The AI Health Index Brief

The week in one line

athenahealth switched on ambient documentation for its primary care users at no added cost, the way a car comes with floor mats, and the standalone scribes spent the same three weeks claiming the order, the code and the posted payment. A dosing checker found more than two in every hundred AI dosing claims wrong, a handheld ultrasound was recalled for needing a signal, and the model behind the product finally got a name. At 120 entries, it is the largest issue the Brief has carried.

This issue covers three weeks, October 10 to October 29, since no issue went out on the 17th or the 24th. It is the largest issue the Brief has carried: 120 entries across 99 vendors, 102 Verified at source and 18 Partially Verified. Product changes led with 59 entries. Clinical evidence followed with 21, EHR integration with 12, safety and governance with 8 and model changes with 7, and the rest were regulatory decisions, pricing, security, two product sunsets and one acquisition.

The scribe came standard

With athenahealth’s Fall 2026 release, athenaAmbient switches on automatically for family medicine and internal medicine providers on athenaOne from October 29. It is included with athenaOne at no added cost, tuned for primary care, and English only for some specialties. Other specialties can join an open beta, and practices that opted in after October 15 activate on December 3.

For a primary care practice on athenaOne, the question stopped being which ambient scribe to buy and became whether to buy one at all. Every standalone scribe selling into that base now answers that question on the first call, which is a hard place to open.

ModMed launched an AI package for general surgery that extends ModMed Scribe. It drafts the note from the patient conversation and works with the EMA EHR to suggest surgical ICD-10 codes for the surgeon to review, with colorectal, breast, vascular, bariatric, endocrine, surgical oncology, and foregut and hepatobiliary surgery named. AI clinical summaries for chart review are planned.

A specialty roadmap used to be the standalone vendor’s best argument against the EHR’s built in scribe. ModMed just published one.

Commure Autonomous Coding received Epic Toolbox designation in the Fully Autonomous Coding category on Epic Showroom, joining the designation Commure Ambient AI already holds. The engine reads the signed note, submitted charges and EHR data and recommends ICD-10, CPT, E&M and modifier codes with a rationale for coder review. OB Hospitalist Group reports clinicians spent 83 percent less time entering charges within three months, with over 85 percent of charges coded by AI.

Chartnote Teams now includes one assistant seat with every clinician subscription at no extra cost and no recording caps. A medical assistant records the intake while rooming the patient, the clinician gets the recording, and AI Scribe folds intake and encounter into one note the clinician creates and signs. Assistants cannot create clinical notes, which keeps the line between who records and who documents.

The standalone vendors spent the same three weeks moving down the pipe, toward the steps after the signature.

Abridge’s October release lets outpatient clinicians on Epic capture orders during the visit. Orders discussed with the patient appear in the Heard in Visit Recording module as the conversation happens, with new orders added, changed orders updated in place and retracted orders marked, all staged for the clinician to review and sign. For inpatient notes, Abridge now suggests diagnoses with their specificity and present on admission status, and accepted ones are written straight into the chart.

Order entry is where much of the after visit work goes. Abridge moved it into the visit and kept the signature where it was.

Vero Scribe added an encounter coding workspace, where a clinician asks the chat to code the visit and reviews an editable table of billing lines, diagnoses, amounts and totals. Two weeks later the chat began returning ICD codes specific to the visit, checked against the code catalog the user selected, and filling an attached PDF form. Vero states each code still needs verification by a clinician or qualified coder.

DrChrono opened EverHealth Scribe BillingAssist to every practice. When Scribe drafts the assessment and diagnoses it now suggests CPT codes that follow billing rules, switched on under Preferences, and it does not yet use DrChrono billing profiles. After Visit Summaries got their own template in early access, with full availability planned for January 2027.

Elation Health added AI payment posting to Elation Billing, the last step in a flow where the signed note drives AI drafted codes, claims are checked against eligibility, coding edits and payer rules, and payments post on their own. Partial payments, complex transactions and denials go to a person with the reason and a suggested next step. It is in beta, available to new customers from November 1, and Elation’s figure of more than 70 percent less manual work per claim covers its AI billing since March, not payment posting alone.

Suki’s iOS app now keeps the HCC insights athenaOne clinicians add during pre charting in the note as problems, with ICD-10 and HCC codes attached, so risk adjustment work done before the visit survives into the documentation. AvodahMed’s release 26.17.14 tightened how notes and orders reach Veradigm: ICD-10 selections are checked before saving, orders that need a procedure result entry rule now save, and launching from Veradigm reuses the matching appointment instead of duplicating it. It also added an OB/GYN specialty and a 14 day free trial for Practice Fusion users. One line in the notes deserves a second read: notes can now be saved to the EHR before anyone has reviewed them.

Our read

When the EHR gives the note away, the note stops being a product and becomes a feature. The standalone vendors answered by claiming the steps after the signature: the order, the code, the claim, the posted payment. That is the right roadmap, and it moves the oversight question with it. A draft note waits for a signature. A posted payment has already happened. Nearly every vendor here kept a review step in the flow, which is why the one release that lets a note save before review stands out. The scribe market is not shrinking. It is being asked to do the job the note was always a prelude to.

The scribe picked up the phone

Scribeberry Pro users can now call patients at Canadian and US numbers, or text them a browser video link that needs no app or account, from the web or the mobile app. The scribe starts when the patient answers and writes the note from the call, and the caller ID can show the Scribeberry number, the clinician’s own verified number or a clinic number. There is no automated announcement that the call is being transcribed, so clinicians need their own consent script.

