The landscape · Healthcare
Healthcare: what AI is really doing
Brief me in 2 minutes
AI in Healthcare.
Ambient scribes and comprehensive imaging triage are the two places AI now touches real money and real patients at scale, running in hundreds of health systems on paid contracts. Drug discovery and prior authorization are one step behind, funded heavily and running in production at specific customers but not yet proven at the scale their backers are pricing in. Patient facing chatbots are the wild card: adoption is enormous and mostly unsupervised, and the safety regulator just named that the top hazard in the field for 2026.
The game: we show you 6 things happening. You call each one real now, early or mostly hype. Then the receipts.
Shift 1 of 6
Ambient scribes go from pilot project to clinical infrastructure
Real now
Abridge alone is processing more than 50 million conversations a year across 150-plus health systems on paid contracts, that is deployment, not a demo.
How it works. A microphone in the exam room records the visit, and a model turns the conversation directly into a structured note inside Epic or Oracle Health, with the clinician editing rather than typing from scratch. The vendors are now selling this bundled with billing and revenue cycle features, not just note drafting.
Overstated: Vendors describe this as giving clinicians their time back. The receipts so far are about after hours charting going down, not about visit volume or patient outcomes changing.
Shift 2 of 6
Imaging AI moves from one algorithm per disease to one model reading the whole scan
Real now
This is FDA-cleared and already running in production reads, not a research preprint; Viz.ai's clearances alone cover 1,700-plus hospitals.
How it works. Older radiology AI was a separate cleared product for each finding, one for pulmonary embolism, one for stroke. Aidoc's CARE foundation model reads a single abdominal CT for 14 conditions at once, and Viz.ai has stacked over 50 separate FDA clearances into one coordinated triage workflow across stroke, hemorrhage and aortic disease.
Overstated: The sensitivity and specificity numbers in the clearance are the trained validation cohort's numbers, not a guarantee of the same accuracy on the atypical case in front of a specific radiologist next Tuesday.
Shift 3 of 6
AI-designed drugs reach the Phase III test that actually decides the argument
Early
Money and molecules are real, but as of this writing no AI-discovered drug has an FDA approval, this segment is still waiting for its first proof point.
How it works. Over 173 AI-originated drug programs are now in clinical development, up from roughly two dozen in late 2023. Insilico Medicine's rentosertib, discovered and designed with generative AI, became the first such drug to reach Phase III, in idiopathic pulmonary fibrosis, and Isomorphic Labs raised $2.1 billion to keep its own drug design engine running.
Overstated: Pitches lean on the 80 to 90 percent Phase I success rate. That number is about safety, not efficacy, and the efficacy number, Phase II, is not currently beating the old way of doing things.
Shift 4 of 6
Payer AI starts reading prior authorization requests, with a federal deadline forcing the plumbing to catch up
Early
Live at named customers with real accuracy numbers, but the federal rule that would force this across the whole industry doesn't bite until October 2027.
How it works. AI models from vendors like Anterior and Cohere Health read the clinical note against the coverage policy and auto-clear requests that match, sending only the ambiguous cases to a nurse or physician. A 2026 CMS proposed rule requires payers to run this over standard FHIR APIs rather than fax and portals, with electronic prior authorization for drugs required by October 2027.
Overstated: "AI never denies care" is the standard vendor line, describing where they've chosen to draw the product's scope now, not a technical ceiling on what the same model could be pointed at tomorrow.
Shift 5 of 6
Chatbots become the default first opinion for doctors and patients, largely unsupervised
Real now
Both sides of this are already at massive daily scale, this isn't a projection, it already happened.
How it works. OpenEvidence, a free, ad-supported clinical question answering tool built for licensed physicians, is now used by more than 40% of US doctors. Separately, tens of millions of ordinary patients ask general-purpose chatbots for medical advice with no clinical guardrails, and no license check, at all.
Overstated: OpenEvidence markets itself as built "for doctors," but a free, ad-supported model answering clinical questions is structurally the same product patients use unsupervised, just with a medical license gate on the front door. The gate doesn't remove the hallucination risk underneath.
