The AI Health Index Brief

The week in one line

Four vendors published outcome numbers in eleven days, spanning a peer reviewed trial, a vendor impact report, and a customer deployment study. They are not the same kind of number, and treating them as interchangeable is how procurement decks become mythology.

Meanwhile, the ambient scribes started placing calls and submitting claims.

Theme 1: The proof came in three grades

Hinge Health published a nonrandomized controlled trial in the Journal of Comparative Effectiveness Research evaluating its digital exercise therapy in adults aged 65 and older. Participants at risk for falls had 37 percent fewer falls and 57 percent lower odds of an emergency room visit over three months than a control group, while physical functioning improved 6.9 points more on the SF-36 scale.

For a Medicare Advantage plan, falls and avoidable emergency visits are exactly the utilization that shows up in total medical spend and Star Ratings. This study will be quoted in every renewal conversation the company has this year, probably before anyone reaches slide four.

VUNO published in the American Journal of Neuroradiology, validating automated brain MRI volumetry for separating frontotemporal dementia from Alzheimer’s disease. The model was trained on 758 subjects, externally validated on 89, reached 91.4 percent accuracy internally, and reduced interpretation time for experienced radiologists.

Two overlapping neurodegenerative presentations, one structured measurement, and a journal reviewer standing between the claim and the reader. A surprisingly effective arrangement.

Heidi Health released an impact report covering five Ontario pilots: 333 active clinicians, more than 111,000 documented visits, an average of 13 minutes saved per visit, and a 66 percent reduction in documentation time at the Ottawa Institute of CBT.

Regard published deployment results from Monument Health. Over eight months across two hospitals, it reported a 56 percent improvement in average Elixhauser scores, a 52 percent increase in HCC capture, and a 44 percent increase in CC and MCC capture. At smaller facilities with no prior documentation improvement program, HCC capture increased 352 percent.

Four sets of percentages, three species of evidence: journal publications with external review, a vendor authored report, and a single customer deployment.

Our verification column already reflects the distinction. The two journal backed entries carry Verified. The two vendor authored reports carry Partially Verified. That is not a slight against either company. It is the method refusing to become emotionally attached to a bar chart.

Our read

All four numbers can be real and still be useful in very different ways. For a journal study, ask whether the studied population matches yours. For a vendor report, ask who chose the denominator and what happened at the pilots that did not make the report. For a customer deployment, ask whether the site was typical or the best case. A 352 percent improvement at facilities with no prior documentation program is a true number about a low baseline, and Regard deserves credit for publishing the context that lets a buyer see that. Context is where impressive numbers go to become informative.

Buyer question

For every outcome statistic in a vendor deck, first ask which of the three grades it belongs to, then ask the question that matches the grade. A percentage without its species is not yet evidence. It is decoration with a denominator.

Theme 2: The scribe stopped just writing

Tali AI shipped three releases in five days. Call from Tali lets clinicians place outbound patient calls from the platform, records both sides, and generates structured EHR ready notes from the audio. A second release generates and submits Ontario Health Insurance Plan claims directly through MC EDT, with one click note export into Med Access and default templates for phone visits. The third is a native integration with the Profile EHR by Intrahealth, allowing the scribe to launch inside the chart rather than beside it.

The scribe is no longer content to take notes. It has begun volunteering for administrative privileges.

The same week, Abridge acquired Altrina, bringing computer use agents into its platform to automate multistep workflows directly inside the EHR. The stated direction is ambient documentation expanding into active workflow execution, with fewer clicks between the note and everything the note triggers.

The change log already holds the standing precedent: Elation Health’s AI Fast Lane, where claims above a confidence threshold bypass the manual review queue and submit automatically. What looked like one vendor’s feature in March is beginning to look like the category’s direction in July.

Our read

The note was the product. It is becoming the input. When a documentation tool starts placing calls and submitting claims, the oversight math changes. A wrong draft note gets edited before signing. A wrong claim gets submitted, and the exposure reaches the billing office before anyone in the exam room notices. This is the moment when “ambient” stops describing where the software sits and starts obscuring what it can do. The autonomy and oversight axis on our vendor grids was built for exactly this shift.

Buyer question

For each documentation tool you run, list what it can now do rather than draft: place calls, assign codes, submit claims. Then name who reviews each action, and the threshold above which nobody does. “Human in the loop” is not an answer until someone names the human, the loop, and the point where both disappear.

Theme 3: What quietly breaks clinical AI

Waiv released two robustified digital pathology foundation models, Phaet and Mascaret, on Hugging Face for research use. Both were fine tuned with a model agnostic technique designed to reduce fragility across slide scanners, laboratory procedures, and tissue staining.

