Epic's AI is now inside the visit — and a third of the conversations its patient agent has happen after hours
August 19, 2026 · 4 items
Epic's AI is now inside the visit — and a third of the conversations its patient agent has happen after hours
Healthcare IT News · Andrea Fox · August 19, 2026Practice operations
Ergo is live. Epic's AI platform — built on Cosmos' 320 million deidentified patient records drawn from 23 billion encounters — now runs Ergo Visit at Ochsner, Central Oregon and Salem Clinic: Art writes the pre-visit summary, then automatically assembles the documentation fields and orders while listening to the conversation. Chart with Art is at 60+ organizations, Art-generated clinician and nurse summaries at 300+ health systems, and 11 systems are live for nursing. More than 1.4 million clinicians are assisted by AI in Epic each month.
The patient agent went live. Sutter was first to go live with Ask Emmie inside MyChart. Trevor Berceau, VP of R&D: more than one third of Emmie conversations happen after hours — "you gave the patient what they needed, when they needed it." Schedule with Emmie arrives in October, the pre-admission testing assistant in November. Emmie also works with Art autonomously to run patient-outreach tasks on the provider's behalf.
Prior auth moved into the workflow. A new API — Coverage Requirements Discovery — lets clinicians check payer requirements without leaving Epic instead of going to an external portal. Ochsner, Froedtert ThedaCare, Denver Health, Summit Health, UnitedHealthcare, Network Health and Aetna have integrated it; 16 payers are testing. Separately (Becker's, Aug 18 — Candidate): Agent Factory lets an organization deploy one of 120 out-of-box AI features or build its own agents without writing code (broad availability 2027), and Health Alerts puts free county- and state-level outbreak detection on EpicResearch.org, public and searchable by state, county or condition.
The pace, and the bill for it. Epic releases went from every 18 months to every 3, with weekly code pushes to its Nebula cloud. CSO Stirling Martin disclosed publicly for the first time that Epic is in Project Glasswing — Anthropic's AI cybersecurity initiative, ~150 organizations — where AI models hunt vulnerabilities human developers can't see. His warning to customers: expect higher-than-normal security fixes, and "turnaround times of 30 days will no longer cut it."
Why it matters for an independent practice: the after-hours number is the one to steal. A third of Emmie's traffic is patients with a question at a time no practice is open — that is precisely the window where an independent regen or ortho practice currently loses the patient to a search engine or a chatbot that has never seen their chart. You don't need Epic to answer at 11pm; you need something that does, wired to real appointment availability. And the prior-auth API is the shape of the next competitive gap: health systems are about to stop losing days to payer portals while independents still burn staff hours on them. For MMR, "we answer after hours" is now a defensible, provable differentiator worth putting on the page.
Two Harvard physicians name the thing your patients are already using: a shadow medical system that borrows medicine's authority and skips its responsibility
STAT News — First Opinion · Arya Rao & Marc Succi · August 19, 2026Practice operations
The scale. More than 40 million Americans ask ChatGPT a health question every day, mostly outside clinic hours — and the models are doing less and less to route them anywhere. A study published this year found the medical disclaimers that used to be standard in chatbot health answers have largely disappeared; today's leading models will ask follow-ups and attempt a diagnosis.
The market that formed around it. Oura sells a 50-biomarker blood panel through Quest for $99. Function Health — valued at $2.5B in November — lets members order 160 lab tests a year, book a full-body MRI, and authorize ChatGPT to read the results. Ro and Hims prescribe off an asynchronous intake. Doctronic, "the world's #1 AI doctor," has run 24 million consultations and now writes AI-generated prescription refills in Utah.
The number that should be in every patient conversation. The authors ran the largest evaluation to date of clinical reasoning across 21 frontier models, published in JAMA Network Open. Given a complete case, the models named the correct diagnosis more than 90% of the time. Given only what a clinician would have at the start of a visit, they failed to produce a comprehensive differential more than 80% of the time — and the authors note that at home, with scattered symptoms and no exam, performance is almost certainly worse than their study shows.
The part that lands on you. "When a physician is wrong, accountability is imperfect but identifiable. When an AI system is wrong, responsibility disperses and, in practice, will often still fall on the clinician." Their objection isn't to the tools — their lab published some of the earliest LLM clinical-decision-support research — it's to "non-inferior to physicians" being treated as the bar for handing over a clinical role.