Shift 6 of 6
FDA starts drafting rules for generative AI devices after clearing over a thousand narrower ones under old rules
Early
The device clearances are real and already happening, but the actual rulebook for generative AI devices is still an open comment period, not a rule.
How it works. The FDA has cleared well over a thousand AI-enabled devices, almost all through the 510(k) pathway built for comparing a new device to an existing one. Generative and agentic AI devices, including foundation models like Aidoc's, don't fit that comparison cleanly, so in August 2026 FDA published a discussion paper asking how to evaluate them, open for comment until October 19, 2026.
Overstated: This is explicitly a discussion paper, not a rule and not binding guidance. Coverage calling it FDA's new "AI regulation" is ahead of what the agency itself has said it is.
The scary part
What happens to jobs in Healthcare
What AI is already taking:
- Writing the visit note from the patient conversation Learning Health Systems (PMC) ↗
- First-pass reading of a CT scan for urgent, incidental findings PR Newswire (Aidoc) ↗
- Reviewing routine prior authorization requests against coverage policy PR Newswire (Anterior) ↗
- Looking up the current evidence for a clinical question at the point of care HLTH ↗
What stays human:
The physical exam, the procedure itself, and the conversation where a clinician tells someone their diagnosis are still entirely human, and no vendor in this briefing is trying to change that. What also stays human, by law rather than by choice, is the signature: courts and regulators are treating the physician's review of AI output as a duty that cannot be handed to the software.
Get better · 15 minutes
Fast, confident auditing of AI output under time pressure, the ability to spot the one wrong dose or missing negative in a note or a scan read before it reaches the patient, rather than rubber-stamping a document that reads as fluently as if you wrote it yourself.
Try this this week:
Pull one AI-drafted note or read from this week and grade it against ECRI's own hazard framing: check every medication dose, contraindication and negative finding against the source encounter rather than skimming for tone, and time how long that actually takes. That number, not the vendor's minutes-saved marketing figure, is your real workload.
Your move
Enterprise ambient scribing is now a three-way war between Abridge, Ambience and Microsoft's Nuance, don't enter it head-on. The open space is narrower: revenue-cycle and appeals automation on the payer side beyond prior auth intake, and agentic patient intake/triage, which is still mostly voice-agent demos, not deployed infrastructure.
Watch retention, not the headline note-drafting demo: Abridge citing over 90% of clinicians who start using the product still using it later is the tell that it works day to day. Red flag in a drug-discovery pitch: any deck that quotes Phase I success rates without also giving you the Phase II number, since Phase II is where AI-discovered molecules are currently performing at ordinary industry rates, not better.
This quarter, pull ten ambient-scribe notes and compare them against your own memory or the recording, specifically hunting for omitted negatives (a symptom the patient denied that never made it into the note), not typos. That is the actual audit gap, and it's a fifteen-minute check per chart, not a research project.
The valuable skill now is fast, confident auditing of AI-generated clinical output, catching a wrong dose or a fabricated contraindication before it reaches a patient, since courts are already treating the physician's review as the point where responsibility sits. First step: read your specialty's entry on FDA's public AI-enabled device list so you know which tools in your own workflow are actually cleared versus just marketed as AI.
What's actually changing
- 1
Ambient scribes go from pilot project to clinical infrastructure
How it works. A microphone in the exam room records the visit, and a model turns the conversation directly into a structured note inside Epic or Oracle Health, with the clinician editing rather than typing from scratch. The vendors are now selling this bundled with billing and revenue cycle features, not just note drafting.
Why it matters. Less after hours charting for clinicians, but health systems are locking into a small number of vendors for a workflow that touches every visit, and the AI's read of the conversation is now the seed for the bill.
What's overstatedVendors describe this as giving clinicians their time back. The receipts so far are about after hours charting going down, not about visit volume or patient outcomes changing.
AbridgeAmbience HealthcareNuance DAX Copilot (Microsoft)Suki
3 receipts
- Abridge's Series E blog post states the company supports over 50 million medical conversations a year across more than 150 health systems.
Abridge is now partnering with more than 150 enterprise health systems across the U.S.
