Spoken confirmations: repeat the risky field, not the whole form

Last updated: 7 Oct 202611 min read

tutorial

A spoken confirmation should focus on fields whose misrecognition changes the requested result. Provide the model with the extracted value, a reason it may be uncertain, and a narrow question that lets the caller correct it. Reading back every known field wastes time and can disclose private details aloud in a shared space. For a book code, say the relevant digits and title only after the account check; ask the caller to confirm when two catalog records are plausible. After a correction, update the structured field and repeat just that field. The backend remains responsible for validating the identifier and eligibility. A polite 'yes' from a nearby speaker is not enough if speaker ownership is uncertain.

Operational case

The caller says 'BK four two one' over a blender. Recognition candidates are BK-421 and BK-427. The agent has not yet opened a private account record, so it asks, 'Was that BK-421 or BK-427?' The patron replies, '421, the atlas.' Once authenticated, the catalog lookup finds 'Harbor Atlas' for BK-421. The agent confirms, 'Harbor Atlas, code BK-421—renew that item?' The patron says yes. It then checks the renewal rule. If the catalog instead maps BK-421 to another title, it stops rather than silently converting the request to the nearest match.

Candidates: BK-421 / BK-427; noisy final digit.
Repair question: 'Was that BK-421 or BK-427?'
Checked title after account lookup: Harbor Atlas.
Effect gate: confirmed BK-421 + eligible renewal + authenticated patron.

Performance and operating cost

If C recognition candidates require a catalog lookup, a naive flow may issue O(C) reads; a focused clarification can reduce reads when identifiers are distinct. Each extra confirmation turn adds speech latency, so reserve it for action-changing uncertainty and privacy-sensitive disclosure. Track wrong-entity actions and unnecessary confirmations separately. Use typed record IDs in the application and speak a human-checkable title only after access control. If the caller cannot confirm, offer a non-effectful way to continue rather than making a probabilistic renewal.

Common Mistakes

Do not use a low-confidence digit to select a record for an effectful call.
Do not recite private account details before access is checked.
Do not interpret a nearby voice as the account holder's confirmation.

Connected lessons

prompt engineering

voice agents