Gem 005: Multi-Language Agent Response¶

Detect your user's language and respond in kind — without asking every time.

Classification¶

Attribute	Value
Category	Personalization
Complexity	⭐⭐⭐ (Moderate — detection logic + instruction adaptation)
Channels	All (detection method varies by channel)
Prerequisite Gems	Gem 001 (for persisting language preference across sessions)

The Problem¶

Enterprise organizations are global. A Copilot Studio agent deployed to 5,000 employees across France, the US, and Japan will receive queries in French, English, and Japanese — sometimes mixed within a single conversation.

The default Copilot Studio experience handles this poorly:

Agent instructions are monolingual: The system prompt is written in one language. If the instructions say "respond professionally," the LLM defaults to the language of the instructions — usually English.
No built-in language detection: There's no System.User.Language variable. The platform doesn't tell you what language the user typed in.
Repeated friction: Asking "What language do you prefer?" every session is wasteful. For a French employee who always writes in French, this should be automatic.
Knowledge source language mismatch: If your knowledge base is in English but the user asks in French, the generative answer may respond in English, French, or an awkward mix — depending on how the LLM processes the retrieved context.
Mid-conversation switching: A user starts in English, then switches to French when asking about French-specific policies. The agent should follow seamlessly.

The core challenge: determining the user's language reliably, applying it consistently, and persisting it across sessions — with minimal user friction and maximum accuracy.

The Ideal Outcome¶

An agent that responds in the user's language naturally:

[ ] Automatic detection: Language is identified without explicitly asking the user
[ ] Accurate matching: Detection works correctly >95% of the time for supported languages
[ ] Persistent preference: Once detected (or chosen), the language preference carries across sessions
[ ] Seamless switching: If the user changes language mid-conversation, the agent follows
[ ] Knowledge source compatibility: Responses are in the user's language even when knowledge sources are in a different language

Approaches¶

Approach A: Graph API Profile-Based Detection¶

Summary: Retrieve the user's preferredLanguage or usageLocation from their Entra ID profile via Microsoft Graph API. Set the language at conversation start.
Technique: Power Automate flow calling Graph API, global variable for language, agent instructions for response language.

How It Works¶

flowchart TB
    A["Conversation start"] --> B["Power Automate:<br/>GET /users/&#123;id&#125;?$select=<br/>preferredLanguage, usageLocation"]
    B --> C["Map to language:<br/>preferredLanguage = fr-FR → French"]
    C --> D["Set Global.ResponseLanguage = French"]
    D --> E["Agent instructions:<br/>Respond in Global.ResponseLanguage"]

The user's Entra ID profile contains preferredLanguage (e.g., fr-FR, en-US, ja-JP) — set by IT administrators. This is the most reliable signal because it's admin-managed and doesn't depend on what the user types.

Implementation¶

Step 1: Retrieve language from Graph API

Extend the GetUserContext flow from Gem 001 (or create a standalone flow):

Action: HTTP to Microsoft Graph
  Method: GET
  URI: https://graph.microsoft.com/v1.0/users/{userId}?$select=preferredLanguage,usageLocation,mail

Action: Map Language (Compose)
  Expression: 
    if(startsWith(outputs('HTTP')?['body/preferredLanguage'], 'fr'), 'French',
    if(startsWith(outputs('HTTP')?['body/preferredLanguage'], 'ja'), 'Japanese',
    if(startsWith(outputs('HTTP')?['body/preferredLanguage'], 'de'), 'German',
    if(startsWith(outputs('HTTP')?['body/preferredLanguage'], 'es'), 'Spanish',
    'English'))))

Output: responseLanguage (string)

Step 2: Set global variable and instruct the agent

Via agent instructions (M365 Copilot compatible):

kind: GptComponentMetadata
displayName: Multilingual Agent
instructions: |+
  # Multilingual Agent

  ## CRITICAL: Language Detection
  At the START of every conversation, call "GetUserContext" to retrieve the user's 
  language preference. Store it as the conversation language.

