Healthcare · Europe · LLM classification · confidence scoring
3,000+ clinical incidents a year, now categorised daily.
Free-text incident reports were reviewed one at a time in each hospital, with no view across the network. Now every report is classified daily, critical events reach clinicians the same day, and buried action plans are lifted out.
Patient found on the floor beside the bed during the night round. Bed rail was down. No visible injury, observed for two hours. Staff to check rails at every round.
- 3,000+/yr
- Incidents classified daily across the network
- 95%+
- Classification accuracy, reasoning shown
- Same day
- Critical events in front of clinicians
A hospital network logged more than 3,000 clinical incidents a year, each written as free text and reviewed by hand, hospital by hospital. Classification varied between sites, severe events could take days to surface, and action plans often sat inside the narrative, never recorded in the formal fields. Nobody could see patterns across the network.
Every new or changed report is now read daily. It is classified, given a severity, and its action plan is extracted, including the ones hidden in the text, with the reasoning shown next to each suggestion. Clinicians review the suggestions side by side with the original record, and low-confidence outputs are flagged for a closer look. Critical events reach clinical review the same day, and patient-safety leaders see patterns across hospitals for the first time.
- 01 · What we foundAssess
Sat with the quality and patient-safety teams. Classification varied by hospital, severity took days to flag, and many action plans existed only in the free text.
- 02 · What we builtBuild
Four coordinated AI agents on top of the existing reporting system: classification, critical-event detection, action-plan extraction, and suggested actions from a menu clinicians approved.
- 03 · Where it standsLive
Every incident across the network reviewed daily, at 95%+ classification accuracy. Critical events surface the same day, and network-wide dashboards are in use.
