaultAzureCredential(), # needs the Cognitive Services User role
)
Swap the package and client class and the pattern holds for Language, Translator, Vision and Content Safety.
Speech
In plain words
Speech listens and talks. It writes down what people say (in real time, from a file, or in bulk), reads text out in a natural voice, translates speech, and can tell you how well someone pronounced a word. It also powers Voice Live, which lets you build an assistant people talk to rather than type to.
Technically
The official name is Azure Speech in Foundry Tools. Its main parts:
- Speech to text in three modes: real-time, fast and batch transcription, plus custom speech models trained on your own vocabulary and audio.
- Text to speech with neural voices, custom voices, SSML for fine control, and a talking avatar.
- Speech translation, language identification and pronunciation assessment.
- LLM Speech, a mode on fast transcription that can transcribe and translate in one call. It became generally available in May 2026.
- Voice Live, generally available since November 2025 and connected to Foundry Agent Service since April 2026. Voice-based prompt agents in Foundry run on it.
You call it through the Speech SDK (C#, C++, Go, Java, JavaScript, Objective-C, Python and Swift), the Speech to text REST API (version 2025-10-15 is the current GA release; v3.0 and v3.1 were retired on 31 March 2026), or the Voice Live SDKs such as azure-ai-voicelive for Python. An agent can use Speech through the Azure Speech MCP server in the Foundry tool catalog. That server needs the Agent Service Enterprise tier and does not support a network-secured Foundry.
Two things have quietly gone: intent recognition in Speech was retired on 30 September 2025, and the Speech SDK dropped speaker recognition in November 2025 because the service was retired. Some Microsoft pages still mention speaker recognition, so treat those lines as stale.
Use Speech, or a model with audio?
Microsoft's own architecture guidance draws the line clearly. Use Speech when you need speaker diarization, custom speech models or custom vocabulary for domain-specific or noisy audio. Use a model with audio input when the speech goes straight into reasoning, summarising or understanding in one call.
Practise it
AI-901 Lab 4, "Add Voice to Meridian's Platform", builds a voice feature with Speech in a real Azure environment. See the AI-901 labs.
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.
Language
In plain words
Language reads text and pulls structure out of it: who and what is mentioned, which parts are private (names, phone numbers, account numbers), what language it is in, and, for medical text, which conditions and medications appear. It is the tool you use when the same text task has to be done millions of times, the same way, and audited.
Technically
The official name is Azure Language in Foundry Tools. Microsoft now splits it in two:
- Core capabilities, recommended for new work: PII detection (in text, conversations and documents), language detection, named entity recognition (prebuilt and custom), and Text Analytics for health.
- Legacy capabilities, retiring: conversational language understanding (CLU), custom question answering, sentiment and opinion mining, key phrase extraction, summarization, custom text classification and orchestration workflow all retire on 31 March 2029. Entity linking retires earlier, on 1 September 2028. Language Studio itself retires on 20 March 2027.
Each legacy page carries the same instruction: direct new projects to Microsoft Foundry models. So if you are about to build sentiment analysis from scratch today, a model prompt or a fine-tuned small model is the path Microsoft is pointing you to.
Developers call Language with azure-ai-textanalytics in Python (the runtime client) or Azure.AI.Language.Text in .NET. PII detection's current GA API version is 2026-05-01, and its synthetic replacement option, which swaps real personal data for realistic fake data, became generally available in June 2026. For agents there is an Azure Language MCP server (preview) with a remote endpoint on the resource and a local version you host yourself.
Use Language, or a model?
For PII detection, named entities and health text, Language is the recommended path: it is consistent, cheaper at volume, and designed for compliance. For sentiment, summaries and intent, Microsoft has already scheduled the retirement and points you to models. That is the clearest signal in the whole Foundry Tools family about where specialised NLP is heading.
Practise it
AI-103 Lab 5, "Understanding Guest Reviews with Azure AI Language", runs sentiment, key phrases, entities, PII detection and summarisation on real guest feedback, and AI-901 Lab 3 covers the same service at fundamentals level. See the AI-103 labs.
