How to create an email triage assistant in Microsoft Copilot

João Barros 23 de September de 2026 4 min read

This tutorial shows how to create an email triage assistant in Microsoft Copilot to help classify, label and suggest automatic actions on incoming messages — useful to reduce inbox management time and ensure consistent responses. The idea is to explain the purpose of the rules and present a practical example of an assistant that identifies priority, request type and proposes reply drafts.

Prerequisites

Microsoft account with access to Microsoft Copilot integrated in Microsoft 365.
Access to Outlook (web or desktop) and permissions to connect Copilot to email data.
Basic concepts of prompt engineering: clear instructions and examples.
A set of 10–20 example emails to train/tune prompts (anonymized).

Step 1: Define triage objectives

Before writing prompts, clarify what the assistant should do: classify priority (High/Medium/Low), identify type (Support/Finance/Sales/Informational), suggest labels and create a reply draft when appropriate. Defining these objectives helps create clear instructions for Copilot and evaluate results.

Step 2: Prepare representative examples

Gather 10–20 example emails that cover variations: urgent requests, technical questions, invoices, newsletters, acceptable spam. For each email, note the desired classification and a reply draft if applicable. These examples serve for few-shot prompts or to tune the assistant’s behavior.

Step 3: Create the base prompt for classification

Write a clear prompt that explains what Copilot should return: structured labels in JSON (useful for automation). Include examples in the prompt itself (few-shot). Basic example you can paste into Copilot Studio or the Copilot interface:

You are an email triage assistant. For the provided email, return a JSON with: {
  "priority": "Alta|Média|Baixa",
  "type": "Suporte|Financeiro|Comercial|Informativo|Outro",
  "labels": ["label1","label2"],
  "suggested_action": "responder|encaminhar|arquivar|pedir mais info",
  "reply_draft": "(optional short draft)"
}

Example 1:
Email: "Tenho um problema com a minha factura de Abril, foi cobrado em duplicado. Preciso de reembolso urgente."
Output: {"priority":"Alta","type":"Financeiro","labels":["Factura","Reembolso"],"suggested_action":"responder","reply_draft":"Lamento o incómodo. Vou verificar a factura de Abril e contacto para confirmar o reembolso até 24h."}

Now process the following email:
Email: "[cole aqui o email a analisar]"

Step 4: Integrate with Outlook and automate actions

Use the Copilot integration in Outlook or Power Automate to apply labels and create drafts. If using Power Automate, create a flow that sends the email body to Copilot (via the appropriate connector) and apply the JSON outputs as conditions for actions (move, label, create draft).

// Exemplo conceptual (pseudo) para Power Automate actions
Trigger: When a new email arrives (Outlook)
Action: Call Copilot API/connector with prompt and email body
Condition: if result.priority == "Alta" then
  Move email to folder "Prioridade Alta"
  Create draft reply with result.reply_draft
Else if result.suggested_action == "arquivar"
  Move to Archive

Step 5: Validate and adjust prompts (common errors)

Test with the 10–20 example emails. Common errors: responses too generic, inconsistency in JSON format, labels outside the vocabulary. To fix, provide more examples (few-shot), constrain the vocabulary in the prompt and ask for format validation (for example: "Return only valid JSON"). Also adjust for the desired language and tone (formal/informal).

Check the result

Verify that, for a sample of emails, the assistant returns valid JSON and that automatic actions (move, label, create draft) occur as expected. Test edge cases (spam, multilingual emails, attachments). Confirm that reply drafts are short, correct and do not contain inappropriate sensitive data.

Conclusion

With a well-structured prompt and integration with Outlook/Power Automate, you can build an email triage assistant in Microsoft Copilot that reduces inbox management time and improves consistency. Next steps: add continuous learning (user feedback) and expand the label vocabulary. Tip: start with simple rules and scale to more complex prompts as you validate results.