Healthcare’s patient communication problem is structural. A mid-size health system with 400,000 annual patient interactions cannot staff its way to coverage — the math doesn’t work, and the nursing shortage makes it worse. The volume of transactional contact — appointment management, benefits questions, refill requests, lab result interpretation, post-visit follow-up — crowds out the time clinical staff need for actual patient care.
Chatbots and virtual assistants address the transactional layer. The distinction that matters in healthcare is the difference between a bot handling structured, deterministic workflows and a bot generating free-form clinical guidance. The first category is where the immediate value and lowest risk live. Scheduling a cardiology follow-up, confirming insurance eligibility before a procedure, walking a patient through a pre-op prep checklist — these are bounded, auditable interactions with clear success criteria.
The architecture I start with in healthcare environments typically involves three layers. The first is identity and authentication — verifying the patient before any PHI is surfaced, usually via date of birth plus MRN or insurance ID, sometimes with integration into the health system’s existing identity provider. The second is integration — connecting to the EHR through FHIR APIs or HL7 interfaces, with an interface engine handling the translation layer. The third is the conversational logic itself, which in regulated contexts should lean on structured dialog flows with LLM capability reserved for natural language understanding, not generation of clinical content.
The obstacles I see most often are not technical. They are governance gaps — no one has defined what the bot is allowed to say, what it must escalate, and who owns the clinical review of conversation templates. Those decisions need clinical informatics and legal involved before a single line of code is written. The second common obstacle is EHR API access: health systems frequently have internal governance processes that add months to API credentialing. Starting that conversation early, and understanding what sandbox environments the EHR vendor provides, is what separates projects that launch from projects that stall.