Translator
In plain words
Translator turns text or whole documents from one language into another, keeping the document's layout. You can teach it your own terms so product names and industry words come out right.
Technically
The official name is Azure Translator in Foundry Tools. It offers text translation, document translation (batch and single-file), Custom Translator, and an adaptive custom translation option that adapts a large model with as few as 5 and up to 10,000 example sentence pairs.
The big change is the 2026-06-06 text translation API, generally available since June 2026. On each request you can choose classic neural machine translation or LLM translation, and add tone and gender controls. LLM translation needs a Foundry resource. The new version is not backward compatible with v3.0, and it drops the BreakSentence, Detect and Dictionary operations. The text SDKs (azure-ai-translation-text, Azure.AI.Translation.Text) still target v3.0, so for the new options you call the REST API directly. Document translation's SDKs (azure-ai-translation-document 2.0.0 in Python) default to API 2026-03-01, which added translation of image files.
Use Translator, or a model?
Microsoft's guidance is that Translator is "more effective and cost effective than general-purpose foundation language models because of its targeted translation models", especially for large document volumes where formatting must survive. The 2026 API now gives you both in one service, which removes most of the reason to build translation on a raw model.
Document Intelligence
In plain words
Document Intelligence reads business paperwork. Give it an invoice, a receipt, an ID card, a bank statement or a tax form, and it returns the fields: totals, dates, names, line items, each with a confidence score. You can train it on your own form layouts too.
Technically
The official name is Azure Document Intelligence in Foundry Tools; its old name was Form Recognizer. The current API is v4.0 (2024-11-30), with SDKs azure-ai-documentintelligence (Python) and Azure.AI.DocumentIntelligence (.NET), both at 1.0.0.
- Analysis models: Read (text) and Layout (text, tables, structure).
- Prebuilt models: invoice, receipt, ID, bank statement, check, contract, credit card, pay stub, US tax and mortgage forms, health insurance card and marriage certificate.
- Custom models: neural, template, composed and classifier.
- Add-ons: high resolution, formulas, fonts, barcodes, searchable PDF, query fields and key-value pairs.
The v2.1 API retires on 15 September 2027 and v3.0 on 30 March 2029, so check which version older code uses. Microsoft now presents Document Intelligence as part of the wider Content Understanding family, but says existing users' "APIs, endpoints, SDKs, and billing are unchanged. No migration is required."
Document Intelligence or Content Understanding?
This is the question most people get wrong. Microsoft's architecture guidance says that where a Document Intelligence prebuilt model exists for your document, use it: it is "more cost-effective and deterministic for these scenarios". Reach for Content Understanding when the document type has no prebuilt, when you need fields that must be inferred rather than read, or when the input is not a document at all.
Practise it
AI-901 Lab 6, "Automate Meridian's Document Processing", extracts fields from business documents. See the AI-901 labs.
Content Understanding
In plain words
Content Understanding is the newest tool and the most flexible. You describe the output you want, for example "vendor, total, payment terms, and whether the contract allows early termination", and it uses generative AI to fill that in from a document, an image, an audio recording or a video. It also tells you how confident it is and where in the source each answer came from.
Technically
The official name is Azure Content Understanding in Foundry Tools. It became generally available in November 2025 (API 2025-11-01), with SDKs azure-ai-contentunderstanding and Azure.AI.ContentUnderstanding generally available since March 2026.
- Analyzers are the unit of work: prebuilt or custom.
- Content extraction: OCR, layout and transcription.
- Segmentation and classification, for example splitting a scanned bundle into separate documents.
- Field extraction with three methods: extract (read it), classify (pick from options) and generate (infer it).
- Confidence scores and grounding back to the source.
- Prebuilt RAG analyzers such as
prebuilt-documentSearchandprebuilt-videoSearch, made to prepare content for retrieval.
One technical detail matters for planning: since general availability you bring your own Foundry model deployment, because Microsoft retired the managed capacity used in the preview. Billing therefore has two parts: Content Understanding's own charges (per 1,000 pages, or per minute of audio and video, plus contextualization tokens) and the tokens your model deployment uses. Recent additions include an agentic mode (preview, July 2026), an Agent Framework package and a command line tool (preview, September 2026).
