The AI Health Index Brief
October 30 to November 13, 2026 · Published November 15, 2026
The week in one line
Healthcare AI spent two weeks building for the bad day. Agents learned to text the patient when the handoff to a human fails, scribes kept recording after the phone locked, and billing software started drafting the appeal when the claim came back denied. Heart imaging supplied the good news, with two randomized trials showing a CT scan can guide the treatment and not only describe the disease. One automation even learned to enter the authenticator code, which puts it ahead of most of us before coffee.
This issue covers two weeks, October 30 to November 13. It also picks up four entries dated October 29 that arrived after the last issue closed. That makes 111 entries across 93 vendors, 75 Verified at source and 36 Partially Verified. Product changes led with 61 entries and clinical evidence followed with 23.
The Brief covers the entries that say something about where the category is going. The rest are in the change log, with sources.
The heart scan started changing the plan
Two cardiology meetings fell inside this window, TCT and the American Heart Association Scientific Sessions. The imaging vendors arrived with trials instead of accuracy tables.
Heartflow had the strongest entry. The P4 trial randomized 1,105 patients at 23 centers and was published in NEJM. It asked whether a stent procedure planned from a CT scan with Heartflow software does as well as one guided by intravascular ultrasound, the imaging catheter threaded into the artery.
It did. Target vessel failure at 12 months was 4.1 percent against 4.6, and the median procedure took 46 minutes against 63. Seventeen minutes a case is the number a cath lab manager will circle. The trial used an earlier version of the planning tool Heartflow now sells as PCI Navigator, which is worth knowing before anyone quotes it.
Heartflow also released one year results from DECIDE, a registry of more than 6,000 matched patients. Adding its plaque analysis to a standard coronary CT went with 34 percent fewer major cardiac events. A registry is not a trial, and these figures are company reported with no paper yet, so read them as a strong hint.
Cleerly presented PARAMOUNT, a randomized trial of 530 patients at 21 sites. It compared care guided by Cleerly’s plaque analysis with usual care guided by a stress test, and it met its primary endpoint at 180 days.
The telling detail is what the doctors did next. Use of cholesterol lowering therapy rose from 31 to 75 percent in the plaque guided group, against 32 to 47 percent with usual care. The results are not yet peer reviewed, and six months is a short horizon.
Caristo Diagnostics got the journal. JAMA Cardiology published an Oxford led study of 19,110 patients on CaRi-Heart, which measures inflammation around the coronary arteries from a routine CT.
Patients with a calcium score of zero are usually told they are low risk. In this study, very high inflammation in those patients was associated with a 14.1 times higher risk of cardiovascular death. A calcium score of zero has been cardiology’s version of a clean bill of health. This study suggests reading the fine print.
Elucid reported 92.9 percent accuracy against invasive measurement across 491 vessels. Then it did something more useful than the headline. It published the result in the gray zone near the treatment threshold, where the call is hardest, and accuracy there was 84.1 percent. A vendor that shows its number where the test is weakest expects to be checked.
The payoff for this kind of work showed up in prostate cancer. In October ArteraAI brought trial evidence to ASTRO. This month it announced that its test for men after prostate surgery is included in the NCCN guidelines, which is what physicians and payers read before they act. The test has been on sale since December 2025. The guideline listing is the news.
Our readFor years the pitch for imaging AI was accuracy, meaning the software agrees with the expert. This period the pitch was consequence. The scan shortened the procedure, changed the prescription, or found risk in a patient who had been told there was none. That is the evidence that moves a guideline committee and then a payer, in that order. For product marketers it also retires the accuracy slide. Once a competitor has a randomized trial, 95 percent agreement with a radiologist reads like a participation trophy.
The agent learned what to do when it gets stuck
Last issue the phone agents got scripts they could not skip. This period the vendors worked on the moment after the script runs out.