Abridge, 23 June 2025 ↗ - Abridge is now building a foundation model for clinical conversations with NVIDIA and took a strategic investment from Eli Lilly, moving beyond note taking into a broader platform.
NVIDIA announced it's also co-developing with Abridge the first foundation model for clinical conversations
Fortune, 11 June 2026 ↗ - Early multi-site studies of ambient documentation show reduced after-hours charting rather than a headline productivity jump.
Early multicenter evaluations consistently show that ambient AI documentation reduces after‐hours charting, improves perceived documentation‐related well‐being, and may reduce clinician burnout at scale
Learning Health Systems (PMC), 10 Aug 2026 ↗
- Abridge's Series E blog post states the company supports over 50 million medical conversations a year across more than 150 health systems.
- 2
Imaging AI moves from one algorithm per disease to one model reading the whole scan
How it works. Older radiology AI was a separate cleared product for each finding, one for pulmonary embolism, one for stroke. Aidoc's CARE foundation model reads a single abdominal CT for 14 conditions at once, and Viz.ai has stacked over 50 separate FDA clearances into one coordinated triage workflow across stroke, hemorrhage and aortic disease.
Why it matters. Fewer incidental findings get missed and radiologists get one alert instead of fourteen separate tools pinging them, but the diagnostic judgment for many conditions now sits inside one company's model instead of being split across vendors.
What's overstatedThe sensitivity and specificity numbers in the clearance are the trained validation cohort's numbers, not a guarantee of the same accuracy on the atypical case in front of a specific radiologist next Tuesday.
AidocViz.ai
3 receipts
- FDA cleared Aidoc's CARE foundation model to triage 14 conditions from a single abdominal CT scan, its first comprehensive AI triage clearance.
This breakthrough was made possible by CARE™, Aidoc's self-developed AI foundation model.
PR Newswire (Aidoc), 21 Jan 2026 ↗ - FDA cleared the tool to triage 14 critical findings in a single abdominal CT scan.
FDA has cleared a tool that can triage 14 critical findings in a single abdominal CT scan
STAT, 21 Jan 2026 ↗ - Viz.ai received the first and only FDA clearance for automatically quantifying subdural hemorrhage size on CT, adding to its stack of stroke and hemorrhage clearances.
Viz Subdural Plus, the first and only comprehensive solution for quantifying the size of collections like subdural hemorrhages (SDH)
Viz.ai, 12 June 2025 ↗
- FDA cleared Aidoc's CARE foundation model to triage 14 conditions from a single abdominal CT scan, its first comprehensive AI triage clearance.
- 3
AI-designed drugs reach the Phase III test that actually decides the argument
How it works. Over 173 AI-originated drug programs are now in clinical development, up from roughly two dozen in late 2023. Insilico Medicine's rentosertib, discovered and designed with generative AI, became the first such drug to reach Phase III, in idiopathic pulmonary fibrosis, and Isomorphic Labs raised $2.1 billion to keep its own drug design engine running.
Why it matters. Phase I success rates for AI-discovered molecules already beat historic pharma averages, but Phase I only tests safety. Phase II, which tests whether the drug works, is running at roughly the same success rate as ordinary drug discovery, so Phase III is the round that tells everyone if the AI premium is real.
What's overstatedPitches lean on the 80 to 90 percent Phase I success rate. That number is about safety, not efficacy, and the efficacy number, Phase II, is not currently beating the old way of doing things.
Insilico MedicineIsomorphic LabsChai Discovery
3 receipts
- Insilico Medicine dosed the first patient in the Phase III trial of rentosertib, describing it as the first Phase III trial of a generative AI-driven drug.
GENESIS-IPF-3, the World's First Phase III Trial of a Generative AI-Driven Innovative Drug
Insilico Medicine, 10 Sept 2026 ↗ - The number of AI-designed drug programs in clinical development surged past 173 in early 2026, up from roughly two dozen in late 2023.
surged to over 173 in early 2026, up from roughly two dozen in late 2023
IntuitionLabs, 31 July 2026 ↗ - Isomorphic Labs raised $2.1 billion in a Series B round to scale its AI drug design engine and pipeline.
Isomorphic Labs announces it has raised $2.1 Billion in Series B funding.
