Controlled customer response

Voice: en-NZ-MollyNeural (natural synthetic voice)

Representative Pūnaha product walkthrough. All data and results shown are fictional.

This representative service workflow begins with a complex customer enquiry about changing an account arrangement. The fictional input contains only the fields permitted by the implementation, and the workflow records which service queue and product area own the request.

A classification stage checks topic, declared sensitivity and urgency. This example is marked sensitive, so the configured policy selects a local model route and prevents the enquiry from being sent to an external model deployment.

Pūnaha uses hybrid retrieval to find approved product guidance, service policy and troubleshooting material. The returned extracts keep their source titles and sections so the service agent can inspect the basis of the draft.

The selected local model prepares a proposed response from the enquiry and retrieved guidance. It follows the requested structure but does not decide whether the response is suitable to send.

A validation stage checks for missing information and conflicting guidance. In this illustrative run, two sources disagree on an effective date, so the workflow marks the conflict and requires human review instead of inventing a resolution.

The service agent reviews the question, sources, conflict and proposed wording. They can amend, approve, reject or escalate. Here, the agent clarifies the date and approves the corrected response.

Only the approved fields can continue through the configured service connection. The model route, retrieved sources, validation state, human amendment and final decision remain in execution history. AI prepared the draft; the accountable person decided what happened next.