Artera’s voice agent now texts the patient automatically when a transfer to a staff member is abandoned. It also reroutes a handoff addressed to nobody instead of dropping it. Anyone who has been transferred into silence will understand the feature.
Artera’s new Intake Hub does the same for paperwork. Whatever the automation cannot verify or finish, such as a failed identity check, lands in a queue for staff.
Luma Health released Agent Launchpad, which runs several agents across a workflow that can last weeks, starting with procedure prep. Clinical questions go to the care team with the full conversation attached, and a nurse can resume, skip or stop the workflow.
Colonoscopy prep is live at UAMS and Mile Bluff Medical Center. Luma says half of Mile Bluff’s patients arrived fully prepared with no staff calls. Colonoscopy prep is a fair test of any communication system, human or otherwise.
Scribeberry launched an AI receptionist that answers clinic calls in English and French. Each job it can do has a switch with three positions, which are off, collect details, and do it for me. Emergency symptoms always get a 911 instruction and a staff alert, whatever the switch says.
That switch is the most candid piece of interface design in the log this period. It admits that trust is granted one task at a time.
Spring Health now lets eligible adult members in its Specialty Care program choose an AI check in with Guide in place of a live session with a human Navigator. A summary of each check in goes to the Navigator. These are higher acuity members, so the question for a buyer is how quickly a person reads that summary.
Smaller releases carried the same idea. Attune’s voice agents now record the words that triggered an emergency escalation. Careswitch’s scheduling agent warns before it double books a caregiver.
Three voice vendors also published numbers on how their agents behave under pressure. Infinitus reports that its agents finish the task 95 percent of the time after a call goes off track, and reach the right place in major payer phone trees 92 percent of the time. That second figure would be a personal best for most humans.
Sully.ai runs Arabic and English speech recognition side by side for Middle East hospital calls, and has the agent ask the caller to repeat when it is unsure. Hippocratic AI says the voice in its new preview release pronounced 90 percent of 629 drug names correctly.
All of these figures are self reported. They are still the right kind of number, because they measure the hard cases.
The best evidence in this group came from nurses. Epic published results for Art, which sorts patient portal messages and drafts replies that nurses review before sending. Advocate Health reports 1.5 minutes saved per message and two full time nurses moved from message triage to patient phone calls. At OhioHealth, messages flagged urgent were read 2.5 hours faster than routine ones.
Our readThe first generation of agent launches described what the agent can do. This one described what happens when it cannot. That is a better sign of maturity than any new capability, because the failed handoff is where patients get lost and where buyers get nervous. For product teams, the exception path is a feature and deserves its own line in the release notes. Several vendors now give it one.
The denial got an agent
Last issue prior authorization got an agent on both ends. This period the agents followed the claim downstream to the part everyone dreads, the denial.
R1 made Phare Denials generally available. It follows a denial from early detection to a submitted appeal, and drafts the appeal from the clinical facts, the payer’s policy and the contract language. R1’s purchase of Humata Health also reached the product, with Humata’s prior authorization AI now running as Phare Auth.
R1 says clinician oversight remains essential for complex clinical denials. It does not describe a required review step before an appeal goes out, and that is the question to ask.
Candid Health launched AgentKits, starting with denials. The pitch is that a practice keeps its existing billing system and plugs the agents in. They call payers, check portals and prepare appeals, and true exceptions go to a biller to approve or override.
The audit log records whether a rule or an agent took each action. That detail matters more than it sounds. When something goes wrong, the system did it is not an answer. An agent did it and here is the log is at least the start of one.
The plans were busy too. Cohere Health reported that an unnamed Medicare Advantage client found more than $10 million in savings in six months by reviewing high cost drug and equipment claims before paying them. Charta Health launched a suite for payers on the same chart review engine it sells to providers. It flags conditions that were documented but never coded, and codes the chart does not support.
Selling the same reader to both sides is a bold move and possibly a healthy one. At least the two parties would be arguing about the same reading of the chart.
