Choose a Zendesk AI chatbot around support ownership
Evaluate a Zendesk AI chatbot with separate checks for policy answers, triage, authorized lookups, human handoff and resolutions your support team can inspect.
Separate the support jobs
Zendesk describes AI agents that answer questions from trusted knowledge sources and take actions in authorized systems. Evaluate a Zendesk AI chatbot by choosing which support jobs should finish automatically and which should reach a person.
A policy answer, ticket triage, a current account lookup and a discretionary support decision are different jobs. Define the approved scope for each. A customer asking where to find a returns form may need an answer, while someone disputing a refused return needs an accountable person. Agree these boundaries with the support lead before comparing vendor proposals or treating every conversation as eligible for automatic resolution. Use separate pilot labels for 'returns link supplied', 'return exception awaiting agent' and 'merchant decision communicated' so triage cannot inflate the resolved count.
Design a handoff staff can act on
Zendesk defines messaging handoff as changing the conversation's first responder from the AI agent to a live agent. In an evaluation, ask the provider to demonstrate the receiving team's view and how it recognizes a request awaiting attention.
The handoff should explain the unresolved question, relevant collected details and any lookup or action already attempted. Identify the person or queue taking responsibility and the fallback when nobody is available. Inspect the request from the recipient's perspective, including a customer who changes their account detail during the conversation. A transfer notification alone does not show that staff have enough context to help. Have the receiving agent locate the customer's cited email and requested deadline exception without replaying the conversation or asking the customer to explain everything again.
Rehearse conflicting answers and exceptions
Hypothetical example: a customer asks support to waive a return deadline after receiving an older policy in an email. The support owner should decide which policy applies and whether an exception is allowed. The assistant's ability to locate policy text does not grant authority to waive it.
Evaluate an approved question, missing information, conflicting sources and a request for a restricted decision. For account-specific information, establish identity and authorized access before a lookup. Keep the explanation given to the customer consistent with the actual state: waiting for clarification, awaiting staff review or completed. Do not count an unanswered exception as resolved because the conversation closed. Include 'Your email says I had longer to return this' as a test question, and check whether the old email date reaches the person deciding the exception.
Measure resolutions over a defined window
Choose the eligible support requests, the source of outcome evidence and the reporting window before the pilot. Report answered policy questions, escalations and completed resolutions separately. Include requests that return to staff after an apparently successful interaction so the team can understand the full workload.
For each unresolved case, identify its owner and next action. If a proposed action times out, reconcile the receiving record before retrying. Compare another proposal against those same handoff and inspection requirements, including the ongoing maintenance each team will perform. Reopen the same disputed-return case after an initial policy answer and track the reopened ticket alongside the original conversation instead of counting two independent successes.
Review support exceptions in Vora
Vora includes a native staff inbox and customer records for agencies evaluating support handoffs. Carry the disputed return deadline, policy revision and outstanding waiver decision into the support brief, then verify any external ticketing destination independently. Use Vora's early-access waitlist to propose that exception-handling evaluation.
Common questions
Which support questions should launch first?
Choose a narrow category with approved answers and clear completion evidence. Keep disputed policies and requests needing discretionary decisions with the responsible support people during the pilot.
What makes a handoff actionable?
The recipient needs the unresolved question, relevant details, prior attempts and ownership. Check that information in the staff view rather than relying on the customer's transfer message.
How should triage be measured?
Measure whether the request reached the right owner with usable context. Report routing separately from the eventual resolution so a correctly assigned but unfinished request remains visible.