💡 Keep in mind: Write it once, where it belongs. When the same rule sits in two fields, the two versions can drift apart.

💡 Keep in mind: 

• Messages are spoken word for word, in the language active at that moment. The model doesn't add to them and doesn't remember them as part of the conversation.
• Don't script a greeting or goodbye in the instructions. The caller would hear it twice.
• Keep the Final Message neutral, e.g. "Thank you, one moment please." It plays after the outcome is chosen, so it can't know what happens next in the Workflow.

💡 Keep in mind: 

• Start language: the Language field sets the language the Virtual User starts in.
• Switching: it changes language only when the caller explicitly asks, e.g. "Can we continue in French?". It never switches because of an accent or a foreign word.
• No own language rules: all profiles handle switching already. Additional rules compete with them. If you want callers to choose, add one line such as "After the greeting, ask whether the caller prefers English or French."
• Voice: stays the same for the whole call. With Azure Voice Live Bots, the voice list depends on the selected Language.

✅ Three-sentence template for the tool instructions:

1. What it does — "Looks up an order by order number in the e-commerce backend."
2. When to call it — "Call this whenever the caller asks about the status, delivery date, or contents of a specific order."
3. What you'll get back — "Returns the order status, expected delivery date and line items, or an empty result if the order number is unknown."

💡 Keep in mind: 

• Tools must exist first: only mention tools that are actually added. A rule about a missing tool confuses the model and may lead to failed workflow outcomes.
• One tool, one job. A customer_lookup and a separate order_lookup are far easier for the model to reason about than one omnibus customer_or_order_lookup.
• Exact names: the model sees names in lowercase with spaces replaced by _ and special characters removed ("Check Balance!" becomes check_balance). Use that form in your instructions.
• Tool Instructions are what the model sees. Each Web Request and MCP Server added to a Virtual User has its own Instructions field. The model does not see the generic Description of the Web Request or MCP Server configuration item.
• Read-only first. If a tool can both read and write, start read-only and add write capability only once the read path is validated.
• Confirm before destructive actions. Encode "confirm with 'yes' before you make the change" rules in the Outcomes & Tools Instructions for any write-capable tool.
• Parameters: attach up to 10, and only what the next Workflow step needs. Give each Parameter its Virtual User Instructions. A warning icon shows on outcomes where they are missing.
• Outcome Name: a short verb phrase, e.g. route_to_billing, up to 64 characters. Names must be unique across outcomes and tools.
• Failure handling lives in the activity, not the prompt. Tool errors take the Failed exit of the Add Virtual User activity — you don't need to write fail-safe branches into every instruction step. Reserve the prompt for graceful behavior (apologize, offer a transfer) once a tool returns an empty result.
  • Exit or not: enable This outcome triggers an exit for results that end the call. Leave it off for steps that only collect or return information.
  • AI Fallback: always added automatically. It is taken when nothing fits the request, the caller stays off topic, or data collection keeps failing.

💡 Keep in mind: 

• When one Virtual User's exit leads to another later in the Workflow, the second one starts a brand-new conversation.
• Good fit: clearly separate phases (identify the caller, then handle the request), or a non-AI step such as a queue in between.
• Avoid: splitting one task just to keep instructions short. The pause and the lost context usually cost more. One Virtual User with more outcomes often does the job.
• If you chain: save everything later steps need as Parameters, keep later greetings short, and test the whole chain end to end. Each Virtual User needs its own license.

💡 Keep in mind: When the Add Virtual User Conversation Handling Activity takes the Failed exit, check:

• Data Privacy Tenant Settings : the Virtual User is allowed for the tenant and its data privacy consent is given.
• License Management: the Virtual User has an active license assigned on its General tab.
• Profile and Bot: the Bot supports the profile, e.g. Open Profile needs a Custom AI Service Bot.
• Exits: every exit outcome of the Virtual User is mapped on the activity. 
  💡 Calls that end early or lose their connection to the AI service also take the Failed exit.