The AI Health Index Brief

The week in one line

CMS approved a New Technology Add-on Payment for Bayesian Health’s FDA cleared sepsis monitor, Xsolis shipped an agent that argues medical necessity back at the payer, and Clearwave put the booking button inside Google Search. Clearance decides whether a product may be used. Payment decides whether it gets used.

This issue covers August 30 to September 5: 10 entries across 10 vendors, seven Verified at source and three Partially Verified from trade press, with one acquisition logged from both sides. A light week by this log’s standards, and an unusually shaped one, since it contains no ambient scribe release at all after a month in which the scribes led every issue. Funding rounds, valuations, and awards are not logged, here or anywhere on this index.

Theme 1: Someone finally said who pays

CMS approved a New Technology Add-on Payment for Bayesian Health’s FDA cleared continuous AI sepsis monitor, creating Medicare payment support for health systems running the platform. This is the rarest event in this index. A clearance establishes that a product may be used and says nothing at all about who pays for it, and the missing payment mechanism is the reason a great deal of cleared clinical AI stalls exactly one budget cycle after a successful pilot.

An NTAP moves the conversation out of the innovation budget and into the operating one, which is a different committee with a much longer memory.

Xsolis came at the same problem from the other end with GenAI Peer-to-Peer Clinical Synopsis, an add on inside Dragonfly Advise. It reads longitudinal EMR data including labs, vitals and nursing notes and produces a structured synopsis of severity indicators and medical necessity rationale ahead of a peer to peer denial review.

The vendor puts the manual version of that task at 15 to 20 minutes of chart reconstruction inside a 24 to 72 hour regulatory window, which is specific enough to check against a physician advisor’s own calendar. Recorded Partially Verified from trade press.

The thing to interrogate in evaluation is what the synopsis leaves out, because a medical necessity argument assembled from a summary is only as good as the exclusions nobody sees.

Clearwave connected its Scheduling Agent to Reserve with Google, putting a Book Online button on a practice’s Google Business Profile so a patient can book from Search or Maps without being handed off to a separate site. Also Partially Verified.

It is a small change with an outcome that can actually be counted, which makes the evaluation question easy: ask for booking completion rates before and after rather than for a description of the integration.

Three entries, three points on the same path. Payment authorised for the treatment, payment defended after the fact, and the patient who generates the payment arriving in the first place. Not one of them is a clinical capability. All three decide whether a clinical capability is still switched on in two years.

Our read

This category talks about two gates, clearance and evidence, and markets against both. The NTAP is a reminder there is a third that nobody puts on a slide. For cleared clinical AI, the variable that best predicts whether the thing is still running after the pilot is not how it performed in validation. It is which budget line pays for it in year two, and whether that line existed before the vendor arrived.

Buyer question

For any clinical AI you are piloting, name the budget it moves to when the pilot ends and the code, payment mechanism or measured avoided cost that funds it. If the answer for year two is still the innovation budget, you have bought a pilot rather than a product, and the renewal conversation will be had by someone who was not in the room for this one.

Theme 2: The report stopped being a finding

Lumea integrated TreatmentGPS’s Alpine Path into its digital pathology platform, so for prostate and bladder biopsies the report itself now carries NCCN based risk stratification, ancillary testing recommendations and a plain language summary written for the patient. Trade press, Partially Verified.

Two things changed in one release. The pathology report moved from a finding to a recommendation, and it acquired a second audience with a completely different tolerance for uncertainty.

A pathologist who has spent a career being exact about what a slide shows is now publishing a sentence about what should happen next, under the same signature, and that is a governance conversation to have before the feature is switched on rather than after.

BrainCheck added BrainCheck Screen, a pre screening tool that decides which patients need fuller cognitive testing, reporting 92 percent sensitivity and 74 percent specificity on a 99 patient dataset. Two operating characteristics and a stated denominator is more than most vendors in this index publish and it is enough to reason with.

The shape is right for a pre screen, where the job is to avoid missing people and to accept that roughly a quarter of those flagged will test normal. The number to ask for next is performance in an unselected primary care population, because that is where the tool will actually run.

bioAffinity Technologies extended CyPath Lung from early detection into surveillance of lung cancer survivors, positioning the noninvasive sputum test alongside CT for monitoring recurrence and for evaluating new or changing nodules after curative intent treatment. Recorded from an SEC filing.

Surveillance is a larger and far more repetitive population than initial screening, so this widens where the test sits in the pathway, and the specific job it takes on is risk stratification at the moment imaging is equivocal, which is when a pulmonologist has to choose between watchful waiting and an invasive procedure on incomplete information.

Three diagnostics, one direction of travel. Each is being asked to carry a decision rather than an observation, and each moves further from the safety of describing what was seen.

Our read

A finding is judged on accuracy. A recommendation is judged on consequence. Vendors are making that move faster than they are upgrading the evidence that supports it, and BrainCheck is this week’s example of doing it in the right order, because it published the operating characteristics and the denominator alongside the claim. The pattern to watch is a product that adds a recommendation layer while its evidence page stays exactly where it was.

