Reducing Patient Intake Time by 68%

Every new patient required staff to transcribe referral documents by hand, verify insurance by phone, and search the EMR for an existing record. The end-to-end process averaged 45 minutes per patient and errors were common: duplicate records were created about 12% of the time, which sits inside the published industry band (a 10% duplicate-record rate is common across healthcare organizations, with some reaching 30%, while AHIMA frames 1% as the well-managed target). Insurance-verification backlogs meant patients sometimes waited days for confirmation. The commercial and clinical stake compounded: misidentification drives roughly a third of denied claims industry-wide, so a 12% duplicate rate is a revenue and a safety problem at the same time.

The network had evaluated three off-the-shelf intake solutions. None could handle the variety of referral-document formats it received from more than 400 referring physicians. The systems assumed a clean, narrow input and the real input was 23 distinct document formats that changed by referrer. A tool that only works on the documents it was demoed with, and quietly mishandles the rest, fails exactly where intake volume is heaviest, and why those evaluations failed was the same in each case: format brittleness, not missing features.

We spent the first three days embedded with intake staff, watching the real workflow and cataloguing the 23 referral-document formats actually in circulation. The system uses Claude Opus 4.8 to parse scanned and digital referral documents, extracting patient demographics, diagnosis codes, referring-physician details, and insurance information. A rules engine handles insurance-verification logic, cross-referencing extracted policy numbers against payer databases in real time. For patient matching, a fuzzy-matching step compares against the existing EMR, and critically it flags uncertain matches for human review rather than auto-merging. The output is a pre-filled intake form that staff verify and submit in about three minutes of hands-on review, the one human step inside an end-to-end intake that now runs about 14 minutes.

Referral documents are parsed by Claude Opus 4.8 with GPT-5.5 as a cross-check on the harder formats, into a structured record. The rules engine validates insurance against payer databases over a FastAPI service backed by PostgreSQL, with documents staged in S3 and the pipeline running on AWS Lambda. The human boundary is the design principle, not a fallback: the fuzzy matcher never auto-merges an uncertain patient match, it routes it to a person, so AI carries the volume and a human carries every judgement call that touches a patient's record. Data handling is built for the regulated setting (PHI scoped, EMR integration, no auto-merge), which is the part that makes the speed defensible rather than reckless.

SECOND-ORDER EFFECTS