Isomorphic Labs, 12 May 2026 ↗
- Insilico Medicine dosed the first patient in the Phase III trial of rentosertib, describing it as the first Phase III trial of a generative AI-driven drug.
- 4
Payer AI starts reading prior authorization requests, with a federal deadline forcing the plumbing to catch up
How it works. AI models from vendors like Anterior and Cohere Health read the clinical note against the coverage policy and auto-clear requests that match, sending only the ambiguous cases to a nurse or physician. A 2026 CMS proposed rule requires payers to run this over standard FHIR APIs rather than fax and portals, with electronic prior authorization for drugs required by October 2027.
Why it matters. This is the fix for the paperwork that the AMA says delays care for most physicians, but it also means an algorithm is now the first reader on a decision that can block a patient's treatment.
What's overstated"AI never denies care" is the standard vendor line, describing where they've chosen to draw the product's scope now, not a technical ceiling on what the same model could be pointed at tomorrow.
AnteriorCohere Health
2 receipts
- Anterior closed a $40 million round, bringing total funding to $64 million, reporting 99.24% clinical accuracy independently validated by KLAS Research and a 75% cut in review cycles at one customer.
99.24% clinical accuracy, independently validated by KLAS Research
PR Newswire (Anterior), 12 Feb 2026 ↗ - CMS's proposed rule requires payers to support electronic prior authorization for medical-benefit drugs over FHIR APIs starting October 2027.
CMS now proposes to require impacted payers to incorporate coverage and documentation requirements into those APIs to support the electronic prior authorization of drugs covered under a medical benefit beginning October 1, 2027.
CMS, 10 Apr 2026 ↗
- Anterior closed a $40 million round, bringing total funding to $64 million, reporting 99.24% clinical accuracy independently validated by KLAS Research and a 75% cut in review cycles at one customer.
- 5
Chatbots become the default first opinion for doctors and patients, largely unsupervised
How it works. OpenEvidence, a free, ad-supported clinical question answering tool built for licensed physicians, is now used by more than 40% of US doctors. Separately, tens of millions of ordinary patients ask general-purpose chatbots for medical advice with no clinical guardrails, and no license check, at all.
Why it matters. Adoption on both sides is real and fast, but the safety regulator ECRI just ranked chatbot misuse the top health technology hazard for 2026, and the first lawsuit over a chatbot's medical advice contributing to a patient's injury is now in court.
What's overstatedOpenEvidence markets itself as built "for doctors," but a free, ad-supported model answering clinical questions is structurally the same product patients use unsupervised, just with a medical license gate on the front door. The gate doesn't remove the hallucination risk underneath.
OpenEvidenceHippocratic AI
3 receipts
- OpenEvidence closed a $250 million round at a $12 billion valuation, with more than 40% of US physicians now using it.
the platform is now used by more than 40% of physicians in the U.S.
HLTH, 22 Jan 2026 ↗ - ECRI named misuse of AI chatbots the number one health technology hazard of 2026, citing 40 million people a day turning to ChatGPT for health information.
More than 40 million people daily turn to ChatGPT for health information, according to a recent analysis from OpenAI.
ECRI, 21 Jan 2026 ↗ - A Florida man sued OpenAI alleging ChatGPT told him to stay 'recliner-bound' instead of seeking care for symptoms that preceded a near-fatal pulmonary embolism.
dismissed both sets of symptoms as relatively minor and suggested he stay home and remain "recliner-bound,"
CBS News, 23 July 2026 ↗
- OpenEvidence closed a $250 million round at a $12 billion valuation, with more than 40% of US physicians now using it.
- 6
FDA starts drafting rules for generative AI devices after clearing over a thousand narrower ones under old rules
How it works. The FDA has cleared well over a thousand AI-enabled devices, almost all through the 510(k) pathway built for comparing a new device to an existing one. Generative and agentic AI devices, including foundation models like Aidoc's, don't fit that comparison cleanly, so in August 2026 FDA published a discussion paper asking how to evaluate them, open for comment until October 19, 2026.