Buyer question

For any diagnostic that now recommends rather than reports: what does the clinician see before signing, is the recommendation attributed to the vendor or to the signing physician, what does the patient facing summary say when the case is genuinely uncertain, and who reviews that language.

Theme 3: The bottleneck was next door

LeanTaaS acquired Aidin and will fold post acute placement into iQueue for Inpatient Flow, putting Aidin’s care transition workflow downstream of LeanTaaS’s discharge readiness intelligence rather than alongside it.

LeanTaaS has sold capacity optimisation inside the hospital walls for a decade and kept arriving at the same wall: a bed that cannot be freed because a placement has not been arranged. Buying the placement workflow rather than integrating with one is a statement about where the remaining length of stay savings live.

Anyone currently evaluating Aidin is now evaluating LeanTaaS, with the contract, roadmap and support relationship that implies.

IQVIA launched Predictive Clinical Development, covering site selection and study design, an automated start up path it calls Push Button Start-Up, and real time data cleaning aimed at shortening the gap to database lock.

Each of the three targets a different well known delay, and the one worth interrogating is the last, because database lock timelines are usually held up by query resolution rather than by query generation.

Generating queries faster against the same resolution capacity moves the queue, not the date. Sponsors should ask which of the three has a measured effect on a real study rather than a modelled one.

Two vendors, two ways of reaching the same conclusion: the delay they are paid to fix is not the delay that binds. One bought the adjacent step and one built it.

Our read

Optimisation inside a single department eventually runs out of room, and the next gain requires owning the handoff. Expect the same move in scheduling, prior authorisation and discharge, where the constraint has always sat between two organisations rather than inside one. For buyers this is mostly good news and it comes with a consolidation cost, since the vendor that owns both sides of a handoff is much harder to replace than the two that owned one each.

Buyer question

When a vendor promises to shorten a cycle time, ask which step in that chain it controls and which it only observes, and ask what the measured improvement was at the step it does not control.

Market notes

Insilico Medicine released a set of small language models trained as scientific specialists through its MMAI Gym for Science framework, covering chemical synthesis, ADMET prediction and potency prediction across GPCR and kinase panels, including a single step retrosynthesis model built on Liquid AI’s 2.6 billion parameter architecture.

Naming the base architecture and the parameter count is rare enough in this index that the entry is graded on model transparency rather than on capability, which says more about the field than about Insilico.

For discovery teams the practical read is that these are narrow specialists rather than a general assistant, which is the right shape for work where a confident wrong answer is expensive to follow.

What the week says about the category

Nine events, and not one of them is a model getting better. A payment pathway, a denial argument, a booking button, a report that recommends, a test that moves into surveillance, an acquisition that buys the next step in the chain, a trial platform aimed at start up delay, and a drug discovery vendor publishing its parameter count.

The category has stopped competing on the model and started competing on the part of the system that decides whether the model’s output is acted on and paid for. That is harder to demonstrate in a sales meeting and considerably easier to grade, which suits an index and, more importantly, suits a buyer.

Index Answer

Which healthcare AI vendors publish evidence that their product works?

Fewer than one in five, and the failure is not where buyers usually look for it. Of the 554 vendors the AI Health Index has assessed on Clinical and Operational Evidence as of August 31, 2026, 98 publish peer reviewed or independently evaluated performance, prospective and multi site where the claim requires it, with the method available to read. Another 188 sit one band below on named deployments with dated outcome figures, or on published research short of independent validation.

The remaining 268, 250 at C and 18 at D, publish something a reader cannot check.

The C band is the most common answer in healthcare AI and it deserves describing precisely, because it does not present as a gap. It is a customer list, or a percentage with no method, no denominator and no reference standard attached. The index records scale of use there and deliberately declines to treat it as evidence of benefit, since a hundred hospitals running something is a fact about a sales team.

A buyer reading that band sees numbers everywhere and has nothing to argue with, which is a more expensive position than reading a page that says nothing at all.

This week’s log supplied the counterexample, and it is instructive because the study is small. BrainCheck published 92 percent sensitivity and 74 percent specificity for its cognitive pre screen and stated the dataset: 99 patients. Publishing the denominator is what converts a number into something a buyer can interrogate, and the interrogation is obvious once the number is there, which is how the tool performs in an unselected primary care population rather than a validation cohort.

A vendor that hides the denominator is not protecting the claim. It is preventing the only conversation in which the claim could earn anything.

The practical move is the one that works on every axis where the field mostly publishes posture: ask for the artifact rather than the claim. Ask for the evaluation report, the population it was run in, the reference standard, the operating threshold, and who performed the analysis.

A vendor graded A or B here can usually send those within a day, because they exist. The distance between B and C is rarely the quality of the product and is very often whether anyone was asked to write the method down.

Full grades, the axis definition and what separates each band are on the Clinical and Operational Evidence page.