Content Understanding, or your own prompts?
Microsoft's comparison page asks exactly this. Build your own solution on Foundry models when you need "complete control over model selection, fine-grained prompt ownership, or integration with a proprietary AI infrastructure". Otherwise Content Understanding saves you the work of building extraction, confidence scoring and grounding yourself.
Practise it
AI-103 Lab 13, "Reading the Paperwork", uses Content Understanding to turn an invoice PDF and a contract into structured fields you can check against each other. Content Understanding is named in the AI-103 skills outline for both vision and information extraction, so it is worth real practice. See the AI-103 labs.
Still here? You clearly care, grab the exam study guide
One-page quick reference plus the hands-on labs that go with it.
Vision and Face
In plain words
Vision looks at a picture and tells you what is in it: a caption, objects, people, text, and a good place to crop. Face detects faces and, for approved customers only, can check whether two photos show the same person or whether a face is live rather than a photo of a photo.
Technically
The official name is Azure Vision in Foundry Tools (formerly Computer Vision). Image Analysis 4.0 provides read (OCR), captions and dense captions, tags, objects, people, smart crop and multimodal embeddings, through azure-ai-vision-imageanalysis in Python.
The important date: the Image Analysis API (v3.2 and v4.0) retires on 25 September 2028, in the cloud and in containers. Microsoft's suggested replacements are Document Intelligence for OCR, the Face API for faces, embedding models such as Cohere Embed for image search, GPT models for description, and Content Understanding. Product recognition, custom image models and background removal were already retired on 31 March 2025. Custom Vision, the service for training your own image classifier, is supported for existing customers until the same date, 25 September 2028.
Face (azure-ai-vision-face in Python) offers detection, attributes, verification (one to one), identification (one to many), find similar, grouping and liveness detection. Identification, verification and liveness are limited to Microsoft managed customers and partners through an intake form, and emotion and gender attributes have been retired as part of Microsoft's responsible AI standard.
Use Vision, or a multimodal model?
Microsoft's guidance is blunt: do not use Vision for "analysis that large, multimodal foundation models like GPT-4o already support". With the 2028 retirement date, new image understanding work belongs on a multimodal model or Content Understanding, and Vision is for existing workloads and narrow needs such as smart crop.
Practise it
AI-103 Lab 11, "What the Photos Say", uses model vision to turn photographs into structured inspection reports and accessibility alt text, and shows where a visual answer stops being reliable. AI-901 Lab 5 covers computer vision at fundamentals level. See the AI-103 labs.
Content Safety
In plain words
Content Safety is the bouncer. It checks what users send in and what the AI sends back, blocks harmful content in categories such as hate, violence, sexual content and self-harm, catches attempts to trick a model into ignoring its rules, and flags answers that are not supported by the source material.
Technically
Learn still calls it Azure AI Content Safety, while Microsoft's pricing page and architecture guidance now call it "Content Safety in Foundry Control Plane". Expect the name to settle on the control plane side.
- Prompt Shields detect direct jailbreak attempts and indirect attacks hidden in documents or tool output.
- Groundedness detection (preview) checks whether an answer is supported by its sources.
- Protected material detection, a task adherence API (preview since November 2025), text and image analysis, and custom categories.
In Foundry you rarely call it directly. The same protections appear as guardrails: collections of controls applied at four points, the user's input, a tool call, a tool's response and the model's output. Called directly, the SDKs are azure-ai-contentsafety and Azure.AI.ContentSafety, billed per 1,000 text records or per 1,000 images.
Practise it
AI-103 Lab 6, "The Stress Test", attacks an assistant with jailbreaks, prompt injection and data leakage attempts, then closes the gaps with a custom content filter and blocklist. See the AI-103 labs.
Azure AI Search
In plain words
Search is how an AI app finds the right paragraph in your own documents before answering. It is the retrieval part of retrieval augmented generation (RAG): instead of guessing, the model is handed the relevant passages.