Why it matters. Companies are already shipping foundation-model medical devices faster than the category built to evaluate them exists. What FDA decides after the comment period closes will set the evidence bar for the next wave of these products.
What's overstatedThis is explicitly a discussion paper, not a rule and not binding guidance. Coverage calling it FDA's new "AI regulation" is ahead of what the agency itself has said it is.
FDAAidoc
2 receipts
- FDA's discussion paper asks for public feedback on evaluating generative AI-enabled medical devices by October 19, 2026.
DHCoE encourages feedback from device manufacturers, clinicians, researchers, the public, and other interested parties, to be submitted under docket number FDA-2026-N-7874 on Regulations.gov, by October 19, 2026.
FDA, 18 Aug 2026 ↗ - An independent tracker found that by late 2023 roughly 97% of cleared AI/ML devices used the 510(k) pathway, versus 2-3% via De Novo or PMA.
by late 2023 roughly 97% of cleared AI/ML devices used 510(k) (vs. 2-3% via De Novo or PMA)
IntuitionLabs, 26 Aug 2026 ↗
- FDA's discussion paper asks for public feedback on evaluating generative AI-enabled medical devices by October 19, 2026.
Who's doing it
10 companies to know, 8 with a round we can source. Newest first where dated.
Search these companies in the startup database ↗ · Explore all research tools
Insilico MedicineAI drug discovery
Uses generative AI to identify drug targets and design small-molecule drugs, running its own candidates through clinical trials.
Latest · 10 Sept 2026Dosed the first patient in Phase III of rentosertib for idiopathic pulmonary fibrosis, calling it the first Phase III trial of a generative-AI-driven drug. Insilico Medicine ↗
Chai DiscoveryAI molecular design
Builds foundation models trained on biochemical structures that let scientists design antibodies and small molecules computationally, licensed to pharma.
Latest · 20 Aug 2026Announced a collaboration with Bristol Myers Squibb to advance AI-driven antibody discovery, its fourth major pharma partnership of 2026. Chai Discovery ↗
$400M Series C, 14 July 2026 source ↗
Hippocratic AIpatient-facing voice agents
Voice AI agents for non-diagnostic, patient-facing work: post-discharge calls, chronic care check-ins, medication reminders, scheduling.
Latest · 13 Aug 2026Launched 'orchestrators', teams of voice agents run by a supervising model, backed by its Polaris safety system of a 700-billion-parameter model plus 30-plus supervising models. PR Newswire (Hippocratic AI) ↗
$444M cumulative to date, 13 Aug 2026 source ↗
Abridgeambient clinical documentation
Turns recorded doctor-patient conversations into structured clinical notes written directly into the EHR, for health systems.
Latest · 11 June 2026Abridge and NVIDIA are co-developing a foundation model for clinical conversations, alongside a strategic investment from Eli Lilly. Fortune ↗
$300M Series E, 23 June 2025 source ↗
Isomorphic LabsAI drug design
Alphabet-spinout drug design company that builds protein-structure and binding-affinity models to design drug candidates for pharma partners.
Latest · 12 May 2026Raised $2.1 billion in Series B funding, led by Thrive Capital with Alphabet, GV and new sovereign-wealth investors, to scale its drug design engine. Isomorphic Labs ↗
$2.1B Series B, 12 May 2026 source ↗
Anteriorpayer prior authorization AI
AI that reviews clinical documentation against coverage policy for health insurers to speed up prior authorization decisions.
Latest · 12 Feb 2026Closed a $40 million round, reporting 99.24% clinical accuracy independently validated by KLAS Research and a roughly 75% cut in review cycles at one customer. PR Newswire (Anterior) ↗
$40M undisclosed (total raised $64M), 12 Feb 2026 source ↗
OpenEvidenceclinical decision support for physicians
Free, ad-supported question-answering tool for licensed physicians to look up evidence-based clinical answers at the point of care.
Latest · 22 Jan 2026Closed a $250 million round led by Thrive Capital and DST, doubling its valuation to $12 billion as physician adoption accelerated. HLTH ↗
$250M Series D, 22 Jan 2026 source ↗
Aidocradiology triage AI
AI that flags urgent findings on CT scans for radiologists and care teams, now built on its own foundation model rather than one algorithm per condition.
