AI Cuts Patient Intake From 18 Minutes to 2 Minutes

The Situation

the client operates a network of outpatient clinics and had recently raised their Series B on the strength of their patient engagement platform. Their technology was modern — but their intake process was stuck in 2010.

New patients filled out paper forms in the waiting room. Staff manually transcribed the information into three separate systems: the EHR (Epic), a custom intake portal, and a billing platform. Each transcription took an average of 12–18 minutes per patient — and introduced errors at every step.

The Challenge

Patient intake is a deceptively complex problem to automate in healthcare:

• HIPAA compliance meant any solution had to handle PHI with strict access controls, audit logging, and encryption at rest and in transit
• Epic's integration capabilities are intentionally limited — direct database access wasn't possible
• The billing system was a legacy product with no official API
• Clinical staff had low tolerance for systems that weren't reliable — one bad week could kill adoption
• Data accuracy wasn't optional: a wrong medication field or incorrect insurance ID had real downstream consequences

Our Approach

We spent the first week deeply embedded with the client's clinical operations team — shadowing intake coordinators, mapping every data field across all three systems, and understanding where errors most commonly occurred.

The core technical challenge was getting data into Epic reliably. We chose to use Epic's HL7 FHIR R4 API (which the client had access to as a licensed customer) rather than any screen-scraping or brittle workaround. For the legacy billing system, we built a controlled integration using their file-export mechanism combined with a custom parser.

For data extraction, we used GPT-4o with a structured output schema — patients could fill out a digital intake form via SMS link before their appointment, and the AI would parse, validate, and normalize the data before sending it to any downstream system.

The Solution

The final architecture consisted of:

• Pre-appointment SMS intake: Patients receive a Twilio-powered SMS with a secure link 48 hours before their appointment. The mobile-optimized form takes 4–6 minutes to complete.
• AI validation layer: GPT-4o parses the submission, flags inconsistencies (e.g., insurance ID format errors, medication name variations), and normalizes to standard clinical terminology before any data is written
• FHIR integration: Validated data is written to Epic via HL7 FHIR R4, creating or updating the patient resource, coverage records, and appointment context
• Billing sync: Insurance and demographics data flows to the billing system via a secure file handoff, with reconciliation checks on both ends
• Coordinator dashboard: Staff see real-time intake completion status and can review/override any AI-flagged fields before the patient arrives

Every data flow was implemented with end-to-end encryption, PHI masking in logs, and a full audit trail. We worked through the client's internal security review before go-live.

Results

Within two weeks of go-live, 78% of patients were completing pre-appointment intake via SMS — exceeding the client's 60% target. Manual data entry per patient dropped from 12–18 minutes to under 90 seconds (for coordinator review and confirmation). Data accuracy, measured against chart audits, improved from 91% to 99.2%.

The intake coordinators — initially skeptical — became the integration's strongest advocates. "I used to dread Mondays," one coordinator told the CMO. "Now I actually get to talk to patients."