FAQ

Why can't I see the option to create an agent?

To create a custom communication agent, contact your account manager for evaluation and activation. The option becomes visible after activation.

The template looks close but doesn't fully fit my use case. Can I go live directly?

No. First copy the template, replace branding and business-specific content, fill in variables and edge cases, then complete debugging, regression testing, and small-sample validation.

Do I need to manually select a branch and version when creating a calling task?

No. When creating a task, you only need to select the communication agent. When each call record enters the scheduler, the system automatically uses the active branch and its latest published version for that agent.

Versions in draft, under review, or review-rejected status, versions in unpublished branches, and historical versions are never used automatically by default calling.

Why do execution records for the same calling task show multiple branches or versions?

This is expected behavior. When a task is created and contacts are imported, versions are not locked for all numbers at once. Each record confirms the published branch and latest published version only when it is actually scheduled.

As a result, if you publish a new version, switch the active branch, or pause and resume the task during execution, records scheduled at different times may use different versions. When troubleshooting or reviewing results, refer to the Branch and Version columns in the execution records rather than just the task name.

Which version is used when a task resumes after the agent was modified during a pause?

Pausing does not update records that are still in pending-call status. When the task resumes and these records re-enter the scheduler, they use the latest published version of the active branch at that time.

If you want resumed records to use the new version, confirm that the new version has passed review and the target branch is active before resuming the task. Conversely, if you want all records in a batch to use the same version, complete publishing, review, and deployment before starting the task, and do not switch the active branch or publish a new available version on the active branch until all pending-call records finish. Pausing a task alone does not freeze the version.

What is the difference between "Recall" in post-call processing and "Retry on no answer" in calling tasks?

The two features serve different scenarios:

Contacts have been imported, but records stay in "Pending" or go to "Failed". How do I troubleshoot?

"Pending" usually means the record is waiting to be scheduled, waiting for a callable time window, or waiting for the next retry. Check whether the task is paused, whether the call plan and time window are configured correctly, and whether the record is in a retry wait period.

"Failed" means the record could not be executed successfully. Open the record to check the failure reason, and pay particular attention to contact list fields and phone numbers, the line or service instance, the task scheduling configuration, and agent dependencies. Do not judge the issue based solely on the overall task status, and do not re-import contacts before the root cause is confirmed.

Why is the SMS not sent even though the rule condition is met?

The SMS action depends on the SMS service and access authorization. Enable the SMS service first, then click Authorize in the post-call processing configuration to complete authorization. Next, verify that the rule is enabled, the trigger condition is matched, the SMS template is available, and all template variables have values.

It is recommended to use a controlled test record to fully verify the end-to-end flow: condition match, template rendering, and SMS delivery.

Why does the agent fail to transfer to a human agent even though it is configured?

First confirm that the human agent line configured in the communication agent is enabled, then check the tool calling instructions and verify that the prompt or conversation flow clearly specifies when to transfer to a human agent and when not to.

Transferring to a human agent depends on both the model's judgment to invoke the transfer and the availability of the line. Before going live, verify successful transfer, fallback handling, and caller experience through call testing.

How do I choose between DTMF key collection and DTMF key playback?

Use DTMF key collection when the contact needs to enter options, numbers, or verification codes via the phone keypad. Use DTMF key playback when the agent needs to send keypad inputs to the remote IVR menu to navigate through it.

The two features work in opposite directions. Use playback only when the remote party explicitly requests key input. Valid key values are 0–9, *, and #. After sending, wait for the next IVR prompt. Do not use key playback as a substitute for collecting contact input.

Why is the agent not citing knowledge base content correctly, or why do answers change after going live?

The knowledge base provides references to the agent through RAG and does not produce fixed, word-for-word responses. First confirm that the materials have been processed and enabled, that the content includes key expressions from customer questions, and that the prompt or node explicitly defines when to retrieve and how to handle cases where no content is found.

The knowledge base is a reusable resource. Modifying shared materials may affect other agents that reference it. Before updating high-risk content such as pricing, commitments, or policies, confirm the scope of references and complete text testing and call testing with real questions.

Does AI self-iteration directly modify live scripts or publish automatically?

No. AI self-iteration only assists in analyzing failed cases and generating optimization suggestions. Human review of the changes is still required, along with reproducing the original issue, regression testing related cases, publishing, review, deployment, and small-sample validation.

For live anomalies, start from the task details to access the corresponding call record, and confirm the branch, version, variables, conversation log, recording, tags, and failure reason that were actually used. Then reproduce and fix the issue in the corresponding branch. Do not judge solely based on the current draft, and do not use unverified modifications for calling.

For related operations, see: Branches, versions, and iterative release, Call task: create, execute, and analyze, Post-call actions configuration, and System tools configuration.

Text testing passes, but why is the call test experience still poor?

Text testing only verifies content and logic. In call testing, focus on the opening line, voice tone, speech rate, pauses, ASR recognition, interruption handling, silence handling, and the experience before and after tool calling.