The AI Health Index Brief

The week in one line

Assort Health’s agents can now write appointments, referrals and diagnoses straight into the NextGen chart, Oracle put voice charting in front of inpatient nurses, and Abridge followed the visit all the way to the bill. Two chest radiograph tools were cleared in the same week by opposite routes, and two vendors signed up to be paid by Medicare on patient outcomes. The agent has moved from answering the phone to editing the record.

This issue covers two weeks, September 6 to 19, plus one entry dated September 4 that arrived after the last issue went out. That makes 19 entries across 18 vendors, 12 Verified at source and 7 Partially Verified. Funding rounds, valuations and awards are not logged, here or anywhere on this index.

The agent got write access

Assort Health expanded what its agents can do inside NextGen Enterprise. Using NextGen’s Platinum API tier, the agents can now read and write appointment slots, referrals, clinical notes, chart alerts, diagnoses, procedures and payment workflows directly in the record. Recorded Partially Verified from NextGen’s marketplace listing.

Read that list again. Appointment slots are what a phone agent has always touched. Diagnoses and procedures are not.

Write access is the line between an agent that answers the phone and one that finishes the job. It is also the line past which every action has to appear in the audit trail with a name on it, and an agent’s name is a new kind of name for that trail to carry.

Oracle opened its Clinical AI Agent to inpatient nurses in the United States. Nurses get voice driven chart navigation, acute nursing summaries and voice enabled discrete charting inside the Oracle Health Foundation EHR.

Ambient AI has so far been sold to physicians, while nurses do more of the documentation and have had less help. The word to notice is discrete. A messy note gets read and corrected. Structured data that lands in the wrong field gets computed on, and nobody reads a field.

Hello Patient acquired Converse Health, adding agents that handle the paperwork around a visit. The combined platform reads incoming faxes, matches or creates the patient record, files the document in the EHR, and then follows up with the patient by call or text. It integrates with ModMed, athenahealth and eClinicalWorks.

Most practice automation stops at the phone, and the fax behind the referral still lands on staff. This one files the fax. The question is what it does with a fax it cannot match to a patient, because a misfiled clinical document is a safety problem before it is an admin one.

One more entry belongs here, dated September 4. Canary Speech’s Canary Ambient for Physicians is generally available inside Microsoft Dragon Copilot, listed on the Microsoft Marketplace under the Dragon Copilot apps and agents programme.

Inside the Dragon workflow it analyses the acoustic and linguistic properties of the recorded encounter to surface screening signals for cognitive impairment, depression and anxiety, with no extra step during the visit.

The change is distribution, not the model. A health system already on Dragon Copilot can now buy voice biomarker screening through a channel its IT and legal teams have already approved, which is the step that has kept this category in pilots. Canary is careful to say the output is decision support for a clinician rather than a diagnosis, and that enablement still goes through Canary rather than a self service button.

Our read

Four entries, and the direction is the same in each: the agent is moving from the edge of the record to the inside of it. Reading the chart was a permissions question. Writing to it is a governance question, and the governance has to live in the record system’s audit trail, because that is the only place a reviewer will ever look.

Buyer question

For any agent with write access to the record, ask which write actions are switched on by default, how each one appears in the audit trail and under whose name, and what the agent does when it cannot match a document to a patient. An agent that files rather than flags in that last case has made a clinical decision nobody asked it to make.

From the exam room to the bill

Abridge launched pre bill review for clinical documentation integrity, coding and revenue cycle teams. It checks inpatient claims before submission, comparing the coded diagnoses and DRG against what the clinical record supports, including what was captured in the encounter, and surfaces discrepancies with the evidence.

Abridge does not change documentation, codes or claim status. The CDI team decides what to hold, correct or release.

That takes Abridge from the exam room into the revenue cycle, where CDI vendors have worked for years. The advantage it claims is the encounter itself: a coder can see what was actually said, not only what was written.

Maverick Medical AI announced a Clinical to Revenue Foundation Model behind its two products. CodeAgent engages physicians at the point of care to flag missing documentation, coding opportunities, medical necessity gaps and payer specific risks before a note is finalised.

mCoder then generates billing ready codes, with further agents checking payer policy before the claim goes out. Maverick reports 85 percent or better direct to bill performance in production. Recorded Partially Verified.

