The AI Health Index Brief

The week in one line

Ambience tied its fees to outcomes, Scribble made oversight a per visit setting, OpenEvidence started handing its answers to patients, and Corti published a sunset date. Four vendors changed the terms of the deal rather than the capability, and terms are the part a health system actually signs.

Theme 1: Two vendors rewrote the bill

Ambience Healthcare introduced the Ambience Standard, a commercial model in which partnership fees are tied to the achievement of measurable clinical, operational, and financial outcomes. Abridge extended Context Aware Clinical Intelligence enterprise wide, so partner systems deploy the agent to every clinician rather than to an early adopter cohort. Both are filed in the log as Pricing / packaging, and both are Verified.

Two pricing entries in one week would be unremarkable in most categories. Here it is close to an event. Of the 449 vendors graded on Commercial Transparency, 31 hold an A and 305 sit at C, and a C on that axis usually means the vendor publishes nothing at all about how it charges. A vendor that ties its fee to a number is not only disclosing a price. It is disclosing which number it believes it can move.

The two moves are different in kind. Ambience shifts risk toward itself. Abridge removes the pilot as a commercial stage, which also removes the checkpoint where a pilot either converts or quietly dies. Both trade a familiar shape for a cleaner one, and both put more weight on measurement than most health systems currently carry.

Our read

A price that moves with results is the strongest transparency signal a vendor can send, and it is also the hardest contract anyone will have to administer. The invoice now depends on a number, and in most health systems nobody owns that number today. Whoever builds the instrumentation writes the definition, and the party proposing outcome based pricing is usually the party best equipped to instrument it.

Buyer question

Which system of record computes the number that sets the fee, who is allowed to audit the query behind it, what the agreed baseline is and when it was measured, and what happens in the first month the two parties disagree. Get the dispute path in writing before the pilot, because after the pilot it is a renegotiation.

Theme 2: Oversight became a setting

Scribble by IKS Health launched Scribble Select, which lets a clinician choose per visit between AI only documentation, AI with human validation, and fully clinician reviewed notes, without leaving the platform. The entry is filed against Autonomy and Oversight Model, where 78 of 449 vendors hold an A and 275 sit at B.

That axis assumes a product has one posture. Scribble ships three and hands the dial to the user. It is a good feature and it is also a policy question wearing a user interface, because a choice offered at the moment of maximum time pressure resolves to whatever the default is. Whoever sets that default sets the documentation policy for the organization, and it will not be the committee that thinks it does.

Corti belongs in the same theme from the other direction. Its Guided Generation endpoints reached general availability, and the Documents Classic endpoint was deprecated with a final sunset of January 18, 2027. Guided generation is a constraint mechanism, structure imposed on a model that would otherwise write freely. The sunset date is the more interesting half. A vendor can only deprecate something people depend on, and publishing the date rather than the intention is the behavior of a company that expects to be integrated against. The entry is sourced to release notes, which roughly one vendor in ten in this index actually publishes.

Oracle Health expanded its Clinical AI Agent with automated professional fee coding, AI chart review, and clinician controlled dictation that transcribes into any text field. Recorded Partially Verified, trade press.

Our read

Autonomy is moving from a product attribute to a runtime setting, and grading will have to follow. A product offering three oversight modes does not have a posture, it has a range, and the honest grade describes the range plus the default. For a buyer the useful question is no longer how autonomous the product is, but how autonomous it will be on a Thursday afternoon in a clinic running 40 minutes behind.

Buyer question

Who sets the default mode, can it be set by encounter type rather than by individual preference, and is the selected mode written into the note metadata so an auditor can later reconstruct which encounters were machine only. If the mode is not recorded, the oversight tier is a marketing claim rather than a control.

Theme 3: The answer left the room

OpenEvidence launched Patient Take Homes, which lets a clinician optionally turn an AI generated answer into educational material for the patient. mdhub shipped three capabilities at once: an AI Care Coordinator named Laura that engages patients through the portal, a Chart Audit tool that reviews documentation against payer requirements, and custom analytics on clinical and financial metrics.

Last issue reported a systematic review in npj Digital Medicine finding that OpenEvidence tends to reinforce rather than alter existing clinical decisions. This week the same platform can turn that answer into a document the patient carries home. A confirmation that stays in a clinician’s head is a habit. A confirmation printed and handed over is a record, and it arrives with the institution’s weight behind it.

