A hospital network says: "Use AI to reduce emergency department wait times." Walk me through how you would approach it.
Think first
Do not propose a model. Clarify the goal, find the bottleneck, define a metric, then pick the smallest valuable slice.
Strong answer
Start by restating the goal as a measurable outcome: median time from arrival to first clinician contact, for which hospitals, measured how today. Then map the patient flow (arrival, triage, waiting, assessment, tests, disposition) and ask where time is actually lost; it is often bed availability or test turnaround, not triage. Ask for the data that exists (timestamps in the EHR, staffing rosters, bed management) and who owns it. Separate options by risk: AI that touches clinical decisions (triage acuity) carries regulatory weight and needs clinician sign-off; AI that touches operations (forecasting arrivals to plan staffing, summarising referral letters, chasing pending test results) is lower risk and faster to prove. Propose one operational slice, e.g. a 4-hour-ahead arrival forecast for staffing, with a baseline, a success metric agreed with the operations lead, and a 6-week pilot at one site. Close with what would make you stop: no reliable timestamps, or no one accountable for acting on the forecast.
Weak-answer indicators
- Jumps to "build a triage chatbot"
- Never defines a metric or baseline
- Ignores clinical safety and regulation
- Proposes a whole-hospital rollout as the first step
Likely follow-up questions
- The CMO insists on AI triage first. How do you respond?
- Your forecast is accurate but staffing does not change. What now?
- What data quality problems would you expect in ED timestamps?