A week earlier Scribeberry brought Berry, its clinical agent, into its iPhone and Android app. With an Accuro connected scribe, Berry answers questions about the chart with dates, fills in the Canadian Disability Tax Credit form from the visit and chart while leaving undocumented answers blank, and drafts a referral and fax once the clinician has reviewed the letter. The fax is reported as queued, with delivery tracked separately. Leaving a blank where the record is silent is a small design choice with a lot of judgment in it.

Nabla released the first beta of a browser SDK for EHR and partner pages that embed its encounter frame. The host page can read the encounter’s state, start and stop recording, trigger note generation and export the note, and listen for events when a note or patient instructions are exported. Nabla switches it on per organization and customers list the sites allowed to embed it. The same week its Connect API added Dutch for the Netherlands and for Belgium.

Enzo Health released Translate for Enzo Scribe and Enzo EHR, so a home health clinician and a patient can each speak their own language during the visit. The transcript keeps the language spoken and the documentation is written in English, with eight languages at launch and a review step before signing. Corti’s coding product made ICD-O-3, the WHO oncology classification used by cancer registries, generally available in English, German and French, returning topography and morphology codes as a pair, and added Norwegian, Austrian and Dutch editions of SNOMED CT in beta.

MedicalScribe.app shipped desktop apps for Mac and Linux that record system audio as well as the microphone, so both sides of a telehealth call are captured, and a week later began accepting up to 10 PDF or Word files per visit, with an option to draft a note from attachments alone. Version 21.4.0 of the Doximity iOS app lets clinicians attach a Doximity Scribe visit note to an Ask conversation and ask follow up questions with that encounter as context.

Our read

With the note itself turning into a line item, the scribes differentiated on everything around it: a phone call, a fax, a government form, a Flemish patient, a German cancer registry. None of these is a model advance. All of them are the paperwork of a particular place, and the scribe that wins a specialty or a country will be the one that knows its paperwork. Product teams planning a 2027 roadmap should notice how much of this period’s news is a form.

The vendor graded itself and posted the score

Doximity launched DoseCheck for Doximity Ask, a rule based check that tests every dosing claim in an answer against a structured drug database built from FDA labeling and professional guidelines. The design choice is the point: a deterministic check against the label, rather than a second model asked to grade the first.

Doximity tested nine widely used AI models on more than 8,000 real physician dosing questions and found at least 2.3 percent of dosing claims wrong, and 43.8 percent of those errors serious or potentially lethal. DoseCheck is free and live for all Ask users. Doximity says it lifts dosing accuracy above 99 percent, and that testing is its own and unpublished.

Two in a hundred sounds small until it is multiplied by every dosing question a clinic asks in a year. Nearly half of those being serious is the number that gets a feature funded.

Sully.ai published how it evaluates its clinical notes. Every prompt, model or harness change is scored offline on fixed encounter sets against rubrics built with clinicians and judged by several LLMs from different families, which agreed 99.04 percent of the time. In production, each clinician edit is classed as a fabrication, an omission, a misinterpretation or a style change, with a severity. Sully reports fabrications down 83.1 percent, omissions down 69.1 percent and misinterpretations down 87.1 percent at one ambulatory clinic, self reported and not peer reviewed.

The clinician’s edits are the ground truth here, which is the one dataset every scribe vendor already has and few have been reading.

Vera Health described its own version. A jury of AI judges built on several model families and tuned on clinicians’ blinded reviews grades every change against a fixed set of questions across specialties and countries, and a change ships only if it does at least as well as the current version. The judges also grade a daily sample of live answers, and flagged answers reach a person each morning. On 50 held out comparisons Vera says its judges matched clinicians 82 percent of the time, against 72 percent for a general purpose grader.

A day later Vera used the method on its competitors, running 100 questions from the public Real-POCQi dataset through Vera and three frontier models and reusing the study’s published answers from OpenEvidence and GPT-5.5. Graded blind by the study’s AI judges, Vera was preferred over OpenEvidence on clinical accuracy 80 percent to 7, and Claude Opus 5.5 led on clinical utility and completeness. All grading was by AI judges, and Vera ran the comparison.

Abridge opened an Evaluation Dashboard in its Enterprise Portal for the governance and clinical informatics leaders at partner health systems. It shows evaluation results by product and quality dimension, gives each measure’s method, refresh frequency and known limitations, keeps monthly metrics apart from published studies, and links each result to the model card. The results describe Abridge’s products overall rather than any one site’s data, and only partners can see them.

Commure Pro, the former PatientKeeper, now ends every AI summary with a disclaimer, on by default and worded per site, that the EHR is the system of record. Its chat audit API gained options to leave out named fields and to drop rows for a given user, and a new AI Incidental Findings dashboard lists patients with findings that need follow up after discharge, switched on by request.

Microsoft’s October release of Dragon Copilot for radiologists moved feature control into Dragon admin center, where admins switch Chat, AI Draft and Optimization on or off at the organization, unit and user level, with lower levels inheriting and a lock to stop overrides. The radiologist product remains in public preview in the US.