One entry showed what the end state looks like. Surescripts and athenahealth switched on automated medication prior authorization for nearly 98 percent of prescribers on athenaOne. When the payer’s criteria are met, approval can come back automatically, and Surescripts cites a median of 18 seconds from 2025. Eighteen seconds raises a fair question about what the previous several days were for.
Then OpenEvidence made part of this free. Its new Access Intelligence adds cost and coverage to its answers, drawing on more than 2,500 patient support programs. It also drafts prior authorization requests, appeals and letters of medical necessity, at no charge to clinicians.
Free is a hard price to compete with. Vendors selling prior authorization tools to practices now have the problem the standalone scribes had last issue, when athenahealth included ambient documentation at no added cost.
Automation only pays if the output survives review, and Regard has a number for that. At a Becker’s roundtable, the medical director of clinical documentation at Cedars-Sinai presented a matched comparison of 793 Regard encounters, which showed a 3.1 percent higher case mix index.
The more useful figure came second. Of the new diagnoses Regard suggested from January to August, 52.3 percent cleared physician, documentation, coding and compliance review. About half is an unglamorous number, and it tells a buyer how much of the suggestion list is real.
For anyone who believes healthcare has moved past the fax machine, Luminai published results from Cleveland Clinic. Its automation handles about 770,000 referral and order faxes a year there. Handling time fell from ten minutes to under one, and a three day backlog is gone. The figures are vendor reported. The fax is not dead. It has acquired a reader.
One note for the security team. GenHealth.ai’s portal automations can now complete sign ins that ask for an authenticator code. That opens more payer portals to automation, and it is worth a conversation about whose second factor it is.
Our readBoth sides of the claim now have software drafting letters to the other side. That will make the argument faster, and it may not make it shorter. The durable products here are the ones that send the result upstream. Candid turns recurring denials into rules that run before submission, and R1 feeds denial patterns back into authorization. Winning the appeal is a service. Not needing the appeal is a product.
The scribe learned not to lose the visit
Four ambient scribes shipped major phone app releases within a week of each other. The release notes read like a list of things that have gone wrong in exam rooms.
Vero Scribe now keeps recording after the phone locks. HealOS holds the recording on the phone until the server confirms each part is saved. Chartnote lets a clinician start on the phone and finish on the computer, and keeps deleted notes for 30 days. Doximity Scribe records with no connection at all and syncs later.
Commure fixed a mistake most clinicians have made at least once. A note recorded against the wrong patient can now be moved to the right one, reusing the recording and regenerating the note with the right patient’s context. Notes already signed in the EHR cannot be moved.
Freed rebuilt its iPhone and iPad app to match the web product, with precharting, coding and letters now in the app. Freed also says plainly that the new version is less resilient on a weak connection than the old one. A limitation stated in the release notes is one a buyer does not have to discover in a basement clinic.
The outcome number in this group came from the emergency department. Ambience Healthcare announced a KLAS report on its use at Memorial Hermann, covering about 100 emergency physicians over 15 months. It cites 1.7 more patient visits per physician per ten hour shift than nonusers, and about $37,500 in added annual productivity per physician.
Our readLast issue a handheld ultrasound was recalled because it sometimes needed a signal before it would scan. The scribes appear to have read the same memo. Nothing in this batch would survive a keynote, and all of it decides whether a clinician still trusts the app in month three. Reliability is the feature nobody demos and everybody churns over. Security teams get homework too, since audio that waits on a phone is audio stored on a phone. Ask how it is encrypted and when it is deleted.
The assistant started taking guests
Microsoft Dragon Copilot 6.0 opened Copilot Chat to apps and agents built by other vendors and by customers. They arrive in public preview for US physicians, loaded by an admin or bought through Microsoft Marketplace. A second preview feature, Show AI, highlights which parts of a note the AI wrote or changed.
