Evaluation criteria

Criteria evaluate specific aspects of a conversation, such as whether the agent confirmed the reason for an inquiry. Define an observable rule and connect it to the relevant agents.

​

Create and attach

1. Open Manage evaluations from the report or organization evaluation settings.
2. Create a criterion with a name, key, and evaluation prompt. Its key cannot be changed after creation.
3. Choose Binary or Percentage and explain the rubric: acceptable evidence, passing behavior, and how to handle insufficient information. Add agent context when needed.
4. Save and select the criterion in the voice agent’s Advanced → Post-call settings. An agent supports up to 15 criteria.
5. Criterion assignments save immediately, without publishing. Test calls that should pass and fail. Connecting criteria requires evaluation_criteria.update and agents.update.

​

Read the report

The report includes a criterion ranking, verdict timeline, and individual results. Open a criterion to filter success, failure, or unknown results; open the associated call to inspect evidence. Criterion pass rate = successes ÷ (successes + failures). Unknowns are excluded from the denominator. Coverage indicates how much material has definitive verdicts. A high pass rate with low coverage therefore does not describe every call. The global metric counts calls whose verdicts were all successful. Percentage criteria also show a distribution and average score. Human ratings and automatic verdicts are separate records. Archive criteria that should no longer be used, retaining the context needed to interpret historical results.

​

Frequently asked questions

​

See also

• Post-call
• Dashboard