Azure Language in Microsoft Foundry Tools

Introduction: Language AI Has a New Home

If you have been building NLP-powered applications on Azure, you already know the drill: stitch together a handful of separate APIs, manage custom routing logic, keep on top of deprecation cycles, and hope your model behaves consistently in production.

Microsoft heard that frustration. The result is Azure Language in Microsoft Foundry Tools a consolidated, agent-native NLP platform that brings Natural Language Processing capabilities under the same roof as the rest of Azure's AI development ecosystem. Language Studio is being deprecated. LUIS has been fully retired (as of March 31, 2026). And the successor stack built around Conversational Language Understanding (CLU), Custom Question Answering (CQA), PII redaction, and a new Model Context Protocol (MCP) server is very much production-ready.

This post walks you through what Azure Language in Foundry Tools actually delivers, where it shines, and how to apply it in real-world AI projects.

What Is Azure Language in Microsoft Foundry Tools?

At its core, Azure Language is a cloud-based service that provides Natural Language Processing (NLP) features for understanding and analyzing text. Think of it as the linguistic intelligence layer of your AI stack: it reads, classifies, extracts, and routes language so your application can act on it.

What makes the Foundry Tools incarnation distinct is integration depth. Rather than calling isolated REST endpoints, developers now access Azure Language capabilities as modular tools inside the Microsoft Foundry portal the same environment where you build agents, manage models, run evaluations, and configure data pipelines.

Three delivery mechanisms are available:

• Foundry Portal (web-based): A no-code or low-code interface for testing, configuring, and deploying language features.

• REST APIs and Client Libraries: For developers who want programmatic control, including newly released .NET SDK packages targeting the 2025-11-15-preview and later API versions.

• Azure Language MCP Server: An open-protocol tool layer available both as a cloud-hosted remote server via the Foundry Tool Catalog and as a self-hosted local server that exposes language capabilities directly to agents and large language models.

Key Features and Benefits

1. Intent Routing Agent (Conversational Language Understanding)

The Intent Routing Agent is one of the flagship additions. It uses CLU to understand what a user wants within a conversation not just a single utterance, but across multi-turn exchanges and routes that intent to the right downstream skill, agent, or action.

What makes this compelling for enterprise AI builders is its deterministic behavior. Unlike purely generative approaches, CLU classifies intent using trained, human-auditable models. The system's decision-making is predictable, which matters enormously in regulated industries.

The latest release also introduces a TRIAGE_AGENT routing strategy: an agent hosted on Foundry Agent Service that uses CLU and CQA together to triage user intent and deliver precise answers, making it straightforward to build multi-skill conversational orchestration without custom glue code.

2. PII Detection and Redaction

Personally Identifiable Information (PII) detection has received significant upgrades and is now generally available (GA) as of the 2026-05-01 API release. Practically speaking, this means:

• Synthetic Replacement: Instead of simply blanking out detected entities, the service can replace them with realistic synthetic values preserving document readability for downstream analysis.

• Confidence Threshold Filtering: Set a minimum confidence score to control how aggressively the model flags entities.

• Entity Synonyms & Custom Regex: Tailor detection to your organization's specific terminology, including custom regex patterns (Text PII container).

• Broad Coverage: Detects names, email addresses, phone numbers, social security numbers, passport numbers, bank account numbers, dates of birth, driver's license numbers, and more.

Coverage extends across text documents, conversational transcripts, native document formats (.pdf, .docx), and healthcare-specific content with FHIR support.

3. Azure Language MCP Server

The Model Context Protocol (MCP) server is arguably the most architecturally significant addition for AI developers. MCP is an open standard for exposing tools and contextual data to language models and agents. Azure Language's MCP server packages the full suite of NLP capabilities PII redaction, intent detection, entity extraction, question answering, language detection, summarization as callable tools that any MCP-compatible agent or LLM can invoke.

This means you can give a Claude, GPT-4o, or any other frontier model the ability to call deterministic, privacy-safe language tools mid-conversation, without hard-coding API logic into your agent scaffolding. The remote server requires no local process management and uses Entra authentication.

4. Named Entity Recognition (NER) and Healthcare NLP

NER identifies entities such as persons, organizations, locations, dates, and events in unstructured text. Healthcare entity extraction goes further, identifying clinical concepts diagnoses, medications, body sites, procedures with FHIR standards support, enabling interoperability with health information systems.

5. Custom Question Answering (CQA) and Exact Question Answering Agent

CQA replaces the retired QnA Maker service and is meaningfully more capable. The Exact Question Answering Agent builds on CQA to deliver word-for-word, curated responses to high-value business questions think FAQ automation, compliance chatbots, and knowledge base query systems where verbatim accuracy is non-negotiable.