Our read

A year ago an accuracy claim was a slide. This period it was a dashboard, a published rubric, a deterministic checker and an admin console. Evaluation has become a product surface, with a login and release notes of its own. The catch sits in the methods: nearly every grader here is another model and nearly every result is the vendor’s own. That is still progress, because a published method can be argued with and a slide cannot. Product teams should expect the RFP question to change by spring, from how accurate the product is to how the vendor checks, how often, and who can see the result.

The phone agent got a script it cannot skip

Artera’s October 21 release of Artera Harmony Federal Edition added a required emergency triage step to its voice AI agents, enforced in code rather than left to the agent’s prompt. When it triggers, the agent speaks a set script and transfers the caller straight to a destination the organization configures. Some updates need switching on by Artera’s project team.

Six days later the same agents could confirm, cancel or request a reschedule of an appointment during a call, with the change written back to the appointment and reports showing the date, time, provider and whether the patient answered by voice or text. The release also lets organizations set a standard AI disclosure message for each agent, off, in shadow mode or on, added to the first outbound text and to voicemail and measured on every send.

An agent on a patient phone line cannot skip the hand off, which puts a hard limit on what it handles alone when a caller may be in danger. The disclosure setting gives a compliance team a record that patients were told who they were talking to.

Doctolib’s AI phone assistant for French practices now asks callers the same qualification questions a practice set for online booking, blocks the booking when an answer fails the practice’s rules, and records the answers in the call log and the appointment. Practices can write their own answers to common requests and choose a voice. Questions that ask for a code, a date or a social security number are not supported yet.

That closes a gap where a phone call could book what the website would refuse. In Germany, Doctolib made the assistant available to hospitals, MVZ outpatient centers and multi site groups, booking, moving and canceling around the clock in the Doctolib calendar and logging every call, including missed and dropped ones. The hospital page sells it as an annual subscription of at least ten licenses, and the release does not mention a connection to hospital information systems.

NeuroFlow’s October 21 release makes a Yes to any of the five ASQ suicide risk screening questions raise an urgent alert for the care team, listing each Yes answer, whether the patient answered in the app, on a clinic tablet, by text or by email. Patients who answer Yes now see a crisis screen with 988 resources, and the tablet flow lets a caregiver identify themselves after the patient is verified, with separate questionnaires and a record of who completed each.

Last issue NeuroFlow moved urgent alerts from the next day to the same day. This issue it made sure the alert fires the same way whatever the channel. The product is closing the gap it exists to close, one release at a time.

Phreesia release 26.14 lets VoiceAI explain a caller’s outstanding balance and collect payment during the call, after verifying identity, using the card on file or a secure payment link. A photographed insurance card now fills in the payer and member ID for the patient to confirm, and early adopters saw 3 percent more patients arrive with eligible insurance. In early access, Referral Automation turns incoming referral faxes into appointment requests, with an AI marker on the fields that need review.

Relatient extended Dash Self so patients can view, book, cancel, reschedule, confirm and update appointments without staff, verified by two factor authentication and with no account to create. Every change is checked against provider preferences, open slots and visit type rules, which is what keeps self service from breaking the schedule. In testing across thousands of appointments, 91 percent of users finished without a call.

Medsender launched Agentic Referral Management. It reads each incoming referral in under 30 seconds, pulling demographics, insurance, referring provider and visit reason, then contacts the patient by text, email and phone with a link to book, and posts the referral to the EHR once the patient books. Only exceptions reach staff, practices pay only for booked appointments, and the first Anne Arundel Dermatology location went from 10 percent to 48 percent conversion in its first month, a single site figure worth testing against your own.

Overjet launched Ivy, an AI insurance coordinator for dental practices that verifies eligibility and benefits, checks coverage limits and exclusions, calls insurers when the answer is not online, and combines the results with the practice’s fees into each patient’s cost estimate. Through Overjet’s ReviewPASS connection with participating payers, Ivy can also say whether a claim can skip manual clinical review. One practice says Ivy saves its sole coordinator 32 hours a week.

Infinitus’s payer calling agents can now identify the right pharmacy benefit manager from the BIN and PCN printed on most member ID cards, rather than from a payer ID mapped in advance, and for some payers pull the patient’s address on file to pass strict identity checks, which Infinitus says raised one customer’s success rate with CVS Caremark by 12 percent. Benefits results now also report how each payer accepts predetermination requests, by phone, fax, mail, website, email or electronically.

Our read

The agents on the phone did more this period and were fenced harder, often in the same release note. Three fences are worth copying: a hard stop written into code rather than a prompt, a rule set the voice agent inherits from the website, and a disclosure setting with a shadow mode and a count of every send. Each turns a safety promise into a feature with a toggle and a report, which means it will be compared in demos. The most important feature a voice agent shipped in these three weeks is the one that makes it stop talking.

Prior authorization got an agent on both ends

ZeOmega launched Jiva PrudenceAI, an AI layer for health plans that puts summaries, recommendations and prepared actions inside utilization and care management work. It covers authorization intake including faxes, guideline help and decision support for utilization review, call preparation, draft care plans and a combined member view. Plans can use ZeOmega’s apps, their own approved models or certified partner tools under one set of governance rules that decide which models run, what each may do, how outputs are checked and when clinicians review, and ZeOmega says any AI decision can be logged, explained and halted in real time.

Payers already on Jiva can switch it on without a platform migration and are not tied to ZeOmega’s own models. The written controls and the statement that member data stays in ZeOmega’s environment are terms a plan can test in a contract.