  ## Language Rules
  - ALWAYS respond in the user's detected language
  - If the detected language is "French": respond entirely in French
  - If the detected language is "Japanese": respond entirely in Japanese
  - Default: respond in English

  ## Mid-Conversation Language Switch
  If the user writes a message in a DIFFERENT language than detected, 
  switch to that language for the remainder of the conversation.

  ## Knowledge Source Handling
  When knowledge sources are in English but the user's language is French:
  - Search in English (knowledge sources are English)
  - Translate your response to French
  - Cite source documents by their original English titles

Or via ConversationStart topic (non-M365 channels):

kind: AdaptiveDialog
beginDialog:
  kind: OnConversationStart
  id: main
  actions:
    - kind: InvokeFlow
      id: getUserLanguage
      flowId: "@environmentVariables('GetUserContextFlowId')"
      inputs:
        userId: =System.User.Id
      outputVariable: Topic.UserContext

    - kind: SetVariable
      id: setLanguage
      variable: Global.ResponseLanguage
      value: =If(IsBlank(Topic.UserContext.responseLanguage), "English", Topic.UserContext.responseLanguage)

Step 3: Persist language preference (link to Gem 001)

Save the detected language using Gem 001's WriteContext pattern so the Graph API call only happens once per user — subsequent sessions load the cached preference.

Evaluation¶

Criterion	Rating	Notes
Ease of Implementation	🟡	Requires Power Automate flow + Graph API. Straightforward but multi-component.
Maintainability	🟢	Language mapping is a simple lookup. Adding languages = adding mapping entries.
Channel Compatibility	🟢	Graph API works regardless of channel.
Detection Accuracy	🟡	Depends on IT setting `preferredLanguage` correctly. Many orgs leave this blank or set to defaults.
Seamless Switching	🟡	Doesn't detect mid-conversation switches. Agent instructions handle this via "if user writes in different language."
Knowledge Compatibility	🟢	Agent instructions can mandate translation of English knowledge to user's language.

Limitations¶

Depends on IT data quality: preferredLanguage is often blank or set to the org default (en-US) for everyone. If your org hasn't populated Entra profiles, this approach returns English for all users.
Fallback chain needed: preferredLanguage → usageLocation → default. Multiple Graph fields may need checking.
One language at start: Detects once. If the user's profile says English but they write in French, the agent starts in English (though instructions can tell it to switch).
Power Automate latency: Adds 1-2 seconds at conversation start for the Graph API call.

Approach B: LLM Auto-Detection from User Input¶

Summary: Let the LLM detect the language of the user's first message and respond in that language automatically. No external calls needed.
Technique: Agent instructions that mandate language matching, with optional Prompt Tool for explicit classification.

How It Works¶

flowchart TB
    A["👤 User: Quelle est la politique de congés ?"] --> B["LLM detects: input is French"]
    B --> C["LLM responds in French<br/>(per instructions)"]
    C --> D["All subsequent responses in French<br/>(until user switches)"]

The simplest possible approach: the LLM naturally detects the input language and mirrors it — you just need to tell it to do so explicitly in the instructions.

Implementation¶

Step 1: Add language-matching instructions

kind: GptComponentMetadata
displayName: Auto-Language Agent
instructions: |+
  # Language-Adaptive Agent

  ## Language Detection Rule
  ALWAYS respond in the SAME LANGUAGE the user writes in.

  - If the user writes in French → respond entirely in French
  - If the user writes in English → respond entirely in English
  - If the user writes in Japanese → respond entirely in Japanese
  - If you're unsure → respond in English and ask:
    "I detected your message might be in [language]. Would you prefer I respond in [language]?"

  ## Language Consistency
  - Once you detect the user's language, maintain it for the entire conversation
  - If the user switches language mid-conversation, follow their switch
  - NEVER mix languages in a single response (except for proper nouns or cited document titles)

  ## Knowledge Source Translation
  Our knowledge base is in English. When the user's language is not English:
  - Search using the user's original query (the search engine handles multilingual queries)
  - Translate the retrieved information into the user's language
  - Keep document titles and reference IDs in their original language

That's it. No flows, no Graph API, no variables. The LLM handles everything.