The failure mode changes with the audience. Clinician facing output passes through an expert who can notice when it is wrong. Patient facing output passes through someone with no way to check it and every reason to trust it, and it persists on a kitchen counter for months after the guidance behind it has moved.

Our read

Every governance model in healthcare AI assumes an expert reader sits between the model and the consequence. Patient facing output removes that reader and keeps the letterhead. This is not an argument against the feature, which is plainly useful. It is the point at which an optional convenience becomes patient communication, and patient communication already has rules, owners, and a review process that no ambient AI committee has been asked to apply.

Buyer question

Who approves a patient facing artifact before it leaves the visit, is that approval logged, does the material carry a date and its sources, and what is the recall path when guidance changes six months later. Ask whether the tool can generate patient material without an explicit clinician action, and if it can, leave that off until the review process exists.

Market notes

Owkin received European approval for two first in class AI diagnostics, one for breast cancer and one for colorectal, using multimodal patient data for biomarker screening and outcome prediction. Verified. This is the third consecutive issue whose headline regulatory entry is something other than a conventional United States clearance.

GeneDx enabled exome to genome reflex ordering inside Epic Aura and other point to point EHR integrations, automatically triggering a genome test when the exome returns negative or inconclusive. The workflow saving is obvious. The quieter fact is that the reflex rule is a clinical decision encoded once at the order level and then executed without anyone revisiting it.

Inspiren launched a daily census sync with August Health, pulling resident identifiers, names, and room assignments so move ins, transfers, and discharges are not entered twice. Second week running that assisted living received a native integration, which remains a small correction to a long standing ordering. Indica Labs shipped version 4.3 of HALO, HALO AI, and HALO Link, with a reworked image copy system and faster model testing. And 4baseCare launched TARGT Indiegene V2 at 2,206 genes, positioned for historically underrepresented populations, reported in trade press only and recorded Partially Verified.

Sword Health published an evidence report on its AI musculoskeletal care model claiming $3,177 in direct savings per member per year, a 3.2x return, and 54 percent of members with moderate to severe surgery intent remaining surgery free. It is logged Verified as a published document and Low impact as evidence, because the analysis is Sword’s own. Last issue’s tiers, one week later, unchanged.

The action item: fix the baseline before you sign the outcome

Outcome based pricing is the best offer a vendor made this week, and it is only as good as the number underneath it. If the vendor’s instrumentation defines the baseline, you have bought a subscription with extra paperwork.

Before the next contract that mentions results, pull 90 days of the exact metrics the contract names out of your own systems. Write down which system of record produced them, who ran the query, and what the number was on the day you looked. Then name the person who will run that same query every quarter, and the path you take when the two answers differ.

If you cannot produce that baseline in an afternoon, you cannot audit the invoice either, and the performance clause becomes a story both parties tell about the same unexamined dashboard.

Index Answer

Which healthcare AI vendors disclose their model providers and subprocessors?

Very few, and the shape of the shortfall matters more than its size. Of the 554 vendors the AI Health Index has assessed on Model Supply Chain Disclosure as of August 31, 2026, 22 name the model providers and subprocessors behind the product clearly enough to grade A. Another 201 publish effectively nothing on the question. The large middle, 90 at B and 241 at C, says something true but not usable: that the product is built on large language models, or that subprocessors are available on request.

The reason this question is hard to answer anywhere else is that it is a question about an absence. A vendor can describe its own posture, and every vendor does. Nobody has an incentive to publish where the whole field actually sits, so the buyer asking it gets a set of individually reasonable answers and no way to tell a strong disclosure from a confident one. Grading the field is the only way to see that 201 of 554 vendors answer it with silence.

The practical move is to stop treating the answer as a claim and start treating it as an artifact. Ask for the current subprocessor list as a document, with a named owner and a change notification commitment, rather than asking whether the vendor discloses subprocessors. The first request is auditable and the second is a yes from everyone. A vendor graded A here can usually produce that document the same day, which is most of why it grades A.

Full grades, the axis definition and what separates each band are on the Model Supply Chain Disclosure page.