A call is a data flow and an action flow
Voice agents process more than a transcript. Audio, caller identifiers, inferred intent, confidence values, summaries, tool results and human notes may each carry a different purpose and retention period. Map these fields to their recipients and declared processing locations. Keep only what the workflow needs.
At first contact, explain that the caller is speaking with AI, the relevant use of their input and how to reach a person. Determine the lawful basis for recording and processing in the specific context. Never shrink the whole analysis to a universal consent checkbox.
Design for mishearing
Test accents, names, dates, medication terms, account numbers, background noise, interruption and language switching in every supported language. Important values deserve confirmation in a form the caller can correct. Confidence thresholds should lead to clarification or escalation — not repeated persuasion.
An unsupported language, an emergency, a distressed caller, a conflicting record or an explicit request for a person each needs a clear route. The handoff should carry enough context to avoid repetition, while excluding unrelated sensitive data. Measure whether the destination can actually respond — not just whether a transfer event was logged.
Lock approval to the exact consequence
A voice agent may gather details and prepare a booking or an external message. In the draft AIA Guard scheme, a person approves the final time, recipient and content before anything is executed. Changes after approval cancel its authority. The same rule covers deletion and clinical case steps.
Test missing approval, refusal, expiry, changed details and reuse of an old approval. The downstream tool must enforce the decision. A sentence in the prompt telling the model to ask permission is not an authorization control.
Leave a record someone can defend
A useful record connects agent version, relevant transcript references, policy decision, tool parameters, human approval, final outcome and any escalation. It should reveal later changes, while limiting who can see sensitive content. Operators need a tested stop and containment path if a connector or customer boundary fails.
Review incidents and representative call samples after release. The draft proposes an evidence review every three months, with immediate reassessment after a serious incident or important change. This rhythm never transfers clinical or financial responsibility to the system.