Step 2 (Optional): Explicit classification via Prompt Tool

For more control (e.g., to store the detected language in a variable for routing):

kind: PromptTool
id: prompt_detectLanguage
displayName: "Language Detector"
description: "Detects the language of user input"
instructions: |
  Detect the language of the following text and return ONLY the language name.

  Supported languages: English, French, German, Spanish, Japanese, Chinese, Korean, Portuguese, Italian, Dutch, Arabic

  If the text is too short or ambiguous, return "English" as default.

  Text: {userMessage}

  Return ONLY the language name, nothing else.

model:
  provider: ManagedModel
  modelNameHint: GPT4Mini
inputs:
  - name: userMessage
    type: string
    required: true
outputs:
  - name: language
    type: string

Step 3 (Optional): Persist detected language (link to Gem 001)

After detection, save the language via Gem 001's persistence so the agent defaults correctly in the next session:

    - kind: InvokePrompt
      id: detectLang
      promptId: prompt_detectLanguage
      inputs:
        userMessage: =System.Activity.Text
      outputVariable: Topic.DetectedLanguage

    - kind: SetVariable
      id: setLang
      variable: Global.ResponseLanguage
      value: =Topic.DetectedLanguage

    # Persist for next session
    - kind: InvokeFlow
      id: saveLangPref
      flowId: "@environmentVariables('WriteContextFlowId')"
      inputs:
        userId: =System.User.Id
        preferredLanguage: =Global.ResponseLanguage

Evaluation¶

Criterion	Rating	Notes
Ease of Implementation	🟢	Instructions only (no Prompt Tool needed). Zero infrastructure. 15 minutes.
Maintainability	🟢	One instruction block. Adding languages = adding a line.
Channel Compatibility	🟢	Works everywhere — LLM-based, no channel dependencies.
Detection Accuracy	🟡	LLMs are excellent at language detection for sentences. Weak for 1-2 word inputs ("Help" could be English or French verb).
Seamless Switching	🟢	Automatically follows user's language change. No explicit switch mechanism needed.
Knowledge Compatibility	🟡	Translation quality depends on the LLM. Complex technical content may lose nuance in translation.

Limitations¶

Short input ambiguity: Single-word or very short messages can be misidentified. "OK," "Merci," "Hello" — some overlap between languages.
First-message cold start: The first message determines the language. If the user's first message is an emoji or a single word, detection may be wrong.
No proactive language: The agent can't greet the user in their language because it doesn't know the language until the user types something.
Translation quality: Having the LLM translate knowledge source content is impressive but not flawless. Technical terminology, legal language, or policy-specific terms may be mistranslated.
No explicit variable: Without the optional Prompt Tool step, the language is implicit in the LLM's behavior. You can't use it for routing decisions or conditional logic in topics.

Approach C: Explicit User Language Selection¶

Summary: Ask the user to choose their language at the start of the conversation (first time only). Persist the choice via Gem 001 and apply it to all future sessions.
Technique: Question node with language options, global variable, persistent storage, agent instructions.

How It Works¶

First conversation:
  Agent: "🌐 Please select your preferred language:
          🇬🇧 English | 🇫🇷 Français | 🇩🇪 Deutsch | 🇯🇵 日本語"
  User selects: Français
  → Saved to persistent storage
  → All responses in French

Second conversation (and all future):
  → Language loaded from storage
  → Agent responds in French immediately
  → No question asked

Maximum accuracy (the user chose it), but adds friction to the first interaction. Subsequent sessions are frictionless because the preference is persisted.