Cohere Health extended its intake, decision and review tools on Cohere Unify to medical benefit drugs, the specialty and provider administered drugs that run through a plan’s medical benefit rather than its pharmacy benefit. Requests that match plan policy are approved in real time and complex cases go to clinicians and pharmacists with the context attached. Pharmacy benefit drugs are not part of the launch, and Cohere says its FHIR connection will support the drug authorization rule CMS has proposed once it is final.

Lyric made expert review generally available in Lyric42, letting health plans hold selected claims during adjudication for a targeted review by an expert. Lyric42 finds and ranks the claims under criteria the plan sets, routes each to the right reviewer and returns a recommendation, while the plan makes the final call. Coding Review is the first review type, and Lyric says review outcomes feed back into new prepayment edits, so it is worth asking how those rules are approved before they go live.

Anterior open sourced a research preview of its symbolic reasoning library and an MCP server under the Apache 2.0 license. The library writes a health plan policy as a graph of narrow criteria, and a deterministic solver walks the graph, asks an LLM agent, a database or a human reviewer one criterion at a time, and returns approve, escalate or deny with a full reasoning trace. Anterior says it uses the approach across thousands of policies for prior authorization, HEDIS, claims and payment integrity.

This is the release to read. It keeps the policy logic outside the model, which is what lets the same facts produce the same decision twice.

Laguna Health’s Companion now shows each member’s top five open HEDIS care gaps in call prep and keeps them in view during the call, read only from the plan’s care gap system. At one national plan, comparing 2,749 calls before with 800 after, care gap conversations rose by 102 per 1,000 calls from a base of 16.4 percent. The study counts conversations rather than closed gaps.

Cencora said it has combined its benefit verification and electronic prior authorization network with Plenful’s automation platform, so health system pharmacy teams can check access, out of pocket costs, formulary and prior authorization in one workflow. Plenful’s claim that customers cut intake and authorization time by more than 75 percent comes without a method.

Our read

Provider side agents have been filing and appealing authorizations for a year. In these three weeks the plan side answered with agents of its own, and they arrived with the governance vocabulary already attached: which models may run, what each may do, who reviews, and a log that can be halted. For product teams on the provider side, the agent you built to argue with the plan is about to meet an agent built to answer it, and both will be quoting the same policy text. Anterior just published the grammar.

The slide started answering the genomic test’s questions

ESMO 2026 produced the heaviest evidence block this log has carried, and the pattern inside it was the ordinary H&E slide doing work that used to need a genomic assay.

Waiv presented an analysis of the TAILORx trial covering 7,584 patients and 532 distant recurrences, comparing the H&E only component of its CE marked RlapsRisk BC with the 21 gene recurrence score. On top of clinical factors the two predicted distant recurrence about equally well, with C indexes of 0.676 and 0.673, and combining them raised it to 0.701. The analysis is retrospective and the 21 gene test remains the guideline standard in the US.

Earlier in the month npj Precision Oncology published the multicenter validation of Waiv’s BRCAura, which predicts inherited BRCA1 and BRCA2 mutations from routine breast cancer slides. In 1,025 patients from six centers and four scanner types it reached an AUC of 0.81 on resections and 0.82 on biopsies, with sensitivity above 93 percent while ruling out about a third of patients without a mutation. It is research use only, the study was enriched for carriers, and a prospective study of about 2,600 patients runs to 2029.

Caris presented two breast cancer models. LDR-B, built with NSABP and NRG Oncology from routine slides plus age, nodal status and surgery type, was trained on 2,271 NSABP B-42 patients; those it scored high saw a 9.39 percent absolute ten year reduction in distant recurrence from extended letrozole, while low scorers saw no significant benefit. CapeAI predicts capecitabine response from exome and transcriptome data, and in 466 ER positive patients predicted responders lived 28.6 months against 19.2. Neither is an orderable test, and CapeAI’s external check in the PEARL trial showed longer progression free survival but not longer overall survival.

Tempus reported an analysis of 60,459 advanced lung cancer cases from 56,969 patients, done with Pfizer, in which an AI panel reading routine slides flagged 12 biomarkers with sensitivity above 80 percent for all of them and above 90 percent for several. The panel is still in development and is meant to triage testing rather than replace it, and Tempus did not say which biomarkers cleared 90.

Lunit showed where the slide and the pathologist disagree. In 993 resected lung cancers, SCOPE PD-L1 agreed with the pathologist’s tumor proportion score 73 percent of the time, and 204 of the 211 disagreements fell in the 1 to 10 percent band, which is the band a lab would want a second read on anyway. All SCOPE products remain research use only.

Veracyte moved Decipher Prostate one step further than last issue. A retrospective analysis of 401 patients from the randomized Phase III STAMPEDE trial showed the classifier predicts who benefits from adding abiraterone in clinically very high risk prostate cancer that has not spread. Adding it cut distant metastasis at five years by 10 points for high scores and 11 for very high, against 2 for lower scores, and the difference was significant in patients without node involvement. An earlier STAMPEDE analysis was recently added to the NCCN guidelines.

Myriad Genetics released MONSTAR-SCREEN-3 data on Precise MRD, its whole genome test for residual cancer. In resectable colorectal cancer it found tumor DNA before clinical recurrence in every recurrence observed, with a median lead time of four months, and more than half of positive results after surgery were below 100 parts per million. Lead time is not yet evidence that acting earlier changes outcomes, which is the data oncologists will want next.