A vendor claiming a foundation model for coding is making a model disclosure claim, and the release gives a buyer nothing to check it against: no architecture, no training data, no evaluation set, no specialty breakdown.

Put the two side by side. Abridge audits the claim against the encounter after the note is written, and leaves the decision with a person. Maverick prompts the physician toward the claim while the note is still open. One of those a compliance team will welcome. The other it will want to read very carefully, because a system that prompts physicians for coding opportunities at the point of care sits close to the line compliance teams draw around upcoding.

Two weeks ago this Brief said clearance decides whether a product may be used and payment decides whether it gets used. This fortnight two vendors signed up for a payment model that pays on results.

AliveCor launched KardiaACCESS under the CMS ACCESS Model, the ten year Medicare initiative for technology supported chronic care. Eligible Original Medicare beneficiaries get clinician guided care and AliveCor’s connected devices for blood pressure, diabetes and heart health, at no out of pocket cost.

RadiantGraph launched Care Guidance Agents, voice agents that support patients with chronic conditions after a diagnosis, and says it too has been approved for ACCESS.

ACCESS pays for outcomes, not activity. Any vendor in it has to show that its agents move blood pressure, pain or depression scores, and it has to hand that data back to the provider who is accountable for the patient. For AliveCor it is also a change of identity, from a device maker to a care provider under Medicare.

Our read

The scribe wants the bill and the coder wants the note, and both are converging on the same fifteen minutes between the visit and the claim. Meanwhile Medicare has started writing contracts that pay on what the patient’s numbers did. Evidence used to be a marketing question in this category. Under ACCESS it is the invoice.

Buyer question

For a pre bill tool, ask how often a flag ends in a corrected claim rather than a physician query, since a query is work and a correction is money. For a point of care coding prompt, ask what happens when the physician disagrees with it. For any ACCESS vendor, ask which tracks it covers and what outcome data comes back to you, because that data is what the payment rests on.

Two chest radiographs, two routes to clearance

AZmed received its sixth FDA clearance, for AZchest to detect and localise pneumothorax and pleural effusion on chest radiographs. That brings its Rayvolve suite to eight cleared findings, and AZchest can now flag up to five findings on a single image. AZmed cites reading time reductions of up to 31 percent for readers outside thoracic radiology.

Three days later DeepHealth received 510(k) clearance for Chest XRay, a detection tool built on a foundation model that detects and localises abnormalities on chest radiographs. It is the next version of the technology DeepHealth acquired with Gleamer, it is commercially available in the United States now, and existing customers are eligible for the update. Recorded Partially Verified from trade press.

Same anatomy, same week, opposite routes. AZmed adds cleared findings one clearance at a time, and now has eight. DeepHealth cleared one model that is meant to cover many.

A foundation model clearance is a real change, because one model covering many findings is a different product from one algorithm per finding. It also opens a new kind of gap. A model that can see everything is only cleared to say some of it, and marketing for foundation models tends to run ahead of the cleared indications.

The AZmed number deserves its own sentence. Pneumothorax is a time critical miss, and the readers who gain most from a flag are the ones outside thoracic radiology, which is exactly the group AZmed measured. Specialist readers gain least, so a trial run only with specialists will understate the tool.

Our read

For two years the count of cleared findings has been the scoreboard in radiology AI. A foundation model clearance changes what the count measures. From here the artifact to ask for is the indications for use statement, because that is the only document that says what the cleared model is allowed to tell a radiologist.

Evidence, and who ran it

Knowtex reported results from a US Department of Veterans Affairs assessment of its ambient scribe at the Kansas City VA, covering 18 primary care providers. Most saved one to two hours of after hours work, and patient experience scores rose nearly three points to 95.8 percent.

In the same announcement Knowtex launched a Frontier AI Lab for Healthcare and KnowBench, an internal benchmark on which it scores its own platform at 97.99 percent. Recorded Partially Verified.

The VA assessment is the part that carries weight: an outside health system, working clinicians, patient scores. Eighteen providers at one site, reported by the vendor rather than published, is still too small to carry an evaluation on its own. KnowBench is Knowtex testing its own product, so treat that figure as marketing until someone else runs it.

Primaa and PathPresenter completed their joint validation of Cleo Skin, