Scenario-based governance case

LLM Triage Risk and Escalation in Pre-consultation Workflow

Post-market decision state: Conditional approval

Problem

A clinic uses an LLM triage chatbot to collect symptoms and history before consultations, classifying patients as low-risk, routine, or urgent review. The LLM may underestimate critical symptoms, miss red flags, or use overly reassuring language that delays care-seeking. The governance focus is not whether the chatbot answers well, but under what conditions it must stop answering and escalate.

Key risks

Governance mechanism

  1. The LLM serves only as a pre-consultation information collection and risk flagging tool; it provides no diagnosis or medical advice.
  2. Red-flag escalation rules (chest pain, breathing difficulty, stroke signs, severe allergy, pregnancy emergencies, suicidal ideation, high fever in children): once triggered, normal conversation stops, urgent advice is displayed, and clinical staff are notified.
  3. All triage outputs carry uncertainty levels and display “AI-generated triage summary, clinician review required.”
  4. Monthly audits of false negatives, complaints, override rates, red-flag trigger accuracy, and escalation response time.

RACI

Responsible: clinical safety lead (red flags and escalation protocol); IT and vendor (triggers, logging, alert routing). Accountable: clinical governance committee. Consulted: frontline clinicians, nurses, legal and compliance, data protection officer, patient representatives, indemnity adviser. Informed: patients, clinic management, clinical teams.

Audit evidence

Red-flag rule list, clinical validation record, triage logs, patient input transcripts, AI risk categories, escalation timestamps, staff acknowledgement records, final clinician assessments, override logs, monthly false negative review, incident reports.

Lesson

In clinical settings, the most dangerous output is not “I don’t know.” It is a critical condition underestimated in fluent language. The core capability of a safe triage system is knowing when to stop answering and hand over to a human.