Paradigm Health reported a retrospective evaluation of its trial screening platform on records from 3,027 lung cancer patients at Japan’s National Cancer Center Hospital East, presented with Fujitsu. The records were translated into English by an LLM and the platform ran entirely in Japan. Of 540 patients relevant to an EGFR trial it flagged 57 for review, 51 of whom were judged appropriate to prioritize, none of the 483 it set aside were trial appropriate, and it surfaced 35 likely RET fusion candidates who had never been tested.

Two lab studies made the economic case. At Pathology Visions, Primaa presented a prospective reader study at 11 French institutions in which 16 pathologists read 313 skin slides with and without its Cleo Skin software; median reading time fell 8 percent for classification and 23 percent for mitosis detection, and requests for extra immunohistochemistry stains fell from 217 to 182 slides per pathologist. Indica Labs published a case study on Source BioScience, which reports histopathology for more than 90 NHS trusts on HALO AP; across 5,721 cases over six months, digital turnaround averaged 1.8 days against about four for glass, with 18 percent reported the same day.

Our read

The H&E slide is the oldest and cheapest artifact in oncology, and in these three weeks three vendors reported models reading it that matched a genomic assay, screened for an inherited mutation or decided who needs a biomarker test. All retrospective, most research use only, every one with a prospective study on the roadmap. The product story is about time and cost rather than accuracy: the slide already exists, and the genomic test costs money and days. The honest marketing line is that the slide can tell you whom to test. The footnote that says retrospective will be the most argued word in every oncology deck for a year.

Three blood tests, three different numbers

Freenome launched SimpleScreen Lung, a laboratory developed blood test for adults 50 to 80 with a smoking history of 20 or more pack years who are not in guideline screening, orderable alone or with SimpleScreen CRC from one draw. In an independent cohort of 172 cases and 436 controls it showed 85 percent sensitivity at 50 percent specificity and a negative predictive value of 99.8 percent. Fewer than 20 percent of eligible adults are up to date with low dose CT, which is the problem the test exists to solve, and the FDA has not reviewed it.

Nine days earlier Freenome presented pivotal validation for an updated SimpleScreen CRC with assay and algorithm changes. Sensitivity for advanced precancerous lesions rose to 18.2 percent from 13.7 in the PREEMPT CRC study, and to 41.9 percent from 30.5 for lesions with high grade dysplasia, with colorectal cancer sensitivity of 80.4 percent at 90 percent specificity. The FDA approved test is the first generation, so the update needs its own regulatory path.

GRAIL brought a PATHFINDER 2 analysis to AMCP Nexus, a managed care audience, reporting Galleri in 15,953 adults aged 65 and older followed for 12 months. The cancer detection rate was 0.73 percent, positive predictive value 56.9 percent, specificity 99.44 percent and episode sensitivity 41.9 percent, and adding the test to recommended screening found about ten times as many cancers through screening, 129 against 13. Older adults are the population a Medicare decision would turn on, and a sensitivity of 41.9 percent means most cancers in the year were not flagged, so the test adds to standard screening rather than replacing it.

Guardant Health presented Shield data at ACG. A lifetime model of adults 45 to 75, using real world rates of follow up colonoscopy, found blood based screening averted 12 colorectal cancer deaths per 1,000 people against 6 for FIT and 9 for stool DNA, and a study at two rural South Carolina clinics reported 99 percent adherence. The death estimates come from a model, not a trial.

Veracyte’s Prosigna Breast Risk of Recurrence test received Medicare coverage through the MolDX program, effective August 19, for postmenopausal women with early stage ER positive disease and up to three positive nodes. Coverage removes the main barrier to ordering it for older patients. The payment rate was not disclosed.

Our read

Each test led with the number its reimbursement argument needs. GRAIL led with positive predictive value and specificity in an older population, because that is the Medicare question. Freenome led with sensitivity and a 99.8 percent negative predictive value, because the lung screening problem is people who never get scanned. Guardant led with adherence, because its case is that a test people finish beats a better test they skip. None of those choices is wrong. The product marketing lesson is to read the number the release did not lead with: 50 percent specificity is a coin flip for the healthy half of the cohort, which is a fine way to get people to a CT scanner and a poor sentence for a headline.

The device needed a signal

The FDA classified Philips’ correction of the Lumify handheld ultrasound as a Class I recall, posted October 16 after an early alert on October 9. A software issue can force the app to register the transducer again, which needs a wireless or cellular connection, and the USB link between transducer and phone or tablet can drop, leaving clinicians unable to scan. Philips reported eight serious injuries and four deaths as of September 29 and has told customers to use Lumify only where a connection is available, unless backup ultrasound is on hand, while it develops a software fix.

Class I is the most serious recall category the FDA has. Teams using Lumify in ambulances, aircraft or field care should check their backup plan, since those are the settings where a connection is least likely.

Clarius released App 13.1 for its handheld ultrasound, which now carries 30 AI models running on the device, up from 16 AI features reported last December. New in the release are automatic carotid intima media thickness measurement, tissue Doppler imaging, a split screen with live B mode beside Doppler or elastography, panoramic imaging, and voice control extended to annotations and measurements. Models that run on the device work without an internet connection, and Clarius notes that availability varies by scanner, market and regulatory authorization.

DeepEcho’s obstetric ultrasound app went live on Butterfly Garden, the first maternal health and first international partner app on Butterfly’s marketplace. It gives clinicians and trainees real time image quality feedback and anatomical guidance to capture standard fetal views, is listed for educational use only, and comes with a 30 day free trial.

Eko Health launched CORE 300, an FDA cleared stethoscope whose single chestpiece twists between traditional listening and three digital modes, with up to 35 times amplification, active noise cancellation and an on device heart rate display. Connected to the Eko App with an Eko+ membership it records and visualizes sounds and runs Eko’s cleared murmur AI. It lists at $249, currently $199, and the membership renews at $119.99 a year.

AIRS Medical shipped version 1.0.0.2 of SwiftSight Body Composition for the cloud ten days after announcing its clearance, adding a fallback for identifying the water and fat images the analysis depends on and producing a report even when height or weight is missing. Sites need the latest gateway to get it.

Our read

One handheld ultrasound was recalled because its software sometimes needs a connection before a clinician can scan. Another shipped thirty models that run on the device so it never does. That is most of the product lesson for AI on portable hardware: the feature that runs locally is a feature, the one that phones home is a dependency, and the places that need a handheld most are the places with the worst signal. Eko made the same split a pricing decision, with the acoustics on the device and the murmur AI behind a membership.

The model behind the product got a name

Corti moved the frontier tier behind its corti-s1 and corti-s1-instant model IDs from GLM 5.2 to GLM 5.3 on its EU infrastructure. US projects stay on OpenAI models hosted on Azure behind the same IDs, and the IDs, the API and the pricing are unchanged. Corti names the model under its product and says where it runs, which few clinical AI vendors do.

Owkin made Claude Sonnet 5.5 the default model in K Pro and replaced Claude Opus 4.6 with Opus 5.5 in its model selector, and its subprocessor list shows Anthropic models served through AWS in Ireland. Days later K Pro learned to stop and ask: when it needs a missing detail to continue an answer it opens a question panel with a suggested option preselected, and both the question and the choice stay visible in the conversation.

Syllable release 26.10.20 added a Tester tab that checks a saved Insights configuration against past sessions or uploaded files before it goes live, prompt caching so several tools can reuse shared conversation context, and a fallback to OpenAI when the primary web search provider fails. The release before it retired the Sheet tools and improved the warnings that flag configurations using models due to be retired.

Xaira Therapeutics introduced X-Scientist, its internal research platform, in which a lead agent plans, delegates and checks the work of worker agents that call Xaira’s own models through APIs and MCP servers. On the 82 task DrugDiscoveryBench it beat Codex by 7.4 points and Claude Code by 13.4 when each ran the same model, GPT-6-Astra or Claude Sonnet 5, and in a Perturb-seq test in liver cells its target picks hit at 7.1 percent against 6.9 for human curation and 0.9 at random, in 20 minutes rather than more than three hours. The benchmark runs used no domain tools, so they measure the harness rather than the biology models.

Two days later Xaira described X-Automate, which joins automation workcells, a camera guided robotic arm built from inexpensive commercial parts, and X-Scientist reading the scene through a vision language model. In its first deployment it kept an HPLC busy 97.2 percent of the time against 53.6, and the safety design is spelled out: a person entering the area pauses the arm and losing the camera feed stops it, both through hardwired controls that do not wait for the language model.

Generate:Biomedicines described an agent harness its team built with Claude Code to design CAR T cell signaling domains. Over three rounds it produced a 45 amino acid intracellular domain, under a third the length of the clinical standard, that raised CAR surface expression and interferon gamma release in primary T cells. The results are in a company hosted preprint, and the harness is internal.

XtalPi released LCMS2Yield on GitHub under the Apache 2.0 license, with model weights, inference code and datasets, so chemists can calculate reaction yields from liquid chromatography data without a pure reference sample of each product; across 178 reactions its average yield error was 4.74 percent. Valar Labs opened a four part series on CHAI, the vision foundation model behind its tests, in which shared models map tissue regions and classify every nucleus and a small formula per clinical question combines measurements a pathologist can check.

HOPPR released a beta of a vision language model that drafts the findings section of a musculoskeletal X-ray report from up to three views, trained on about 674,000 outpatient studies and reachable through its Foundry API. It is not cleared and not for diagnostic use, so it is a starting point for products rather than a tool radiologists can use.

Veeva’s Vault AI release opened the Query Agent to programmatic use through the Vault API and MCP, without a person in the loop, and let Vault AI agents call other agents’ actions as tools during orchestration. That moves Vault AI from a chat feature toward a platform, and programmatic queries nobody reviews need governance rules of their own. Eolas Medical’s app now shows which population and sources each Ask Eolas chat searched, and splits answers between the organization’s own guidance and external guidance.

Aignostics deprecated its Python SDK, which covers the Launchpad desktop app, the command line tool, the client library and the example notebooks. It gets only bug and security fixes until April 30, 2027, after which the repository is archived, and users are pointed to the REST API and the browser Console. A dated end of life with a named replacement is how a deprecation should read.

Our read

The Brief counted the vendors that name their model providers back in August and found very few. This period the count moved. Corti and Owkin told customers the model changed on a date, Xaira and Generate named the frontier models in their harnesses, and Syllable published which provider catches the fallback. The news inside the news is that the model behind a stable ID can change on a Tuesday. For product teams integrating any of these, a stable model ID is a promise about the interface, not the behavior. Ask for the model changelog, and plan to re run your evaluation set when it moves. It is the software equivalent of your barber changing without mentioning it: same chair, same price, different haircut.

Named hospitals, named numbers

Qventus published HonorHealth’s results from moving discharge rounds at its large hospitals to an asynchronous model built on QCard, a Qventus panel in Epic where hospitalists set the expected discharge date, disposition and barriers before rounds. Average length of stay fell from 5.41 to 4.74 days over 12 months, with 0.40 of a day attributed to the switch, and rounds shrank from an hour to 30 to 35 minutes. HonorHealth’s smaller hospitals still round in person.

LeanTaaS published a Rush University Medical Center case study for iQueue across its 38 main campus ORs: a 20.5 times return on investment, 8 percent more surgical cases, 6 percent higher prime time utilization and 63 percent fewer unused block minutes, without adding rooms. The figures are vendor reported and no measurement period is given.

Caregility published results from Rome Health, a 128 bed community hospital in Central New York running virtual nursing on its platform. From February through August, virtual nurses completed 1,437 admission and discharge workflows, an estimated 718.5 hours returned to bedside nurses at 30 minutes each, while the hospital’s Press Ganey score for discharge information rose from the 51st to the 98th percentile and discharge order to departure time fell by 20.5 minutes.

Cadence reported an interim result from Hackensack Meridian Health, where more than 6,000 patients are managed at home with connected devices and a Cadence clinical team working under the health system’s own protocols; after six months, the share of patients at the blood pressure goal rose 76 percent in relative terms, without a starting rate. Andor Health said the virtual specialist program it runs with Mass General Brigham now connects more than 20 community and rural hospitals in four states for telestroke and teleneurology, on pace for more than 16,000 consultations a year, a projected run rate with no baseline attached.

Sentara Health deployed a case management solution developed with Regard, extending Regard’s platform from physicians to hospital case managers, who can ask questions about a patient’s chart in plain language and see what changed in the last 24 hours before discharge rounds, where weekend caseloads can exceed 100 patients each. Sentara’s estimate of nearly 90,000 hours and almost $5 million saved assumes expansion across acute and ambulatory care, so it is a projection.

Our read

A customer’s number is the best asset a vendor has, and the index reads it on two things: whether it states the baseline and whether it states the period. HonorHealth did both and attributed only 0.40 of a 0.67 day drop to the workflow, which is the most credible sentence in the set. A multiple without a period, a relative lift without a baseline and a projected run rate each ask the reader to supply the missing half. For the product marketer writing the next case study, the restraint is the credibility.

Market notes

Azra AI completed its acquisition of Blackford Analysis, whose platform offers more than 130 imaging AI applications from over 45 partner vendors. All Azra customers now have the Blackford Platform under one contract and a single IT connection, and Blackford’s existing partners and customers carry on as before. It is Azra’s second acquisition in under six months, and Blackford customers should watch how the platform’s neutrality across vendors holds up under new ownership.

Distribution kept doing the integrating. AZmed’s cleared fracture and chest X-ray apps and Optellum’s Virtual Nodule Clinic both arrived on RapidAI’s enterprise platform, which RapidAI says more than 2,500 hospitals use, so a health system already on it can switch them on without a separate integration. Deciphex’s Diagnexia reporting service joined the Sectra Image Exchange Portal, so NHS trusts short of pathologists can send cases over a connection they already have. Techcyte linked Fusion to NovoPath, a lab system used by more than 300 labs, with NovoPath staying the system of record, and launched a bundle through Fisher Healthcare that combines Fusion AP software, scanners, five years of service and 30 days of cloud storage for a single fee per slide, a unit published without its figure.

Aignostics launched HETA Clinical, four pan cancer products that run on whole slide images in the pathology lab: quality control, tumor detection at scanning so likely tumor cases are read first, mitosis counting across the whole tumor area, and tumor cellularity, which measures tumor cell content and proposes a region for molecular testing. Aignostics says 10 to 20 percent of molecular tests fail for lack of tumor tissue, and all four are research use only. PictorLabs introduced a platform that takes a single H&E slide, generates virtual stains and combines them with clinical and molecular data into a tissue score, announced on LinkedIn rather than in a dated release. Aiforia released Create 8.0, its no code tool for building pathology models, with training, annotation and whole slide analysis in one workspace and centralized audit logging for the data behind each model.

Notable launched Flow Compass, an assistant inside Flow Studio that answers operators’ questions about their automated workflows, runs and costs, scopes and builds new flows, generates test cases before go live and traces failed runs to a cause. Only the health system’s own team can publish workflow changes, which keeps a person accountable for everything that reaches production. symplr launched Time and Scheduling, joining Smart Square’s predictive scheduling with its timekeeping, with an AI schedule balancing feature that recommends changes for managers to approve.

Counsel Health launched a GLP-1 weight management program in 44 states, where its physicians assess eligibility and prescribe while its medical AI checks in on progress and side effects between visits and hands off to a physician when needed; visits cost $29 each or $300 a year with a membership. Sword’s remote pelvic health rehabilitation can now be prescribed at no cost to patients by physicians in Portugal’s national health service, a rare national reimbursement for digital pelvic floor therapy. UpToDate Expert AI’s iOS app now offers a 14 day free trial, up from seven, and lists an in app personal subscription at $89.

