After go-live, the Companion Care Group service agent answers "can I move my appointment to Thursday" with a friendly but generic explanation of the rescheduling policy and never actually changes the booking. The Appointment Management topic is being selected correctly. What is the most likely cause?
Select an answer to reveal the explanation.
Short Explanation
The agent is talking because talking is all it can do. If a topic has no action wired to it, the reasoning engine can only generate language from its instructions, so you get a polite pamphlet instead of a rescheduled surgery slot. Actions are what make an agent do things rather than describe things.
Full Explanation
An agent can only take an action that has been assigned to the currently selected topic. When a topic carries instructions but no action, the reasoning engine has nothing to invoke and falls back to generating a natural-language answer from its instructions and grounding, which is precisely the symptom described: correct routing, polite prose, no record change. The fix is to assign an action backed by a Flow or an invocable that updates the appointment record. The role and company description are wrong because they shape persona, not capability, and the topic is already being selected so routing is demonstrably fine. Trust Layer masking is the most attractive distractor because masking genuinely does replace sensitive values before they reach the model, but masking is applied to personally identifiable data and would not silently disable a write operation; it would at worst affect how a value is echoed back. An over-broad scope statement can cause the agent to answer things it should not, but it never suppresses an action that exists. Worth knowing for the exam: instructions describe how to behave, actions are what the agent can actually do, and the two are configured separately.