Three AI layers on one chart: the week OpenAI plugged ChatGPT into Epic, OpenEvidence shipped its own models, and $220M went to the layer nobody was fighting over
Three things happened this week in the same place: the chart a clinician opens before a visit. On September 1, OpenAI announced ChatGPT Health can pull read-only data from Epic — an EHR covering more than 325 million patients — with UCSF Health the only named pilot so far (TechCrunch, 2026-09-01). On September 3, OpenEvidence — valued at $12B and claiming daily use by more than 40% of US physicians — shipped its own family of clinical models (STAT Health Tech, 2026-09-03). And the biggest cheque of the week, on September 2, went to Scan.com for automating imaging bookings: $220M against a $165M annualised run rate that doubled year over year (Fierce Healthcare, 2026-09-02). Three AI layers land on one workflow, and the hospital paying for all of them buys the same job three times — while the one layer nobody is fighting over, and the one reliably collecting cash, sits outside the chart entirely.
01 — Top Stories
OpenAI wires ChatGPT Health into Epic charts: read-only, no write-back, UCSF first
On September 1 OpenAI announced an Epic connector for ChatGPT Health, letting clinicians import visit notes, lab results, medications and specialist documentation to summarise records, surface changes in history and prepare for appointments. The integration is read-only; the AI writes nothing back to the chart. Epic covers more than 325 million patients, and ChatGPT Health — opened to all US consumers only last month — already fields roughly 300 million health-related queries a week (TechCrunch, 2026-09-01). UCSF Health is the only named clinical pilot (healthsystemCIO, 2026-09-01).
This is the one item this week that sends a hospital CIO straight to a spreadsheet. The same job — pre-visit chart review — is now sold by three vendors at once: Epic's native AI, the ambient documentation vendor deployed two years ago, and now ChatGPT. Three layers on one workflow means duplicate licensing, separate validation processes, and clinicians splitting by preference (healthsystemCIO). The governance question is not whether the model is good; it is whether this connection counts as a sanctioned partnership or an ordinary third-party interface — the two clear entirely different approval gates.
Every safety figure OpenAI cites is self-reported: 99.1% of responses rated safe across 4,363 physician ratings on 27 clinical use cases, with data-connector accuracy of 93.2%–98.6%, none independently audited (healthsystemCIO). Meanwhile OpenAI still faces lawsuits alleging its consumer product gave harmful medical advice, including dosing errors (TechCrunch). Whether the enterprise path requires a BAA, and how it is walled off from the consumer product, is a question each system must ask for itself.
OpenEvidence ships its own clinical model family: from interface-on-someone-else's-model to its own
STAT Health Tech reported on September 3 that OpenEvidence has released a new family of clinical AI models. The company closed a $250M Series D in January 2026 at a doubled $12B valuation (CNBC, 2026-01-21; Fierce Healthcare), and claims daily use by more than 40% of US physicians across 10,000-plus hospitals, with annualised revenue estimated at $300M as of July 2026 (Sacra).
The timing is not a coincidence. OpenEvidence's moat has always been content licensing — the NEJM, JAMA, Wiley, Cochrane, ACC, ADA and NCCN bundle (Sacra) — not the model. With OpenAI walking into the Epic chart carrying its own model, half of the "interface plus licensed corpus" bundle is suddenly exposed. Training your own is how you take the last outsourced piece back in-house, and how you take gross margin back off someone else's API bill.
The STAT piece sits behind STAT+; this report is written from the publicly visible headline and deck only, and does not have model names, sizes, benchmarks or pricing. OpenEvidence's past capability claims — the perfect USMLE score, for instance — are company-published (OpenEvidence announcement), and no established mapping runs from exam scores to clinical outcomes.
Scan.com raises $220M: 85% of imaging scans are still booked by fax or phone
On September 2, medical imaging network Scan.com announced $220M: $90M of Series C equity plus $130M in debt. Noteus Partners led the equity, with Aviva, Concord Health Partners, YZR Capital and Oxford Capital participating; VerisFi Capital and Atempo Growth provided the debt. Founded in the UK in 2017 and in the US since 2023, the company has served over 900,000 patients globally, with revenue doubling year over year to a $165M annualised run rate (Fierce Healthcare, 2026-09-02).