Caris received FDA approval for a supplement that moves the bioinformatics pipeline behind MI Cancer Seek from its own computing cluster to Amazon Web Services, a rare public record of an approved test’s analysis moving into a commercial cloud, with no change control plan authorized. Tempus’s xR RNA sequencing assay received IVDR certification as a Class C device, opening a regulated path for European labs. Luma Health earned the HITRUST AI Security Certification, which assesses how a company uses, trains and tunes AI models, still rare among patient engagement vendors. Wysa narrowed what its privacy policy treats as outside personal data, so information about an identifiable person now counts whether or not it is public. Zus Health added a managed role for integration clients that act on behalf of users and removed that permission from Builder Admin, narrowing who can reach patient records in a clinician’s name.

Insilico posted a preprint on the models that score trials in inClinico, trained only on what was known when each trial was registered, reaching an AUROC of 0.828 for efficacy on trials started from 2021, and set up a forward benchmark of 2,040 ongoing trials to be checked as they read out. Brainomix’s e-Lung flagged progression before clinical diagnosis in 47 to 83 percent of progressive pulmonary fibrosis patients across three US sites, in Annals of the American Thoracic Society, and also met progression thresholds in 42 to 67 percent of patients judged clinically stable, so sites need a plan for those flags.

GeneDx added Report Guide to its Provider Portal, plain language guidance beside eligible exome and genome reports, without saying how the guidance is generated. Theator’s assistant can now assess suturing on a recorded case, returning seven needle handling checks with the clip to play, and states the checks are not a validated skill score. Lumea added Nonagen’s three Oncuria bladder cancer urine tests to its ordering workflow, with each test’s requisition requirements built into the order. Inspiren’s Yardi integration syncs resident census daily so fall and call alerts follow move ins and room changes, one way for now. Weave went live with CareCredit, so front office staff can see a patient’s available credit before the appointment and start a payment in the office or by text.

HealOS published a command line client for its API covering patients, notes, eligibility checks and prior authorization tracking, with a dry run on every command. xCures added Care Team endpoints that return the providers found in a patient’s records as structured data, and removed two checklist fields that integrators will need to stop reading. Upheal added reusable weekly availability with up to 10 locations, so a booking picks up the right address, place of service and billing modifier. Medsender let admins switch on AI document filing themselves, at half a transmission page per processed page, and added authenticator app sign in codes. Current Health labeled its 9.1 app a critical fix: blood pressure readings from Welch Allyn cuffs now reach the care team even if the cuff disconnects, and setup screens arrived in 15 more languages.

DrFirst opened its Peerson assistant to staff who are not prescribers, so practices should decide what those users may ask it. Imprivata’s AI alert summaries now carry a severity score from 0 to 10 and drop straight into investigation memos. Finite State’s binary scans detect five more embedded libraries in device firmware and give 2026 kernel vulnerabilities a reachability result. BenchSci put Patsnap’s patent and clinical status data inside EMET, for teams that already pay for Patsnap. DynaMed’s app plays videos inside Dyna AI answers, and OpenEvidence added the American Society of Addiction Medicine’s guidelines to the content behind its answers, without listing which documents.

What the three weeks say about the category

One hundred and twenty entries, and the theme is the second step. The note became a line item, so the scribes went for the order and the posted payment. The accuracy claim became a dashboard. The voice agent learned to act on an appointment and, in the same release, to stop talking. The plan got an agent to meet the provider’s. The slide started answering questions a genomic test used to own, and the model under the product got a name and a date.

Each second step raises the bar in a different place. When the scribe posts a payment, the review step has to be written down. When the vendor grades itself, the method has to be published. When the agent on the phone can change a booking, the hard stop has to live in code. When the model changes behind a stable ID, the changelog has to exist.

The category spent three weeks shipping the step after the one it was known for, and every one of those steps arrived with the same question: who reviews it, and where is that written.

Index Answer

Which healthcare AI vendors publish what their product costs?

Fewer than one in ten, and this period showed what the exceptions look like. Of the 673 vendors the AI Health Index has assessed on Commercial Transparency as of October 30, 2026, 57 publish enough for a buyer to establish a pricing basis before contacting sales: a unit, a tier, a per provider or per encounter rate, or a self serve trial with a price attached. Another 76 publish a partial basis, usually the unit without the figure or a starting price with no statement of what it includes.

The remaining 540, 485 at C and 55 at D, cannot be priced without a sales conversation. The index records contact us for pricing as what it is rather than as a failure, because opacity is a legitimate way to sell enterprise software. The D grade is reserved for a published pricing claim the available evidence contradicts, which is a different offense. Commercial Transparency is the second most closed of the fifteen axes the index applies.

This issue’s log carried more published prices than most. athenahealth published a price of nothing, by including ambient documentation in athenaOne for primary care. Eko published a list price of $249 for its CORE 300 stethoscope and the $119.99 a year membership its murmur AI needs. UpToDate published an $89 personal subscription for Expert AI, and Counsel Health published $29 a visit or $300 a year for its GLP-1 program. Techcyte published its unit, a single fee per slide, and not its figure, and the index grades a published unit above a published number, because the unit is the part that survives contact with a real contract.

The practical move for a buyer is the same on every axis where the field mostly publishes posture: ask for the artifact rather than the claim. Ask what the unit is and whether it scales the way your volume does, what is excluded from the quoted number, particularly implementation and interface fees, and what happens to the rate at renewal. A vendor graded A here has usually answered all three on its own website, which is most of why it grades A.

Full grades, the axis definition and what separates each band are on the Commercial Transparency page.