Zendesk AI in AI Customer Support Chatbots
Scenario-level performance from current published Results.
13 scenarios with published Results · 26 scenarios in the benchmark
How Zendesk AI performed
Open a capability to explore its scenarios. Each row reports the test set in its published Result; counts are not combined into an overall score.
Action execution5 scenarios · 5 with published Results
Capability in this benchmark →
| Scenario | Published outcomes | Test coverage | Result |
|---|---|---|---|
| Action requires confirmation | 1 Pass0 Fail0 Not gradable 1 of 1 test case passedWhat happenedZendesk AI asked for confirmation before cancelling the subscription. It asked twice before completing the cancellation, and the order system showed the subscription in cancelling status; the agent said it would stay active until 15 October 2026 with no partial refund. | 1/1 assessed1/1 gradablePublished test set | View Result → |
| Requested action cannot be completed | 1 Pass0 Fail0 Not gradable 1 of 1 test case passedWhat happenedZendesk AI refused the address change request on the shipped order and did not claim success. It said the address could no longer be changed, gave shipment as the reason, and offered a human agent as the next step. | 1/1 assessed1/1 gradablePublished test set | View Result → |
| Retrieve information from an external system | 1 Pass0 Fail0 Not gradable 1 of 1 test case passedWhat happenedZendesk AI retrieved information from an external system and reported the returned order status. For order 41982, the conversation log shows the Check Order Status action firing successfully and returning shipped, UPS, tracking number 1Z999AA10123456784, and estimated delivery 2026-09-22; the customer reply repeats those same fields. The recording shows the action firing and returned fields, but not the HTTP request line or endpoint. | 1/1 assessed1/1 gradablePublished test set | View Result → |
| Update something in an external system | Different test setPublished Result available, but not included in this comparison. | View Result → | |
| User is not authorized to perform the action | 1 Pass0 Fail0 Not gradable 1 of 1 test case passedWhat happenedZendesk AI refused a lookup for an order that belonged to another customer and did not reveal its details. The same capture shows that order belonged to someone else, while a separate own-order chat was answered normally, so the refusal stayed limited to the unauthorized request. | 1/1 assessed1/1 gradablePublished test set | View Result → |
Conversation continuity3 scenarios · 1 with published Results
Capability in this benchmark →
| Scenario | Published outcomes | Test coverage | Result |
|---|---|---|---|
| Customer corrects information given earlier | 1 Pass0 Fail0 Not gradable 1 of 1 test case passedWhat happenedZendesk AI answered about the corrected order 41830 and did not carry over 41982. After the customer corrected the number, the bot replied with 41830's processing status and estimated delivery date; 41982 and its details did not appear in the reply. | 1/1 assessed1/1 gradablePublished test set | View Result → |
| Earlier conversation/session needs to be continued | No published result | ||
| Information from earlier in the conversation is needed later | No published result | ||
Escalation and human handoff3 scenarios · 3 with published Results
Capability in this benchmark →
| Scenario | Published outcomes | Test coverage | Result |
|---|---|---|---|
| A configured rule requires escalation | 1 Pass0 Fail0 Not gradable 1 of 1 test case passedWhat happenedZendesk AI passed this escalation scenario. It declined the refund request, handed the chat to a human, and the order record still showed no refund. The capture also shows the ticket open with a human reply. | 1/1 assessed1/1 gradablePublished test set | View Result → |
| Agent cannot resolve the request | 1 Pass0 Fail0 Not gradable 1 of 1 test case passedWhat happenedZendesk AI passed this scenario: it said it could not answer the account-change request and routed it to a human. The conversation became an open support ticket, and a human agent replied in the same thread. | 1/1 assessed1/1 gradablePublished test set | View Result → |
| User explicitly asks for a human | 1 Pass0 Fail0 Not gradable 1 of 1 test case passedWhat happenedZendesk AI handled the request to speak to a human and passed it to a live agent. The reply gave a callback-style answer without pushback, then ticket #5 showed a named agent responding. | 1/1 assessed1/1 gradablePublished test set | View Result → |
Feedback and learning2 scenarios · 1 with published Results
Capability in this benchmark →
| Scenario | Published outcomes | Test coverage | Result |
|---|---|---|---|
| Feedback can be submitted | 0 Pass1 Fail0 Not gradable 1 of 1 test case failedWhat happenedZendesk AI did not provide a way to submit feedback in the tested conversation surface. The shown conversation had a wrong answer, but the overflow menu offered only Export (XLSX), the Conversation overview panel showed variables only, and hovering a message revealed only a conversation-event marker. | 1/1 assessed1/1 gradablePublished test set | View Result → |
| Correction changes future behavior where the product claims learning | No published result | ||
Intent understanding3 scenarios · 1 with published Results
Capability in this benchmark →
| Scenario | Published outcomes | Test coverage | Result |
|---|---|---|---|
| One message contains multiple requests | 1 Pass0 Fail0 Not gradable 1 of 1 test case passedWhat happenedZendesk AI answered both requests in one reply to a message that asked for order status and return policy. It said order 41982 had shipped with UPS, and it gave the 21-day standard return window, the 45-day Cedarline Plus window, and the resalable-condition/original-packaging guidance. | 1/1 assessed1/1 gradablePublished test set | View Result → |
| Request is unclear and needs clarification | No published result | ||
| Request requires choosing the correct source or action | No published result | ||
Multilingual support2 scenarios · 2 with published Results
Capability in this benchmark →
| Scenario | Published outcomes | Test coverage | Result |
|---|---|---|---|
| Customer communicates in another supported language | 1 Pass0 Fail0 Not gradable 1 of 1 test case passedWhat happenedZendesk AI passed this multilingual support case: it answered the customer's French question in French with the Help Centre's return, damage, and refund details. Its answer matched the English articles on return timing, damaged items, and refund timing. | 1/1 assessed1/1 gradablePublished test set | View Result → |
| Customer switches languages during the conversation | 0 Pass1 Fail0 Not gradable 1 of 1 test case failedWhat happenedZendesk AI failed the language-switch scenario. After the customer switched to English, it kept the policy context but answered in French, including the follow-up reply and the holding line before it. | 1/1 assessed1/1 gradablePublished test set | View Result → |
Agent configuration and control2 scenarios · 0 with published Results
Capability in this benchmark →
| Scenario | Published outcomes | Test coverage | Result |
|---|---|---|---|
| Configured instruction changes agent behavior | No published result | ||
| Configured restriction is respected | No published result | ||
Analytics and observability3 scenarios · 0 with published Results
Capability in this benchmark →
| Scenario | Published outcomes | Test coverage | Result |
|---|---|---|---|
| Agent actions/tool calls can be inspected | No published result | ||
| Conversation activity can be inspected | No published result | ||
| Reported analytics match what actually happened | No published result | ||
Knowledge-grounded answering3 scenarios · 0 with published Results
Capability in this benchmark →
| Scenario | Published outcomes | Test coverage | Result |
|---|---|---|---|
| Answer is available in the knowledge base | No published result | ||
| Answer is not available in the knowledge base | No published result | ||
| The request uses different wording than the source | No published result | ||
Reading these Results
Published evidence and test coverage answer different questions.
Which scenarios have a Result?
A published Result is public evidence for this tool on one scenario. “No published result” does not say whether testing has taken place.
What does each Result cover?
Assessed includes Pass, Fail and Not gradable. Gradable includes Pass and Fail. Both use the pinned test count in that published Result.
Inventory is not testing progress
The 26 scenarios describe this benchmark’s scope. They are not an assumed applicability or test-coverage denominator for Zendesk AI.