This is the most honest item of the week. Scan.com's AI does not read images, make diagnoses or write notes — it matches referrals against network availability and pricing, because 85% of scans are still booked by fax or phone. Co-founder Charlie Bullock puts the position plainly: "Labs got that decades ago with Quest Diagnostics and Labcorp. Imaging never did, and that is what we have built." While everyone else races to be the physician's second opinion, someone went and became the switchboard — a business with a clear unit of billing, a clear payer, and revenue that doubled in a year.
The $165M is a company-stated annualised run rate, not audited annual revenue, and $130M of the $220M is debt rather than equity. The report does not disclose a post-money valuation for the round.
Cureety acquires Reimagine Care: AI triage meets virtual oncology care across 250 cancer centres
Announced September 2, terms undisclosed. Cureety builds therapy-specific symptom monitoring and triage; Reimagine Care runs a 24/7 virtual care centre with an AI assistant called Remi that resolves roughly 95% of patient needs virtually. Combined, they cover more than 250 cancer centres in five countries and over 100,000 patients. Recent implementation data showed 97% of patient interactions resolved without provider escalation and only 2.4% referred to the emergency department (Fierce Healthcare, 2026-09-02).
Cureety COO Misha Kaur's line is worth copying down: "The test is not whether AI can have an impressive conversation… it is whether the right human enters the conversation at the right moment." It is the week's crispest product definition — in oncology the value of AI is not generation but routing. It also explains why these companies buy rather than build: you can write a triage algorithm yourself, but a 24/7 oncology-trained nursing bench is faster to acquire.
The 95%, 97% and 2.4% are all company-stated implementation figures, with no peer review or third-party audit; "resolved without escalation" is a vendor-defined measure, and no false-negative rate is disclosed — in oncology that is the number that actually matters. With terms undisclosed, the valuation cannot be assessed.
1,357 AI devices cleared; three were tested on whether patients live
A systematic analysis published in PLOS Digital Health on August 19 found that of 1,357 AI/ML-enabled devices cleared by the FDA through December 5, 2025, only 34 (2.5%) had a registered prospective clinical trial, 12 (0.9%) posted results on ClinicalTrials.gov, and three (0.2%) evaluated patient-centred outcomes such as mortality or readmission. Of the trials that exist, nearly 74% enrolled fewer than 500 participants and 68% were US-only, with pregnant women, non-English speakers and paediatric patients systematically excluded (PLOS Digital Health, 2026-08-19).
Set this beside the four items above and the week takes shape: distribution is being solved (OpenAI into Epic), the model problem is being solved (OpenEvidence training its own), the money problem is being solved (Scan.com's $165M run rate) — and the proof problem is not being solved at all. The authors are blunt: readiness should no longer be defined by FDA clearance alone, and they propose a three-phase framework of retrospective validation, prospective safety studies and multi-centre outcome trials.
First Fierce AI Innovation Awards names 18 winners — more drug-side than bedside
Announced September 2, one winner in each of 18 categories: Recursion (drug discovery), Xaira Therapeutics (preclinical), Unlearn (trial design), Inato (trial operations), Innovaccer (data interoperability), Verantos (real-world evidence), Insight Health AI (healthcare delivery), Regard (point-of-care engagement), Implicity (digital health and diagnostics), Komodo Health (commercial strategy), Codoxo (regulatory and compliance), Vitea (data governance) and others (GlobeNewswire, 2026-09-02).
The list is itself an industry map: nine of the 18 categories sit on the pharma value chain (discovery, preclinical, trials, medical affairs, marketing, supply chain), and only three — Regard, Insight Health AI, Implicity — are anywhere near a bedside. Not one of the year's loudest names — OpenEvidence, Abridge, Ambience — appears. The awards were judged on "measurable, real-world impact", which is precisely what clinician-facing AI still lacks.
The awards are judged by the organiser from submitted entries, so absence signals a company that did not enter, not one that is behind. The list reflects the submission pool, not market share.
Nourish puts a genAI assistant in the patient app: doctor visits up 2–3x, dietitian visits flat
On September 2 nutrition care platform Nourish moved its 24/7 generative AI assistant from beta to wide rollout, covering meal planning, lab insights, appointment prep, prescriptions and insurance support. Company-stated figures: 50% of active patients use it daily; users saw 15% greater weight loss at 30 days and 30% at 60 days; weight tracking up 20% and daily meal logging up 15%; lab scheduling doubled; and about 50% of dietitian sessions use the AI "Assist" feature (Fierce Healthcare, 2026-09-02).
