Verify and monitor production
A successful draft test proves only the draft test path. Every publication needs a new production-path smoke test.
Voice-agent smoke test
- Confirm the intended version is marked Live.
- Confirm the inbound phone number is assigned to the correct agent.
- Confirm the agent is Active.
- Call the assigned number from an external phone.
- Test the greeting, one core task, one high-risk action, and the safe ending or transfer.
- End the call and open its exact record in Call Logs.
Verify:
- routing and answer delay;
- recognition and audio in both directions;
- prompt and knowledge behavior;
- function input, result, and external side effect;
- transfer or hangup behavior;
- transcript and duration;
- outcome, summary, and contact enrichment when processing completes; and
- credits charged against the published rate.
Version pinning
A live voice call loads the published configuration when the call starts. Publishing while it is in progress does not rewrite that call. Start a new call to verify the new version.
Website Chat is Beta. Its public sessions similarly pin the published version selected at session creation; later turns in an existing session do not switch versions after a publication.
Monitor after release
Compare a defined post-release window with the baseline:
- connection and failure rate;
- task completion and classified outcomes;
- transfer frequency and destination success;
- unknown-answer and escalation patterns;
- function errors and duplicate actions;
- call duration and latency;
- transcript recognition errors; and
- credits per call and total usage.
Post-call analysis and actions are asynchronous. Allow processing time before treating an immediately missing summary or contact update as a final failure.
Decide whether to continue or roll back
Continue when critical scenarios pass and metrics remain within the release threshold. Pause delivery or restore a known version when the release causes unsafe actions, material misinformation, broken routing, uncontrolled cost, or a significant outcome regression.
Record the decision, affected version, evidence, and follow-up work.