Technically
Microsoft describes Azure AI Search (formerly Azure Cognitive Search) as "a fully managed, cloud-hosted service that connects your data to AI". It supports full-text, vector, hybrid and multimodal queries, AI enrichment during indexing, and agentic retrieval through knowledge bases and knowledge sources, which is generally available with some parts in preview. It underpins Foundry IQ, Foundry's managed knowledge layer for agents, and it is one of the few tools an agent can use natively, both directly and inside a Toolbox. A serverless pricing model is in preview alongside the dedicated tiers.
Practise it
AI-103 Lab 8, "Solara's Internal Knowledge Vault with Foundry IQ", builds a knowledge base from internal documents and attaches it to an agent, and AI-901 Lab 7 introduces Foundry IQ. See the AI-103 labs.
The retirement calendar in one place
Microsoft publishes these dates on separate pages. Together they show the direction: narrow NLP and image analysis move to models, while speech, translation, document extraction, safety and search stay as dedicated tools.
| Date | What retires |
|---|---|
| 31 March 2025 (done) | Vision product recognition, custom image models, background removal |
| 30 September 2025 (done) | Intent recognition in Speech |
| November 2025 (done) | Speaker recognition removed from the Speech SDK |
| 31 March 2026 (done) | Speech to text REST API v3.0 and v3.1 |
| 20 March 2027 | Language Studio |
| 15 September 2027 | Document Intelligence API v2.1 |
| 1 September 2028 | Language entity linking |
| 25 September 2028 | Vision Image Analysis API (v3.2 and v4.0); Custom Vision support ends |
| 31 March 2029 | Language CLU, custom question answering, sentiment, key phrases, summarization, custom text classification, orchestration workflow; Document Intelligence API v3.0 |
How to choose, in four questions
- Is there a prebuilt model for exactly this input? An invoice, a receipt, an ID: use the Document Intelligence prebuilt. A known language pair: use Translator. Deterministic and cheaper beats clever.
- Is the task on Microsoft's retirement list? Sentiment, summaries, key phrases, intent, image captioning: build new work on a Foundry model, because the dedicated feature has an end date.
- Do you need fields that must be inferred, or mixed media? Use Content Understanding, and budget for the model tokens it consumes.
- Is it audio, safety or retrieval? Speech, Content Safety (through guardrails) and Azure AI Search remain the right tools, and agents can use them through MCP or the Toolbox.
Frequently asked questions
What is the difference between Foundry Tools and Foundry Models?
Foundry Models are general-purpose models such as GPT-5, Llama and Phi that you deploy and prompt. Foundry Tools are prebuilt APIs for one job each, such as speech to text or invoice extraction, billed per transaction. Agents can use both.
Are Azure AI Services and Foundry Tools the same thing?
Yes. Microsoft renamed Azure AI Services to Foundry Tools, and each service now carries "in Foundry Tools" in its name, for example Azure Speech in Foundry Tools. The APIs, SDKs and endpoints did not change because of the rename.
Is Azure AI Language being retired?
Not all of it. PII detection, language detection, named entity recognition and Text Analytics for health remain core. Most other Language features, including sentiment analysis, summarization and conversational language understanding, retire on 31 March 2029, and Microsoft asks new projects to use Foundry models instead.
Should I use Document Intelligence or Content Understanding?
Use a Document Intelligence prebuilt model when one exists for your document, because Microsoft calls it more cost-effective and deterministic. Use Content Understanding for document types without a prebuilt, for fields that must be inferred, or for images, audio and video.
Which Foundry Tools are on the AI-103 exam?
The AI-103 skills outline names Content Understanding for vision and information extraction, Azure Translator and Speech for text and speech solutions, Azure AI Search for retrieval, and safety controls such as guardrails and Prompt Shields. Language-model text analysis with Foundry Tools also appears.
Related reading
- Microsoft Foundry explained: the platform these tools live in.
- What Is Azure AI Foundry Used For?: eight things people build with it.
- Microsoft Foundry vs Azure AI Studio vs Azure OpenAI: the naming history.
- AI-103 Study Guide: every exam domain.