Implementation¶

Step 1: Create a language selection topic

kind: AdaptiveDialog
beginDialog:
  kind: OnRecognizedIntent
  id: main
  intent:
    displayName: Set Language Preference
    triggerQueries:
      - "change language"
      - "switch language"
      - "changer de langue"
      - "言語を変更"
  actions:
    - kind: Question
      id: askLanguage
      variable: init:Topic.SelectedLanguage
      prompt: "🌐 Please select your preferred language / Veuillez choisir votre langue :"
      entity: ChoicePrebuiltEntity
      choiceOptions:
        - value: "English"
          synonyms: ["EN", "english", "anglais"]
        - value: "French"
          synonyms: ["FR", "french", "français", "francais"]
        - value: "German"
          synonyms: ["DE", "german", "deutsch", "allemand"]
        - value: "Japanese"
          synonyms: ["JA", "japanese", "日本語", "japonais"]
        - value: "Spanish"
          synonyms: ["ES", "spanish", "español", "espagnol"]

    - kind: SetVariable
      id: setLanguage
      variable: Global.ResponseLanguage
      value: =Topic.SelectedLanguage

    # Persist for future sessions
    - kind: InvokeFlow
      id: saveLanguage
      flowId: "@environmentVariables('WriteContextFlowId')"
      inputs:
        userId: =System.User.Id
        preferredLanguage: =Global.ResponseLanguage

    - kind: SendActivity
      id: confirmLanguage
      activity:
        text:
          - "✅ Language set to **{Global.ResponseLanguage}**. All future conversations will use this language.\n\nYou can change it anytime by saying \"change language\"."

Step 2: Auto-prompt on first conversation only

In agent instructions:

instructions: |+
  ## Language Handling

  At the START of every conversation, call "GetUserContext" to check if a 
  language preference exists.

  - If `preferredLanguage` is set → use it immediately, DO NOT ask
  - If `preferredLanguage` is empty (first-time user) → ask the user to choose 
    by calling the "SetLanguagePreference" topic

  ## Response Language
  ALWAYS respond in Global.ResponseLanguage.

  ## Language Change
  If the user says "change language" or equivalent in any language,
  trigger the SetLanguagePreference topic to let them choose again.

Step 3: Combine with ConversationStart for non-M365 channels

kind: AdaptiveDialog
beginDialog:
  kind: OnConversationStart
  id: main
  actions:
    - kind: InvokeFlow
      id: loadContext
      flowId: "@environmentVariables('ReadContextFlowId')"
      inputs:
        userId: =System.User.Id
      outputVariable: Topic.UserContext

    - kind: SetVariable
      id: setLang
      variable: Global.ResponseLanguage
      value: =Topic.UserContext.preferredLanguage

    - kind: ConditionGroup
      id: checkLanguageSet
      conditions:
        - id: hasLanguage
          condition: =!IsBlank(Global.ResponseLanguage)
          actions:
            - kind: SendActivity
              id: greetInLanguage
              activity:
                text:
                  - =If(Global.ResponseLanguage = "French", "Bonjour ! Comment puis-je vous aider ?", If(Global.ResponseLanguage = "Japanese", "こんにちは！何かお手伝いできますか？", "Hello! How can I help you?"))
      elseActions:
        # First-time user — ask for language
        - kind: GotoTopic
          id: gotoLanguageSelect
          topicId: SetLanguagePreference

Evaluation¶

Criterion	Rating	Notes
Ease of Implementation	🟢	Question node + persistence flow. Simple and clear.
Maintainability	🟢	Choice list is easy to update. Adding a language = adding a choice option.
Channel Compatibility	🟢	Multiple choice questions work in all channels.
Detection Accuracy	🟢	100% accurate — the user chose it explicitly.
Seamless Switching	🟢	"Change language" topic allows switching anytime.
Knowledge Compatibility	🟢	Known language enables explicit translation instructions.

Limitations¶

First-session friction: New users must answer a language question before anything else. For an agent that serves 90% English-speaking users, this is unnecessary for most people.
Choice list maintenance: Adding less-common languages expands the choice list, potentially overwhelming users. Consider showing only languages relevant to your org.
No detection of unexpected languages: If a user writes in Korean but Korean isn't in your choice list, the agent responds in whatever the user previously selected.
ConversationStart limitation: The language selection prompt doesn't fire in M365 Copilot unless handled via agent instructions.

