To implement the transfer rules for the scenarios you described, when the chatbot does not understand the question, when the request falls outside the knowledge base scope, or when no answer is found, the most appropriate place is in the Guidelines. Guidelines are operational instructions that tell the agent what to do during customer interactions, and this is where you define the decision logic, for example: "When searching the KB, if the confidence level is low or if no results are returned, trigger the transfer tool to route the chat to a human agent." This approach allows the agent to interpret the context and take the correct action dynamically and intelligently.
GuardRails, on the other hand, serve a different purpose. They establish rigid and inviolable boundaries that the agent can never cross, regardless of the conversation context. They are safety and compliance rules, such as "never share sensitive customer data" or "never provide medical diagnoses." They are not designed to manage operational workflows like transferring a chat, but rather to ensure the agent operates within ethical, legal, or corporate limits. Therefore, for your use case, Guidelines are the correct tool, while GuardRails would act as a protective layer over these rules without replacing them.
Regarding the priority between Role and Guideline, it is important to clarify that in most platforms, both operate within the space of flexible instructions, meaning the agent is trained to interpret them together without a strict hierarchy between them. The Role defines the assistant's identity and tone, for example, "you are a polite and product-specialized attendant," while the Guidelines detail the procedures and workflows. Since the two are complementary, one does not carry more weight than the other; they combine to shape the agent's personality and expected behavior.
The priority conflict that actually exists and is documented is between Instructions/Guidelines and GuardRails. In this case, GuardRails always take precedence. If a Guideline says "answer all customer questions," but a GuardRail determines "never reveal internal company information," the GuardRail will be obeyed when the sensitive topic arises. For your implementation, you will need to configure the KB tool with a confidence threshold and define in the Guidelines that, below that level, the agent should call the transfer function. This way, the workflow remains clear, secure, and aligned with best practices for virtual agent platforms.
------------------------------
Camila Meneghini
------------------------------