Doximity Ask added a Check with UpToDate button. A clinician asks a question, then taps to see an UpToDate answer beside Doximity’s own. A product that invites a second opinion on itself is confident, or wise, or both.
Abridge went the other direction and brought the customer’s content in. Its assistant can now pull a health system’s own care pathways into an answer, with a link to the source document.
Google’s Gemini for healthcare picked up two guests on the same day. Infinitus and Synthpop both launched connectors, so staff can start benefits checks, prior authorization follow ups and referral intake from inside Gemini. Owkin made K Pro drivable from outside assistants such as Claude, so a researcher can start an analysis without leaving the chat they already have open.
Commure took the older route to the same place. Its documentation and coding AI is now available inside iKnowMed, Ontada’s oncology EHR, which makes Commure the guest and the EHR the host.
Our readIn October the Brief said distribution was turning into the product. This period the assistants started acting like it. Dragon Copilot is becoming a storefront, Gemini a lobby, and Doximity a place where two reference sources share a screen. Product teams now have to decide whether they are the host or the guest, and for most the answer is guest. Buyers inherit a new review task. When an agent from one vendor runs inside the assistant of another, someone has to say whose rules apply and whose name is on the answer.
Patient data got an expiry date
Phreesia set a six year limit on the patient session data it stores, after which sessions are permanently deleted. Data already sent to the EHR is not affected. Heidi Health gave team admins automatic deletion for its evidence chats, with a retention window they choose.
Both are small settings. Both replace a policy of hoping someone remembers.
Axonius for Healthcare added a switch that controls whether its own support staff can reach the monitoring appliances inside a hospital network. Turning it off takes effect immediately, and every access session is logged.
The reminder of why this matters arrived on schedule. Myriad Genetics disclosed unauthorized access to its network on November 2 and suspended testing and ordering systems while it responded. Its customer notice says systems are restored. The investigation into whether patient records were accessed continues.
Vendors also kept saying where the model runs. Corti moved the efficient tier behind its mini model IDs to Qwen 3.8 in the EU, while US projects stay on OpenAI models hosted on Azure. Dyania Health said its models run on NVIDIA hardware inside the hospital’s own network. Sully’s Arabic decision model runs in the UAE.
Our readRetention used to live in a policy document nobody could find. It is turning into a setting with a number in it, which is the only form a security team can check. The Index Answer below counts the vendors that publish that number.
Market notes
Cellarity’s AI drug discovery platform now belongs to someone else. Flagship Pioneering formed a new company, Etcelera, to carry the platform forward, and Cellarity continues as a clinical company built around its sickle cell candidate. Anyone evaluating the platform is now evaluating Etcelera.
Needletail AI published its prices, which remains rare enough to report. Dental insurance verification now costs $399 or $499 per location per month with 400 verifications included, then $1.00 or $1.25 each.
AI4Eyes said it can now sell its eye care platform in the United States after completing FDA registration and listing. Registration and listing is not a clearance, and buyers should not read it as an FDA review of the AI.
GE HealthCare added two cleared AI tools to Vscan Air SL. One guides a user through a lung exam, and GE’s validation has nonexpert users capturing diagnostic quality exams 95.2 percent of the time against 96.8 for experts. The other estimates moderate to severe aortic stenosis.
Ceribell launched delirium monitoring on the same headband ICUs already use for seizure detection, with no outcome results reported yet. Butterfly Network said it achieved FedRAMP authorization, which opens the door to federal buyers such as the VA.
Apella published results from Baptist Memorial Hospital in Memphis, where first cases starting on time rose from 36 to 78 percent in nine months. The side finding is the fun one. Apella’s cameras had 36 percent of first cases starting on time, while the manually entered records said nearly half. The room and the chart had been keeping different clocks.
LeanTaaS launched iQueue for Virtual Care, its first product built on the computer vision it acquired with LookDeep Health. It watches enrolled patient rooms continuously and escalates to bedside staff.
