Built for the people most often left out

Do ambient AI scribes work equally well in every language? Not in the largest evaluation to date. Across 54,000 encounters, 33,000 patients, roughly 150 providers and 28 specialties at Cambridge Health Alliance, non-English encounters were 21 to 25 percent less likely to reach the vendor's adequacy threshold of 80 percent effort reduction. The odds ratios were 0.76 for Spanish, 0.80 for Portuguese and 0.75 for Haitian Creole.

Does an interpreter change the result? No significant difference appeared between interpreter-mediated visits and visits with a bilingual provider. That raises a real-time interpretation question these tools may already be answering in practice, and that no one has studied.

What do the failures look like in an FQHC exam room? Macrobid transcribed as Migrabid. A medication name surfacing with no clinical context. Detail in a note that was never discussed in the room. Laterality dropped from an exam. Notes that read tidier than the visit actually was. The benefits are real at the same time: lower cognitive load, more eye contact, more room for health education.

Why is consent different in the safety net? A 2020 study found speech recognition error rates roughly twice as high for Black speakers. In a safety-net visit a patient may disclose immigration status, intimate partner violence, or a mental health crisis while the room is being recorded, so consent has to be meaningful, offered in the patient's own language, and explicit about retention and model training.

How strong is the evidence base? One randomized trial has been published to date. Follow-up across the literature runs 30 to 90 days. There are no patient-outcome trials, and no published studies from federally qualified health centers at all.

Hosted by the Digital Health Equity Collaborative at UCSF. Co-presented with Hannah Galvin, MD, Chief Health Information Officer and Interim Chief of Pediatrics at Cambridge Health Alliance.

The math of visit-centric care is unsustainable: a single primary care physician would need 26.7 hours a day to deliver full guideline care to a 2,500-patient panel. The visit is the beginning, not the boundary.

The evidence is strong across five conditions. Telehealth lowered systolic blood pressure by 4.9 mmHg and HbA1c by 0.42% across 75 trials and more than 100,000 patients. Remote monitoring paired with nurse coaching moved average blood pressure from 152/85 to 132/74 in a year. Home monitoring is no longer an add-on: the 2025 AHA/ACC hypertension guideline and the 2026 ADA standards move home blood pressure and continuous glucose monitoring to the center of care.

What makes programs fail is routing raw device alerts into the physician's inbox. Role clarity, not technology, sustains the program. And the equity test is simple: if scaling widens the gap, it is not scaling, it is sorting.

The clinic is where care is planned. The community is where health is made.

Hosted by California Telehealth Resource Center.