Why it matters for an independent practice: this is the single most useful patient-education asset produced this month, and it's free. Every regen practice is already competing with a patient who arrived having run their symptoms, their labs and their treatment options past a chatbot at 1am. The 90%-with-the-full-case / 80%-failure-at-triage split is the exact framing that respects the patient's intelligence without pretending the tool is a doctor: the machine is very good once someone has already done the hard part — figuring out which questions to ask. Put that in the intake packet, the pre-consult email and the "what to know before your visit" page. Two Harvard faculty saying it is worth more than a practice saying it alone.
The AI Mode queries Google won't officially give you are leaking into Search Console anyway — and there are now four ways to pull them out, one of them free
The gap. Google shipped Search Generative AI performance reports in Search Console on June 3, 2026 — impressions inside AI Overviews and AI Mode. What it does not give you is the queries, and it's UI-only: the author re-verified on his own property on August 11, 2026 that neither the Search Analytics API nor the BigQuery bulk export exposes the generative-AI data. Officially, there is no query-level AI Mode tracking.
Except the queries show up anyway. Fragments of real AI conversations — follow-ups, replies like "yes go on," whole pasted prompts — appear as ordinary queries in the performance report. Google's John Mueller confirmed the data had always been there. Four extraction methods now exist: Glenn Gabe's full-inventory pull via the API into Excel; Jean-Christophe Chouinard's custom regex, published Aug 14, that you paste straight into the report's Query filter for free; Amin Foroutan's Chrome extension; and the author's Search Console / BigQuery MCP servers.
The free tool. He trained a classifier — FacebookAI's xlm-roberta-base, pre-trained on 100 languages, fine-tuned on labelled Search Console conversation exhaust, validated across just over 120,000 queries — and put it online with no signup and no email. Drop in a CSV, it sorts every query into seven buckets (full conversational, short replies, follow-up pivots, rank-tracker probes, agent-harness prompts, pasted strings, ordinary search) with a confidence score, and exports labelled with no row cap. Handles 100,000 unique queries per run; it's beta and will slow under load.
The honest caveat, stated by the author. Google anonymizes rare queries, conversation strings are rare, and part of the pool never surfaces — treat every output as an undercount. His own BigQuery export showed 57.7% of impressions sat in the anonymized pool over 59 days, measured Aug 11.
Why it matters for an independent practice: every AI-visibility conversation this feed has carried for two months has been measurement about you sold by a vendor. This is the opposite: your own Search Console data, your own device for the deterministic pass, free, today. For a practice this answers the question no keyword tool can — what are people actually typing into an AI when they end up on my site? Run it on a client property, and the conversational-query list becomes the FAQ page, the pre-consult email and the service-page H2s. That's a deliverable MMR can produce this week for the cost of an export.
What is actually verified. Organic clicks down ~42% since Google's AI Overviews scaled — Define Media Group, Search Console data across 64 sites, current as of March 2026. Pew: only 8% of searchers click an organic result when an AI Overview appears, versus 15% when it doesn't. LinkedIn organic reach down 47% YoY since the platform rebuilt ranking around 360Brew, a 150-billion-parameter model that reads expertise and dwell time instead of counting reactions. Pangram Labs scanned 1M+ posts: 41% of LinkedIn long-form content is now fully AI-written — and Pangram calls that a lower bound.
What isn't. The "Business Insider traffic down 85%" figure that circulated everywhere came from a Wall Street Journal chart the Journal corrected on July 22 — the real trailing-12-month decline is 43%, alongside 28% for USA Today and 44% for The Washington Post. The bad chart had already been shared to tens of thousands of views before the correction caught up. Jarboe's point: citing a corrected number without noting the correction "erodes trust in every other number in the piece."
The distinction the whole piece hangs on. Decline means the thing you do has no version worth saving. Structural disruption means the function still matters and the form is changing underneath you. Tony Uphoff's line, which Jarboe endorses: "The machines aren't taking anything from you that you haven't already agreed to rent out."
The three moves. Separate channel metrics from function metrics — build the dashboard around whether AI systems cite you, whether email subscribers convert, whether direct traffic to owned pages grows. Audit what you actually own (list, first-party data, direct relationships) and treat single-platform dependence as a risk to fix this quarter. Stop measuring content by whether it ranks and start measuring whether it earns a citation, a save or a reply.
Why it matters for an independent practice: zero medicine in this one, and it maps onto MMR twice. First, operationally: a practice whose new-patient flow rents its visibility from one platform is one ranking change from a bad quarter — the reviews, the email list, the referral relationships and the site are the parts nobody can reprice. Second, and sharper: this is a discipline lesson for us. We hand doctors numbers about AI search. The 85%→43% correction is a live example of an AI-era statistic going stale inside three weeks while the industry keeps repeating it. Every figure in a client deck should carry its source and its date, or a physician who checks one will stop trusting all of them.