The failure being engineered away is a quiet one: a model performs beautifully at the institution that trained it, then loses accuracy at the next site’s scanner without the courtesy of announcing the decline.

PathAI shipped AISight Dx 2.21 with customizable case dashboards, expanded metadata support for bulk slide ingestion, and Amazon S3 Glacier Deep Archive support for moving older digitized slides into cold storage. Digital pathology’s dirty secret is the storage bill. Archive tiering is the vendor acknowledging it in the release notes, where all expensive realities eventually surface.

Ordr expanded its integrations, including one with TruAsset that maps FDA recalls, CVEs, and network anomalies into the clinical engineering and asset management systems where infusion pumps and imaging equipment are actually managed. Security context now lands beside the maintenance record instead of inside a separate console nobody in biomed opens unless something is already on fire.

Our read

None of this demos particularly well, and all of it is where deployed clinical AI actually degrades. Models drift across scanners and stains. Device fleets accumulate recalls and vulnerabilities. Slide archives grow into a storage line item with its own budget meeting. A vendor doing unglamorous work on a failure mode is usually a vendor that has met it in production. Reality rarely arrives wearing a conference badge.

Buyer question

Ask any imaging or pathology vendor what happens to accuracy when the scanner, stain, or site changes, and who pays to store the evidence afterward. The answers separate vendors who have deployed from vendors who have piloted, and vendors who have piloted from vendors who have prepared a very attractive PDF.

Market notes

Heartflow’s PCI Navigator entered commercial use, with The Valley Hospital becoming the first US institution to deploy it. The software gives interventional cardiologists a personalized 3D model of vessel sizing, plaque, and blood flow to plan angioplasty and stent placement before entering the catheterization lab. A vendor built on noninvasive diagnostics now sits inside procedure planning. Setting expansion is the quiet way a platform’s scope grows. First it informs the decision. Then it begins arranging the room.

Counsel Health launched Counsel Studio, a white label layer that embeds its AI enabled virtual care interface into partner applications through APIs, iFrames, or WebViews, with the partner’s own clinical protocols ingested. Care delivery becomes an embeddable component, and the brand on the front of the app may no longer be the company doing the triage. When the interface is yours and the AI is theirs, decide where clinical accountability sits before launch. Postlaunch is a poor time to discover that everyone assumed it belonged to someone else.

From the back catalog

The change log predates the Brief. Four standing entries a buyer should not miss while we are introducing ourselves:

• Guidelines: the ArteraAI Prostate Test entered the NCCN prostate cancer guidelines, becoming the first AI enabled risk stratification tool recommended there, with Level 1B evidence under Simon Criteria and a Category 2A recommendation. Guideline inclusion means a panel with no commercial stake reviewed the evidence, which is a different class of validation from anything a vendor can publish about itself, however tasteful the typography.
• Reimbursement: Bunkerhill Health secured FDA clearance for coronary and aortic valve calcium scoring on contrast enhanced, routine nongated chest CT, alongside a national CMS billing code under the Hospital Outpatient Prospective Payment System. Most healthcare AI has no reimbursement pathway at all; a national code moves a tool from cost center argument to billable service line, which is the dialect finance prefers.
• Triage at scale: Aidoc received FDA clearance for a 14 indication abdomen CT triage workflow powered by one foundation model, reporting roughly an order of magnitude reduction in false alerts compared with single condition tools. Triage AI lives or dies on the false alert rate. Radiologists eventually stop listening to software that behaves like a smoke alarm near a toaster.
• Throughput: Subtle Medical cleared SubtleHD for PET, supporting up to 75 percent faster PET imaging on existing scanners across all FDA approved radiotracers. Scanner capacity is a hard constraint. Faster acquisition adds throughput without new capital equipment, a sentence hospital operators tend to read twice.

The action item

If an ambient scribe evaluation is on your calendar this quarter, price the tier you would actually deploy.

Two standing entries in the log frame the issue. Heidi Health restructured its pricing upmarket in February: the main paid tier became Clinician at 150 dollars per user per month billed annually, with push to chart EHR integration and an organization wide Business Associate Agreement reported as available only from the Practice tier upward. Nabla appears to have withdrawn public pricing entirely in favor of a sales conversation, recorded in the log as Partially Verified. Nothing says transparent software market quite like a button labeled “Contact us.”

The self serve evaluation path this category was known for is narrowing. Run the pilot on the tier you would put into production, and get Business Associate Agreement inclusion and the ability to disable session data training in writing before the pilot begins, not after it succeeds. A free tier that cannot legally touch protected health information is a demo, not an evaluation.