You add a plugin tool to a Copilot Studio agent that can retrieve customer account balances. Despite being configured correctly, the generative orchestrator rarely invokes the tool when users ask about their balance. What is the most likely configuration issue?
Select an answer to reveal the explanation.
Short Explanation
The orchestrator decides what to call by reading the tool description, so if it's vague or too generic, it won't know to call it for balance questions. You need a specific description that says exactly what the tool retrieves.
Full Explanation
When Copilot Studio's generative orchestrator decides which tools, topics, or child agents to invoke for a user's request, it uses the natural language description of each registered tool to determine relevance. If the tool's description is vague (for example, 'retrieves information') rather than specific (for example, 'Retrieves the current account balance, available credit, and last payment date for a customer account given the customer ID'), the AI cannot match it to user queries about balances. The fix is to write a detailed, specific tool description that names the exact use cases, the data it returns, and the inputs it needs. This is called 'tool prompt engineering' and is critical for reliable generative orchestration. The missing API key option is wrong because if the API key were missing, the tool would fail when called, returning an error, not fail to be selected; the symptom here is non-invocation, not invocation failure. The tools-only-from-topics option is wrong because tools can be discovered and invoked by the generative orchestrator directly when their descriptions match user intent. The explicit tool name option is wrong because generative orchestration is intent-based, not name-based, and users should never need to know tool names. Exam caveat: poor tool description means the orchestrator does not invoke the tool; invest time in clear, specific tool descriptions. This is analogous to writing good system prompt tool descriptions for function calling in LLM APIs.