6. Conversation Summarization and Sentiment Analysis

Conversation summarization condenses long multi-turn exchanges into structured summaries highly useful for call center analytics, meeting recap generation, and customer support triage. Sentiment analysis and opinion mining round out the picture, providing granular positive/negative/neutral signals at the sentence and aspect level.

Real-World Applications and Case Studies

Financial Services: Compliance-Safe Customer Support

A regional bank wants to deploy a conversational AI assistant for customer inquiries about loans, accounts, and transactions. The challenge: every interaction potentially involves sensitive financial data, and regulators require auditability.

How Azure Language fits: The Intent Routing Agent handles intent classification deterministically loan inquiry goes to the loan workflow, account balance query routes to the account management skill. PII redaction with synthetic replacement strips personal data from logs before they reach the analytics pipeline. CQA handles the 200 most common compliance-sensitive questions with exact, lawyer-approved verbatim answers. The entire flow runs inside a Foundry Agent with Entra-authenticated MCP server calls fully auditable and reproducible.

Healthcare: Clinical Document Processing

A hospital system processes thousands of clinical notes daily. Extracting medication names, diagnoses, and procedures for downstream EHR enrichment is manual, slow, and error-prone.

How Azure Language fits: Healthcare entity extraction identifies clinical concepts with FHIR output format, feeding directly into the hospital's interoperability pipeline. PII redaction with document-level support (native .pdf and .docx) ensures patient identifiers never leave the secure processing environment. Conversation summarization converts physician-patient transcripts into structured encounter summaries.

Customer Experience: Intelligent Ticket Routing

A SaaS company's support desk receives thousands of tickets per day across email, chat, and web forms in multiple languages.

How Azure Language fits: Language detection automatically identifies the input locale. CLU routes tickets by intent (billing, technical, account) without relying on keyword rules. Sentiment analysis flags high-frustration interactions for immediate human escalation. The MCP server integration means the company's existing GPT-4o-based agent gains all of this without re-architecture.

Best Practices and Tips

1. Use GA API versions in production. Preview API versions (-preview suffix) are great for testing new features but are not covered by service-level agreements. For production workloads, pin to the generally available API version (currently 2026-05-01 for Text PII and 2025-11-01 for core capabilities).

2. Separate deterministic and generative tasks deliberately. Azure Language's strength is predictable, auditable output. Use CLU and CQA for intent classification and high-stakes Q&A where consistency is mandatory. Layer generative models on top for open-ended synthesis or creative tasks, not for routing logic.

3. Calibrate PII confidence thresholds per use case. A compliance workflow may need aggressive detection (lower confidence threshold). An analytics pipeline may prioritize precision over recall to avoid over-redaction. The confidenceThreshold parameter gives you that control.

4. Leverage the MCP server for agent composability. If you are building multi-step agents, the Azure Language remote MCP server lets you modularize NLP capabilities without embedding API calls in every agent skill. This simplifies maintenance and makes capabilities reusable across projects.

5. Migrate from LUIS now. LUIS was fully retired on March 31, 2026. If you are still running any LUIS-dependent workloads via archived APIs, they are no longer operational. Migrate to CLU the migration tooling is documented in the official Microsoft Foundry migration guide.

6. Use agent templates as scaffolding, not finals. The Intent Routing and Exact Question Answering agent templates, downloadable via Azure Developer CLI (azd), give you a working starting point. Customize the CLU project configuration and CQA knowledge base to reflect your domain before going to production.

7. Plan for multi-language from day one. Azure Language supports a broad range of languages and locales. If your user base is multilingual particularly relevant for East African developers serving diverse linguistic communities configure language detection at the entry point and design your CLU and CQA projects to handle locale-specific intent patterns.

Conclusion: Build with Confidence, Not Just Intelligence

The shift from Azure AI Services to Microsoft Foundry Tools is not just a rebranding exercise it represents a genuine architectural convergence. Azure Language in Foundry Tools now gives AI developers a single, composable, and governance-ready NLP layer that integrates directly with the agent development lifecycle.

The capabilities that matter most deterministic intent routing, enterprise-grade PII protection, MCP server composability, and healthcare NLP with FHIR support are production-grade and actively evolving. For teams building customer service agents, compliance systems, healthcare applications, or any NLP-heavy workflow, this is the stack to build on.

The era of stitching together disparate language APIs is ending. The unified platform is here.

Ready to Explore?

• Visit the Azure Language in Foundry Tools overview on Microsoft Learn

• Try prebuilt capabilities in the Microsoft Foundry portal — no account required for initial exploration

• Review the What's New page to stay current with monthly releases

• Scaffold your first Intent Routing Agent with: azd ai agent init -m azureml://registries/azureml-staging/agentmanifests/intent_routing_agent/versions/1

Written for AI enthusiasts and developers working at the intersection of cloud infrastructure and intelligent application design. Last updated: June 2026.