Call task analysis
Locate call issues through execution records, call details, and key metrics, then save real problems as test cases to drive continuous improvement of the communication agent.
The four-stage analysis loop
Call task analysis follows a four-stage loop that turns production issues into regression coverage:
Narrow the execution records — Filter batch results to surface representative problem calls.
Inspect individual call details — Review labels, variables, recordings, and node transitions to confirm the root cause and decide on optimizations.
Match symptoms to actions — Map observed symptoms to the right checks and next steps.
Save problematic sessions as test cases — Add validated fixes to the regression suite so the same issue cannot recur.
Optimized configurations must be validated through testing, and problematic sessions must be saved as regression test cases to prevent similar issues from recurring in production.
Step 1: Narrow the execution records
In the execution records, filter by execution status, batch number, phone number, call result, primary label, call duration, and time range. Identify representative samples first, then move on to individual call details. Do not adjust scripts based on summary metrics alone.
Step 2: Inspect individual call details
For each representative call you selected in Step 1, review the following areas.
| Area | What to verify |
| Output labels and call summary | Whether the primary intent, structured fields, and call summary match the actual conversation. |
| Call variables | Whether the variable values passed in by the task, such as name, city, and business information, are correct. |
| Call recording | The opening sentence, speech speed, voice tone, pauses, interruptions, and overall listening experience. |
| Call log and node transitions | The customer's exact words, the agent's responses, and the actual path taken through the nodes. |
Step 3: Match symptoms to actions
Use the following table to map an observed symptom to the checks and next step that address it.
| Symptom | Check first | Next step |
| High early hang-up rate | Opening sentence, identity introduction, speech speed, and voice tone. | Shorten the opening and re-test the first sentence. |
| Intent recognition errors | Prompts, node conditions, and label rules. | Mark representative sessions as test cases, update the rules, and run regression. |
| High drop-off at a specific node | Node script, entry conditions, and exit paths. | Make targeted fixes in the canvas or prompts, then run regression. |
| Inaccurate call summary | Label definitions, examples, and generation rules. | Compare call summary tests against real records and add edge case coverage. |
| Increase in failed tasks | Number format, line quality, time slot, and concurrency. | Pause scaling and investigate by failure category. |
Step 4: Save problematic sessions as test cases
In the call details, identify the problematic response or incorrect label.
Click Mark as test case and select Model's next response.
Specify the expected logic, correct and incorrect examples, and how variables and labels should be evaluated.
Add the test case to the regression suite for the current branch, then re-run it after making fixes.