The most interesting number is the one that did not move. Doctor visits rose 2–3x and lab scheduling doubled, while dietitian visit frequency stayed flat. That suggests the assistant is not substituting for the human dietitian but pushing patients toward more expensive parts of the system. Whether that is savings or new spend depends on whether those extra visits and labs were indicated in the first place — a question company-reported data cannot answer.
02 — Product Analysis
ChatGPT Health（Epic connector）
A general model reaching down into the chart · OpenAI (US)
Function and position. Read-only import of Epic visit notes, labs, medications and specialist documentation for pre-visit synthesis and change detection; in some deployments it is embedded directly in the chart workflow so the clinician never leaves the patient record. The buyer is the health system, but the product's real weight sits in habits that already exist — ChatGPT Health opened to US consumers only last month and already handles roughly 300 million health queries a week (TechCrunch).
- Strength : distribution costs almost nothing. Epic covers more than 325 million patients, and clinicians already use ChatGPT — no market to educate, just a wire to connect (TechCrunch, 2026-09-01).
- Strength : read-only is a smart self-limit. Writing nothing back to the chart draws the liability boundary outside the documented record and drops the compliance review a full notch (healthsystemCIO).
- Concern : every safety figure is self-evaluated with no health-system validation — 99.1% safe across 4,363 physician ratings on 27 use cases, connector accuracy 93.2%–98.6%, all vendor-supplied (healthsystemCIO). The company also carries live lawsuits over harm from consumer-side medical advice (TechCrunch).
- Concern : the customer list has one name on it. UCSF Health is the only disclosed pilot — and in healthcare the gap between "available" and "one site is trying it" runs about two years (healthsystemCIO, 2026-09-01).
OpenEvidence
A licensed-corpus answer layer monetised by advertising · OpenEvidence (US)
Function and position. Free to verified US physicians, monetised through pharma and device advertising plus enterprise subscriptions — CPMs of $70 to $1,000-plus against social media's $5–15, working out to roughly $124 average revenue per user (Sacra). That model let it bypass healthcare's ~18-month sales cycle, and explains revenue going from $7.9M in 2024 to an estimated $300M by July 2026. Its own clinical model family shipped September 3 (STAT).
- Strength : the content licences are genuinely hard to copy. The NEJM, JAMA, Wiley, Cochrane, ACC, ADA and NCCN bundle is not simply purchasable — publishers are far readier to license a small company that will not become their competitor than a general-purpose model vendor (Sacra).
- Strength : it is already inside the chart. Epic deployments at Sutter Health (Feb 2026), Mount Sinai (March) and Cedars-Sinai (May) mean it walked the path OpenAI took in September six months ago (Sacra).
- Concern : the tension between ad revenue and clinical neutrality has no third-party referee. When pharma pays $70–$1,000 CPMs for a physician's attention at the moment of decision, no public mechanism audits where the answer ends and the advertisement begins (Sacra).
- Concern : a $12B valuation against roughly $300M annualised revenue is about a 40x multiple, at the moment half the moat — the model — is being commoditised by a general-purpose vendor (CNBC). The new family's size, benchmarks and pricing are not public, so this report cannot judge whether it actually closes that gap.
