Voice AI for Healthcare

The agent listens to the patient encounter, either as an ambient listener in the exam room or as a post-encounter dictation interface, and generates a structured SOAP note draft from the conversation audio. Transcription runs on Deepgram Nova-2 Medical with custom vocabulary tuning applied to your specialty's terminology: drug names, procedure codes, anatomy terms, and condition-specific language that general-purpose transcription models consistently mishandle.

The draft note is structured against your EHR's template fields before pre-population. FHIR R4 DocumentReference or the EHR's clinical note write API receives the draft where write-back is supported. A mandatory human-in-the-loop checkpoint requires clinician review and sign-off before any note is finalised, the agent drafts, the clinician approves. CDS Hooks integration can trigger the documentation agent from an EHR context event, so the draft appears in the clinician's queue as the encounter closes.

Note generation scope is defined and reviewed by your clinical team during a two to four week validation phase before the agent goes live. This is not optional, it's the step that catches the gap between what a general LLM produces and what your documentation standards require.

The agent handles inbound patient calls before an appointment or telehealth visit. It collects chief complaint, symptom onset and duration, severity, relevant medical history, current medications, and insurance information in a structured guided conversation. The question sequence branches dynamically on the chief complaint: a patient presenting with chest pain gets a different intake path than one presenting with a skin condition, without requiring either to answer irrelevant questions.

Symptom patterns are screened against urgency criteria defined by your clinical team. Red-flag combinations, chest pain with dyspnea, sudden severe headache, stroke indicators, trigger immediate routing to emergency escalation rather than standard appointment queuing. The triage logic is set by your clinical team before deployment; the agent applies it without deviation.

Collected intake data is structured as FHIR R4 Patient, Condition, Observation, and MedicationStatement resources and submitted to the EHR via the FHIR write API. The provider arrives at the encounter with structured intake already in the record. Every patient response and every routing decision is captured in the audit trail per 45 CFR 164.312 access control requirements.

Prior auth calls are the most time-intensive administrative workflow in most medical groups. The agent handles the end-to-end process: extracting relevant clinical criteria from patient records via FHIR R4 resource types, matching them against payer-specific guidelines, placing or routing the submission, and tracking status on the defined follow-up schedule.

Pending requests approaching payer response deadlines are escalated automatically. Incomplete submissions where the payer requires additional documentation are flagged with the specific gap noted, not a generic denial, so the staff member handling follow-up has the information to act immediately. Denials assessed as viable for appeal are surfaced with the denial code and available clinical documentation attached.

The workflow is implemented using LangGraph for stateful process management. The multi-step prior auth process, extract, match, submit, track, escalate, is modelled as a directed graph with explicit human-in-the-loop checkpoints before any submission action. The audit trail captures every step per 45 CFR 164.312.

The agent handles inbound scheduling requests and outbound appointment reminders across provider calendars. Availability is retrieved in real time from the scheduling system via FHIR R4 Schedule and Slot resources where the EHR exposes them, or via the scheduling system's native API. Appointment type rules, which types can be booked by voice, minimum notice requirements, and provider-specific block scheduling, are configured during setup.

Confirmed bookings write the FHIR R4 Appointment resource to the EHR and send a patient confirmation with details, location or telehealth link, and preparation instructions. Outbound reminder calls run at configured intervals, typically 72 hours and 24 hours before the appointment, with voice-activated confirm, reschedule, or cancel options. Cancellations received outside the no-show window trigger a slot release and an offer to waitlisted patients who match the appointment type and provider criteria.

The agent escalates to front desk staff for requests outside its configured scope: complex multi-provider scheduling, patients with active care plans that require clinical input on scheduling decisions, and requests where the patient describes a symptom suggesting a different appointment type than requested.

Inpatient nurse call systems generate a high volume of non-urgent requests, pain medication timing, meal preferences, room temperature, TV controls, and general questions, that consume nursing staff time disproportionate to their clinical complexity. A voice AI agent handles these requests directly: answering common questions from a structured knowledge base, routing non-urgent service requests to the appropriate floor staff, and reserving nurse escalation for requests that match clinical intervention criteria.

Escalation logic is defined by your nursing leadership before deployment. Requests that contain clinical indicators, reported chest pain, difficulty breathing, fall events, or specific medication questions, route immediately to the nursing station. The agent does not attempt to handle clinical assessment. It handles the volume of routine requests that represent 40 to 60 percent of nurse call interactions in most inpatient units.

Integration connects to your nurse call system and, where supported, to the EHR for patient context retrieval via FHIR. Every interaction is logged with timestamp, request type, routing decision, and response delivered, providing the audit record needed for quality review and compliance documentation.

Accurate medical coding requires mapping clinical documentation to the right ICD-10, CPT, and HCPCS codes, a process that drives billing accuracy and revenue capture. A voice AI agent assists coders by listening to dictated encounter summaries or reviewing documentation already in the EHR and suggesting relevant code candidates with the supporting clinical rationale extracted from the note.

The agent uses BioBERT or a fine-tuned clinical NLP model to extract diagnosis, procedure, and service information from unstructured documentation. Code suggestions are ranked by confidence and presented to the coder with the specific supporting language from the note highlighted. The coder reviews, selects, and submits, the agent surfaces candidates and rationale, the human makes the final coding decision.

FHIR R4 Claim and ExplanationOfBenefit resources are used for integration with the billing system where supported. The audit trail records every suggestion, every human decision, and every code submitted, providing the documentation required for coding quality audits and payer reviews.