Checking human handoff when an answer is uncertain

An uncertain answer needs a route that transfers the conversation, not just a “Contact support” prompt. Run each test from the customer’s message through to the receiving queue; check that the customer does not have to repeat the problem and the agent can see the conversation and why it was escalated.

In Khoros Flow, check the three documented handoff triggers: Text, when a user asks for an agent; Unknown, when the bot does not recognise the input; and Feedback, when the user says the bot did not answer. Also test cases where the script is broken, the request is unidentified or a resolution is not found.

Try a missing answer, ambiguous description, policy exception and account-specific action. Check that the bot asks a useful clarifying question where needed, does not invent an answer and offers a reachable handoff. Flow can gather required details such as an email address, date or specific order number before redirecting the question.

Check that the handoff procedure redirects the question to the applicable department or workflow. In Khoros Flow, Handoff Reply sends the conversation to a customer service tool; tickets can be created through Zapier or custom tickets and integrations using APIs.

Test the route against channel and availability rules. WhatsApp and Facebook Messenger allow messaging within 24 hours, while live chat has higher priority than asynchronous messaging; during business hours an available live agent can follow up, and outside business hours a ticket can be created in a CRM for later follow-up.

A passing test record identifies the customer’s question, the trigger (such as Unknown), the relevant source if one was used, and details gathered, such as an order number. A failing record says only “Contact support” and leaves the agent without the conversation or reason for escalation.

Open the receiving tool as an agent. Confirm the conversation is visible, it reaches the applicable department or workflow, and a human can take over; the agent should be able to inspect the conversation without treating a possibly wrong AI conclusion as fact. Fail the test if context is missing, no human can pick it up or the customer must repeat the problem.

In Zendesk, distinguish customer-facing AI agents from agent-facing auto assist. Auto assist suggestions appear in the ticket interface and require agent approval before being sent; auto assist does not work on AI agent tickets, so approving a suggestion is not proof that a customer-to-human handoff works.

Zendesk AI agents use connected knowledge sources, including Zendesk help centres and external sources. For restricted help-centre content, their responses respect article permissions: authenticated customers may receive relevant restricted articles, while unauthenticated customers can receive only public articles.

For each test, record the trigger, intended destination, whether the conversation and details arrived, and the failure cause. Use the existing causes—missing source, routing, permissions or unclear wording—to identify what needs fixing before increasing the share of questions answered autonomously.

AI Agent vs Auto Assist in Zendesk

Handoff Testing Requirements