Latest · 21 Jan 2026Received FDA clearance for CARE, a foundation model that triages 14 conditions from a single abdominal CT scan. PR Newswire (Aidoc) ↗
$150M Series E, 29 Apr 2026 source ↗
Ambience Healthcareambient clinical documentation and coding
Ambient AI platform for documentation, medical coding and clinical documentation integrity across ambulatory, ED and inpatient settings.
Latest · 29 July 2025Raised $243 million in a Series C to scale its ambient documentation and coding platform for health systems. Ambience Healthcare ↗
$243M Series C, 29 July 2025 source ↗
Viz.aicare-coordination imaging AI
AI that detects time-critical findings like stroke and hemorrhage on imaging and automatically alerts the care team to coordinate treatment.
Latest · 12 June 2025Received the first and only FDA clearance for automatically quantifying subdural hemorrhage volume, thickness and midline shift on CT. Viz.ai ↗
Your move
Our read, built on the receipts above.
Enterprise ambient scribing is now a three-way war between Abridge, Ambience and Microsoft's Nuance, don't enter it head-on. The open space is narrower: revenue-cycle and appeals automation on the payer side beyond prior auth intake, and agentic patient intake/triage, which is still mostly voice-agent demos, not deployed infrastructure.
Watch retention, not the headline note-drafting demo: Abridge citing over 90% of clinicians who start using the product still using it later is the tell that it works day to day. Red flag in a drug-discovery pitch: any deck that quotes Phase I success rates without also giving you the Phase II number, since Phase II is where AI-discovered molecules are currently performing at ordinary industry rates, not better.
This quarter, pull ten ambient-scribe notes and compare them against your own memory or the recording, specifically hunting for omitted negatives (a symptom the patient denied that never made it into the note), not typos. That is the actual audit gap, and it's a fifteen-minute check per chart, not a research project.
The valuable skill now is fast, confident auditing of AI-generated clinical output, catching a wrong dose or a fabricated contraindication before it reaches a patient, since courts are already treating the physician's review as the point where responsibility sits. First step: read your specialty's entry on FDA's public AI-enabled device list so you know which tools in your own workflow are actually cleared versus just marketed as AI.
What's next
- 19 Oct 2026 Comment period closes on FDA's discussion paper on regulating generative AI-enabled medical devices fda.gov ↗
- 1 Oct 2027 CMS deadline for payers to support electronic prior authorization for medical-benefit drugs over FHIR APIs cms.gov ↗
- Sept 2027 Primary endpoint window closes for Insilico's 52-week Phase III trial of rentosertib (GENESIS-IPF-3), the first AI-designed drug in Phase III insilico.com ↗
Other industries
- FinanceBanks move agentic AI from chatbots to running back-office and dealmaking work
- Software engineeringSpaceX buys the leading AI coding startup outright, for $60 billion
- LegalLegal AI vendors go from research assistant to agent that produces the deliverable
- Real estateLeasing and resident calls move to voice AI agents that never sleep
- Sales and marketingCustomer support agents now get paid per resolved ticket, not per seat
- Jobs, HR and recruitingAI runs the first-round interview, not just the resume filter
- InsuranceClaims move from adjuster-first to AI-first, humans supervise the exceptions
- CybersecurityAutonomous AI agents take over penetration testing and vulnerability hunting
- Consulting and accountingAI-written client reports become a recurring, public Big Four embarrassment
- Retail and e-commerceShopping agents get a checkout button, and Amazon spends 2026 deciding who's allowed to press it
- EducationNYC and LA ban AI for students, not for teachers
- Media and creativeAI video splits into model labs that train and aggregators that resell
- Manufacturing and roboticsHumanoid robots move from demo livestreams to paid warehouse shifts
- Transport and robotaxisRobotaxis go from demo to metered fare across a dozen-plus US cities
- Energy and data-centre powerGrid rules get rewritten mid-buildout, state by state
- GovernmentGSA turns AI adoption into a government-wide subscription, then flips 'free' to metered
- Agriculture and foodAutonomous tractor fleets learn to run the tools behind them, not just the wheel