AI Patient Intake Assistant

Business Challenge

Meridian Health's front-desk staff across 12 clinics were spending roughly a third of every shift on repetitive intake work — insurance verification, symptom pre-screening, and appointment confirmations — almost all of it collected by phone or paper form and re-typed into their EHR by hand.

Objectives

• Cut manual intake data entry without adding headcount
• Keep every interaction HIPAA-aware, with a clear human escalation path
• Integrate directly with their existing EHR rather than replacing it

Approach

We started with two weeks shadowing front-desk staff across three clinics to map every intake variant — new patients, returning patients, referrals, and same-day bookings each had a slightly different flow. That map became the spec for the conversational assistant.

Deliverables

A web and SMS-based intake assistant, an internal review queue for edge cases, and a lightweight analytics dashboard tracking completion rates and escalation reasons.

Solution

Planning

We scoped a phased rollout: one pilot clinic for four weeks, then a staged rollout to the remaining 11 once the escalation rate stabilized below 8%.

Design

Conversation flows were designed around plain-language prompts rather than form-style fields, tested with actual patients during the pilot for comprehension before wider rollout.

Development

Built on Next.js with a PostgreSQL-backed conversation state machine, Twilio for SMS delivery, and a HIPAA-aware data layer that keeps PHI inside Meridian's existing infrastructure rather than a third-party store.

Testing

Two weeks of shadow-mode testing — the assistant ran in parallel with staff intake, with outputs compared but not yet acted on — caught several ambiguous-symptom cases that needed clearer escalation rules before go-live.

Deployment

Rolled out clinic-by-clinic over six weeks, with a live dashboard so clinic managers could watch completion and escalation rates in real time.

Optimization

Post-launch, we tuned escalation thresholds twice based on staff feedback, cutting false escalations by roughly half without missing any cases that genuinely needed a human.