|
3, 2, 1: Health AI Brief
Every Friday
April 3, 2026
|
|
|
AI is reshaping healthcare fast. Below are 3 key AI developments, 2 studies, and 1 takeaway for this week to help you better lead with AI. Target read time: 5 minutes. |
|
Eli Lilly signed a deal with Hong Kong-listed Insilico Medicine worth $115 million upfront and up to $2.75 billion in development, regulatory, and commercial milestones. Under the agreement, Lilly receives an exclusive license to develop, manufacture, and commercialize preclinical oral drug candidates identified through Insilico's Pharma.ai generative platform. So what?
$2.75 billion for preclinical compounds that an AI platform identified. Not AI tools to help scientists work faster — AI output that Lilly is licensing directly. That's a different category of bet. The March 31 deadline under CMS-0057-F arrived, and payers published their first-ever prior authorization metrics. KFF's initial analysis of UnitedHealthcare's data shows approximately 80% approval for marketplace plans, 92% for Medicaid/CHIP, and 95% for Medicare Advantage. But the reporting lacks service-level breakdowns, denial reasons, and prescription drug authorization metrics. KFF concluded the aggregated data "offers little insight into what gets approved or denied." So what?
Approval rates look high in aggregate. But without service-level detail or denial reasons, there's no way to tell whether that 80% reflects appropriate gatekeeping or unnecessary burden. Transparency (without granularity, benchmarking, and supply-side obfuscations) is still a good start. ElevenLabs, valued at $11 billion as of February 2026, is integrating with Epic, Cerner, and Athenahealth for HIPAA-compliant patient triage, appointment scheduling, and prescription management. Potential applications include multilingual discharge videos and voice cloning for clinician follow-up calls. HLTH's analysis flagged the trust risk: if a patient can't tell whether they're hearing their actual clinician or an AI reproduction, that's a trust problem — not a feature. So what?
Voice is the next modality entering the clinical workflow, after text and ambient listening. The operational upside is real. But the evolution of cloned clinician voices will be worth watching, especially in cases where patients can't distinguish real vs. fake. |
|
A multi-site observational study published in JAMA tracked 1,800+ clinicians using AI scribes across 5 academic medical centers, with 6,770 controls. Led by Dr. Lisa Rotenstein (UCSF) and Dr. Rebecca Mishuris (Mass General Brigham), the study found AI scribes saved 16 minutes of documentation time and 13 minutes of EHR time per 8-hour shift, with approximately 0.5 additional patient visits per week. But clinicians who used the tool in more than 50% of encounters saw notably larger gains — 21 fewer minutes in total EHR time and 27 fewer minutes on documentation, roughly 1.5-2x the average benefit. Only about one-third adopted at that intensity. After-hours "pajama time" did not improve. Why it matters
16 minutes a day is real but modest. The more interesting finding: power users saw roughly twice the benefit, and two-thirds of clinicians never reached that threshold. The tool works, but the adoption isn't fully there, yet. Notably absent from the conversation, though: patients. Whether a patient is comfortable with an AI listening to their encounter rarely gets measured, yet I would argue their agency matters a lot here, too. A systematic review and meta-analysis of 50 studies across 17 medical specialties — led by Waldock et al. at Imperial College London — found a marked gap between technical validation and real-world implementation of AI clinical decision support. 76% of studies used retrospective datasets; only 24% involved prospective deployment. 64% reported exclusively on technical metrics (sensitivity, specificity, accuracy) without documenting workflow integration, clinician adoption, or patient outcomes. The tools showed moderate overall accuracy (0.765) with substantial variability across specialties. Why it matters
The technical results themselves were mixed: specificity of 0.82 but sensitivity of only 0.66, meaning these systems miss roughly 1 in 3 positive cases. And most of the evidence base comes from retrospective studies reporting technical accuracy — not from prospective deployments measuring whether clinicians actually use the tools or patients do better. |
|
The Data Is Out. Now What?
This was supposed to be a transparency milestone. On March 31, CMS-0057-F forced payers to publicly report prior authorization metrics for the first time — volumes, approval rates, turnaround times. The data is now live. KFF's initial review: it "offers little insight into what gets approved or denied." Approval rates look high, but there's no service-level detail, no denial reasons, no Rx data. The numbers exist. What we do with them determines their impact. We've seen this before. When CMS first released health plan "report cards," the expectation was that transparency would drive quality improvement. My own research found the opposite — health plans responded by increasing advertising to compensate for poor scores, not by improving care. Publishing quality data, on its own, didn't change behavior. Prior auth transparency doesn't have to follow the same path. But avoiding it requires changes on both sides — demand and supply — not just more disclosure. Takeaway
For prior auth transparency to matter, patients and providers need to use it to make different choices — in choosing plans, negotiating contracts, challenging denials — and plans need to feel the pressure, competitively or regulatorily, to respond. Without both sides moving, the data just sits there. More data doesn't automatically mean better outcomes. |
|
|
Know someone who'd find this useful? Share |
The Data Is Out. Now What!?
3 signals, 2 studies, 1 insight for health leaders
Apr 4, 2026
•
1 min read