03 — Companies & Competition
| Company | Recent state & numbers | Position & moat |
|---|---|---|
| OpenAI General model moving into the clinic |
Epic read-only connector went live Sept 1, across an EHR covering 325M-plus patients; ChatGPT Health handles ~300M health queries a week; UCSF Health is the sole named pilot (TechCrunch, 2026-09-01). | The moat is installed habit plus raw model capability; the thin spots are zero clinical publications, zero third-party audit, negligible customer scale, and an unsettled compliance status. |
| OpenEvidence The physician answer layer |
$12B valuation on roughly $700M raised (CNBC, 2026-01-21); daily use by 40%-plus of US physicians across 10,000-plus hospitals, 65k new clinician sign-ups a month, ~$300M annualised revenue as of July 2026 (Sacra); own model family shipped Sept 3 (STAT). | The moat is the journal licence bundle plus habit built on free access; the thin spots are an unaudited conflict of interest in ad monetisation and a model layer a general vendor just caught up on. |
| Epic The chart itself |
Its EHR covers 325M-plus patients while it also runs native AI summarisation and opens the chart to third parties — making it, this week, simultaneously OpenAI's channel and its competitor (healthsystemCIO, 2026-09-01). | The moat is being where the data lives — everyone who wants the patient goes through it; the thin spot is that if its native AI lags, it gets demoted to plumbing. |
| Abridge Ambient documentation |
$5.3B valuation on ~$830M raised, including a $316M Series E extension in April 2026; deployed at 250-plus health systems including Kaiser Permanente, Mayo Clinic, Duke Health and Johns Hopkins (ValueAdd VC compilation, 2026). | The moat is contract inertia and depth of revenue-cycle integration; the thin spot is that pre-visit chart review is exactly the overlap OpenAI and Epic both aimed at this week. |
| Scan.com Imaging booking infrastructure |
Raised $220M on Sept 2 ($90M equity plus $130M debt), led by Noteus Partners; $165M annualised revenue, doubled year over year; over 900,000 patients served globally (Fierce Healthcare, 2026-09-02). | The moat is network effects and supply-side coverage, not a model; the thin spot is that this is channel aggregation, and a large payer building its own matching engine routes straight around it. |
| Hippocratic AI Voice agent fleet |
Launched Agentic Orchestrators on Aug 13, described as a 700B-parameter primary model with 30-plus supervising models, validated by 7,700 US-licensed clinicians across 775,000 calls at 99.89% correct advice and zero severe harm; 250M cumulative clinical interactions on $444M raised (PR Newswire, 2026-08-13). | The moat is voice labour substitution at scale priced against payer outcomes; the thin spot is that every safety figure is company-reported, with neither customers nor pricing disclosed. |
| Cureety ＋ Reimagine Care Oncology triage |
Acquisition closed Sept 2, terms undisclosed; combined footprint of 250-plus cancer centres in five countries and 100,000-plus patients, with implementation data showing 97% of interactions resolved without escalation and 2.4% referred to the ED (Fierce Healthcare, 2026-09-02). | The moat is therapy-specific clinical logic plus a 24/7 oncology bench; the thin spots are small scale, undisclosed terms, and no published false-negative rate — the safety number that matters most. |
Today's competitive structure is an inverted funnel. The layer closest to the clinician's cursor — answers and summaries inside the chart — holds three or more contenders with the highest valuations, the wildest revenue multiples and the thinnest evidence. The layers furthest from it — booking, oncology triage, outbound voice — have the fewest competitors, the clearest units of billing and the most real cash. That PLOS 0.2% sits squarely at the crowded end: what is being fought over is attention, not outcomes.
04 — Taiwan Angle
(1) The interoperability timetable decides whether Taiwanese hospitals hit the same three-layers-on-one-workflow problem. Health Minister Stone Chung-liang's "333 policy" centres on unifying data structures across medical information systems, with a "FHIR Box" format standard; the plan is nationwide medical-centre record interoperability by end-2026, extending to regional and district hospitals plus a public system for clinics and health stations across 2027–2028 (UDN; iThome). This week's US argument only exists because Epic is already one enormous pool; Taiwan is still connecting the pools, which is a rare ordering advantage — governance rules can be written before the layers stack, rather than retrofitted the way US CIOs are doing now.
(2) The order of magnitude means Taiwan cannot afford to buy the same job three times. The 2025–2029 Deepening Health Taiwan Plan allocates NT$48.9 billion (about US$1.5 billion) over five years across four pillars including precision and telemedicine (GeneOnline; Executive Yuan). For scale: that is the national five-year figure, while Scan.com took $220M in a single round and OpenEvidence has raised roughly $700M. Taiwan cannot win a valuation race at the answer layer — and does not need to. Duplicate licensing is the waste Taiwanese hospitals can least afford, and nailing down interoperability standards first is the cheapest way to avoid it.
(3) The draft AI-in-medicine rules should write the evidence bar into the text. The health ministry has signalled draft regulations for AI in medicine (CNA, 2026-08-06) alongside three national smart-healthcare centres (MOHW Smart Healthcare Center). The PLOS paper hands over a ready-made reference: 1,357 cleared, three measured against patient outcomes. Rules that merely copy "cleared means deployable" inherit that evidence gap intact. Demanding multi-centre trials with patient outcomes as the endpoint is slower, but it is the one differentiation a small market can actually build — the completeness of Taiwan's single-payer claims data is already a scarce international asset.
