An AI router can send a patient message to the right queue and still miss the symptom that matters.

Jivesh Sharma, M.D., and Talia, our AI cohost, examine message routing, emergency triage, prior authorization and breast screening. Animated diagrams explain how software allocates clinical attention, when it should escalate, and what to measure beyond accuracy.

Share a question or study at https://www.healthit.com/feedback. Please leave out patient-identifying information. Thank you for watching and listening.

CHAPTERS
00:00 Not urgent is not the same as not important
01:18 Deciding where things go
02:34 The least expensive intelligence that can safely do the job
04:14 Route by consequence
05:02 Triggers, and which way to err
05:46 Layers, not a single filter
06:11 Consumer triage
07:33 Hospital alerts and missed cases
08:11 The label was cost, not need
09:05 Prior authorization and incentives
10:40 Human review
11:28 How much attention, not whether
12:29 What made it work
13:39 Same risk, different design
14:06 Five rules for an AI router
15:13 Pay for accuracy
15:38 Put the most important thing first
16:21 Not urgent should never mean not important

SOURCES AND SCOPE
OPTIC (preprint; agreement with GPT-4 labels, not patient outcomes): https://arxiv.org/abs/2503.05701
Johns Hopkins team overview: https://bids.jhmi.edu/news/triaging-patient-messaging-with-ai/
ChatGPT Health triage test, Nature Medicine (2026): https://www.nature.com/articles/s41591-026-04297-7
Methodological counterpoint (preprint): https://arxiv.org/abs/2603.11413
Earlier Epic sepsis model, retrospective validation (2021): https://jamanetwork.com/journals/jamainternalmedicine/fullarticle/2781307
Cost-proxy bias (2019; simulated change): https://doi.org/10.1126/science.aax2342
Medicare WISeR: https://www.cms.gov/priorities/innovation/innovation-models/wiser
CMS timing and appeal-payment details: https://www.cms.gov/priorities/innovation/files/document/wiser-model-frequently-asked-questions
EFF release and corrections: https://www.eff.org/deeplinks/2026/09/new-records-reveal-problems-medicares-ai-prior-authorization-experiment
CMS records via EFF: https://www.eff.org/document/wiser-foia-response-combined-records
Jessica Fu, The Seattle Times, Sept. 16, 2026 (syndicated; vendor denominators and response): https://www.arcamax.com/business/businessnews/s-4299526
Early Washington and Arizona rates are descriptive, not adjusted comparisons of accuracy or clinical appropriateness. The 83-day delay covers the full request pathway; CMS's 72-hour target applies to direct vendor submissions. EFF is an advocacy organization that challenged the program.
MASAI interim analysis (2023): https://pubmed.ncbi.nlm.nih.gov/37541274/
MASAI interval-cancer analysis (2026): https://pubmed.ncbi.nlm.nih.gov/41620232/
MASAI studied one Swedish workflow against double reading. Non-inferiority does not establish that automation bias cannot cause errors.

DISCLOSURES
The patient example is fictional. Jivesh Sharma's voice is AI-generated with his authorization; Talia is an AI cohost. Jivesh Sharma, M.D., is CEO of Nexgen Precision, a healthcare AI venture studio. No Nexgen Precision product is discussed. Proposed routing rules are editorial recommendations, not a validated clinical protocol.
Research cutoff: September 28, 2026. Sources checked September 30, 2026.