Voice Notes to Incident Draft

In practice

First, the feature gate. The AI reads a voice note recorded by the staff member reporting the incident, plus the incident record's definition. It writes one thing: a draft report in the defined fields. It cannot submit, classify for regulators or assign severity. Second, the boundary. Without review, a wrong draft would enter the hospital's safety record, so the draft sits firmly on the suggestion side: the reporter reviews and submits it. Third, the risks. Transcription can mishear drug names and doses, so every number and medication in the draft is highlighted for the reporter to confirm. Voice notes may include patient identifiers, so the work runs in the hospital's governed system, and the draft uses only the fields the definition allows. Fourth, grounding. The prompt uses the incident definition and tells the model to leave a field empty and ask, rather than infer. That rule stops the model inventing details such as a time or a harm level. Fifth, review levels. Severity and reportability are never drafted by AI; they stay with the quality and risk staff, as the case insists. Sixth, monitoring. Reporters' corrections are logged, and the rewrite ratio and the most common corrections are reviewed monthly with the patient safety team.