Senior living got a two way connection. Sage and August Health finished the second phase of their integration, so care tasks reach the phones caregivers carry and completed documentation flows back into the record. SafelyYou’s new link with ECP runs one way for now, with fall details into ECP described as the next step.
Guardant Health had a study in ESMO Open of 130 patients having surgery for colorectal cancer metastases. Residual disease found in blood after surgery went with a 4.7 times higher hazard of recurrence or death. GeneDx reported in conference posters that 62 percent of positive exome and genome results in a very large cohort were medically actionable.
GRAIL addressed the standing objection to cancer blood tests, which is the worry. In a study of 35,350 adults, average mental health scores stayed in the normal range whatever the result. Participants with a true positive dipped after the test and moved back toward baseline by 12 months.
Oatmeal Health posted its first LungAI study as a preprint. Combined with standard Lung-RADS scoring, it cut the cancers sent to a 12 month rescan by 46 percent. The paper also states that its accuracy gain over Lung-RADS alone was not significant, and credit is due for saying so.
In drug discovery, Edison Scientific released ChefNMR2, a model that works out a molecule’s structure from its NMR spectrum. It ranked the correct structure first 82.4 percent of the time against 51.9 for the best general model tested, and Edison published the weights. Schrödinger moved Generative Glide from beta to full release.
PhaseV and Causaly both launched research agents that begin by writing out a plan for scientists to review. It is the research version of showing your work.
OpenEvidence also launched Learn, a free board prep and continuing education product. Synapse Medicine opened its MedGPT assistant to clinicians in ten more European countries, in eight languages, applying the guidelines of the country where the user practices.
Corti raised Spanish to its top language tier, with a reported 28.4 percent relative drop in word error rate. It also released a beta tool that turns a free text note into structured FHIR data, with each value linked back to the source words.
What the two weeks say about the category
One hundred and eleven entries, and the theme is the bad day. The agent got a plan for the handoff that fails. The scribe got a plan for the locked phone and the wrong chart. The billing system got a plan for the denial, and the intake tool got a date on which it forgets.
None of that is exciting, and that is the point. A category shows it is growing up when its release notes describe failure in specific terms. The vendors that can say what happens when things go wrong are usually the ones worth trusting when things go right.
The good day had a strong showing too. Heart imaging brought randomized trials, and a guideline committee added an AI test to its recommendations. The pattern holds across both halves. The claims that travel are the ones a skeptic can check.
Which healthcare AI vendors say what happens to patient data?
Nearly all of them say something, and very few say it precisely. Of the 687 vendors the AI Health Index has assessed on AI Safety and PHI Stewardship as of November 16, 2026, 27 grade A. An A requires retention windows, training use and de identification stated specifically enough that the statement could be proven wrong. It also requires the safety engineering in writing, meaning guardrails, hallucination mitigation and what happens after a safety event.
Another 245 grade B. The remaining 415, 366 at C and 49 at D, offer general assurances or less. A promise that privacy is taken seriously grades C, because it answers the question buyers asked before AI and skips the ones AI raises.
This period supplied examples of the precise kind. Phreesia put a number on retention, six years, after which stored intake sessions are deleted. Heidi gave admins a deletion window to set for evidence chats. Scribeberry says its phone receptionist does not record calls and keeps transcripts in Canada for a year by default. Each of those is a sentence a security team can check, and that is the whole test.
The practical move is to ask for the window per artifact. Ask how long the audio, the transcript, the output and the logs are each kept, and whether any of it trains a model belonging to anyone in the chain. A vendor graded A here has usually answered in writing already.
Full grades, the axis definition and what separates each band are on the AI Safety and PHI Stewardship page.
The AI Health Index Brief is published weekly by AI Health Index, an independent reference for evaluating AI vendors in healthcare. No vendor pays for inclusion, placement, or rating. Compare any indexed vendors by capability at Compare and read the evaluation standards at Methodology.