Comparison Matrix¶

Dimension	Approach A: Graph API	Approach B: LLM Auto-Detect	Approach C: Explicit Choice
Implementation Effort	🟡 Medium (2-3 hours)	🟢 Low (15 min)	🟢 Low (1 hour)
Detection Accuracy	🟡 Depends on IT data	🟡 Good for sentences, weak for short input	🟢 100% (user chose it)
First-Session Friction	🟢 Zero (automatic)	🟢 Zero (automatic)	🔴 Language question before first use
Mid-Conversation Switch	🟡 Via instructions only	🟢 Automatic	🟢 "Change language" command
Proactive Greeting	🟢 Greet in detected language	🔴 Can't greet (don't know language yet)	🟢 Greet in saved language
Infrastructure	🟡 Power Automate + Graph API	🟢 None	🟡 Persistence flow (Gem 001)
Best When...	IT maintains Entra profiles well	Quick deployment, LLM handles it	Maximum accuracy, users accept one-time setup

Recommended Approach¶

For most scenarios: Combine Approaches B + C with Gem 001 persistence.

First session: Use Approach B (LLM auto-detect) — the agent silently detects the user's language from their first message and responds accordingly. After detection, persist the language using Gem 001.
Subsequent sessions: Load the persisted preference. The agent greets and responds in the saved language immediately.
Override: Provide Approach C's "change language" topic for users who want to explicitly switch.

This gives you: zero friction (no question asked), reasonable accuracy (LLM detection), persistence (no re-detection), and an escape hatch (explicit switch).

Add Approach A when: Your IT organization maintains Entra profiles well (preferredLanguage is populated). Use it as the primary signal with LLM detection as fallback for users without profile data.

Use Approach C alone when: You support fewer than 3 languages and/or regulatory requirements demand the user explicitly confirms their language choice.

Platform Gotchas¶

Warning

ConversationStart doesn't fire in M365 Copilot (see Gotchas Compendium).
If you load language in ConversationStart, it only works in Web Chat and Teams. For M365 Copilot, use agent instructions: "Before responding, load the user's language preference."

Warning

LLM translation quality varies by language pair.
English → French/Spanish/German: excellent. English → Japanese/Korean/Arabic: good but may miss nuance in technical or legal terminology. For high-stakes content (HR policies, legal notices), consider human-translated knowledge sources in each language rather than LLM translation.

Note

Generative answers search works cross-lingually.
Copilot Studio's SearchAndSummarizeContent can find English documents when the user asks in French. The semantic search layer handles this. The challenge is the response language — use customInstructions to mandate: "Always respond in the user's language."

Note

System.User.Language does not exist.
There is no built-in system variable for user language in Copilot Studio. You must detect or store it yourself.

Gem 001: Persisting User Context Across Sessions — Provides the persistence layer for language preference
Gem 002: Persona-Adaptive Agent Instructions — Language is a facet of persona. Combine language adaptation with persona-specific tone and vocabulary.

References¶

Gem 005 | Author: Sébastien Brochet | Created: 2026-02-17 | Last Validated: 2026-02-17 | Platform Version: current

Gem 005: Multi-Language Agent Response¶

Classification¶

The Problem¶

The Ideal Outcome¶

Approaches¶

Approach A: Graph API Profile-Based Detection¶

How It Works¶

Implementation¶

Evaluation¶

Limitations¶

Approach B: LLM Auto-Detection from User Input¶

How It Works¶

Implementation¶

Evaluation¶

Limitations¶

Approach C: Explicit User Language Selection¶

How It Works¶

Implementation¶

Evaluation¶

Limitations¶

Comparison Matrix¶

Recommended Approach¶

Platform Gotchas¶

Related Gems¶

References¶