05 — Further Reading
-
Regulatory clearance without clinical validation of AI/ML-enabled medical devices — PLOS Digital Health (2026-08-19)
The only primary research with a methods section this week. Read the three-phase evidence framework in full — it is the passage most worth copying into a procurement contract.
-
ChatGPT's Epic Integration Stacks a Third AI Layer on the Chart, Leaving CIOs to Sort the Overlap — healthsystemCIO (2026-09-01)
A rare piece written from the buyer's chair. Its clearest contribution is the governance fork — sanctioned partnership versus ordinary third-party interface — which matters more than any product write-up.
-
OpenEvidence revenue, valuation & funding — Sacra
The most complete public teardown of OpenEvidence's economics, including CPM ranges and an ARPU estimate. If you want to argue about the $12B, start with these numbers.
-
Hospitals Are All In on AI, but Testing and Oversight Haven't Caught Up — MedCity News (2026-08)
The complement to the PLOS paper: one covers the clearance-side gap, this one the hospital-side gap. Read both and this week's three product stories self-adjust downward.
-
Do Ambient Scribe Startups Have a Future Now That Epic Launched Its Own Tool? — MedCity News (2026-02)
A question asked six months ago that OpenAI asked again this week. Read it as a dress rehearsal: the argument structure is identical, only the challenger has more money.
06 — References
- ChatGPT Health adds Epic integration for clinicians to import patient data. TechCrunch, 2026-09-01. techcrunch.com
- ChatGPT's Epic Integration Stacks a Third AI Layer on the Chart, Leaving CIOs to Sort the Overlap. healthsystemCIO, 2026-09-01. healthsystemcio.com
- STAT Health Tech: OpenEvidence launches new family of AI models for clinicians. STAT News, 2026-09-03. statnews.com
- Scan.com raises $220M to build largest medical imaging network in the US. Fierce Healthcare, 2026-09-02. fiercehealthcare.com
- Cureety acquires Reimagine Care for AI-driven precision oncology. Fierce Healthcare, 2026-09-02. fiercehealthcare.com
- Nourish embeds generative AI assistant in patient app for 24/7 support. Fierce Healthcare, 2026-09-02. fiercehealthcare.com
- Forge's Fierce Life Sciences Announces the Winners of the First-Ever Fierce AI Innovation Awards. GlobeNewswire, 2026-09-02. globenewswire.com
- Abulibdeh R, Cajas Ordóñez SA, Celi LA, et al. Regulatory clearance without clinical validation of AI/ML-enabled medical devices. PLOS Digital Health, 2026-08-19. journals.plos.org
- OpenEvidence revenue, valuation & funding. Sacra, 2026. sacra.com
- OpenEvidence, the 'ChatGPT for doctors,' doubles valuation to $12 billion. CNBC, 2026-01-21. cnbc.com
- OpenEvidence clinches $250M Series D as AI platform sees explosive growth with doctors. Fierce Healthcare, 2026. fiercehealthcare.com
- OpenEvidence Creates the First AI in History to Score a Perfect 100% on the USMLE. OpenEvidence announcement. openevidence.com
- Abridge Valuation 2026: $5.3B, $100M+ ARR, and How It Beat Nuance and Ambience. ValueAdd VC, 2026. valueaddvc.com
- Hippocratic AI Announces Next Generation of Healthcare AI: Orchestrators Focused on Outcomes, Not Tasks. PR Newswire, 2026-08-13. prnewswire.com
- 高醫大論壇揭示AI醫療新局！衛福部推「333政策」. 聯合新聞網, 2026. udn.com
- 行政院BTC預備會議聚焦AI生技和智慧醫療，再提FHIR Box打造全臺病歷互通. iThome, 2026. ithome.com.tw
- 健康臺灣深耕計畫（114–118 年）. 行政院全球資訊網. ey.gov.tw
- 衛福部「114-118 年健康臺灣深耕計畫」489 億元四大主軸. GeneOnline News. geneonline.news
- 厚生會成立智慧醫療委員會 衛福部擬提AI醫療細則草案. 中央社, 2026-08-06. cna.com.tw
- 臺灣智慧醫療三大中心. 衛生福利部. aicenter.mohw.gov.tw
- Hospitals Are All In on AI, but Testing and Oversight Haven't Caught Up. MedCity News, 2026-08. medcitynews.com
- Do Ambient Scribe Startups Have a Future Now That Epic Launched Its Own Tool? MedCity News, 2026-02. medcitynews.com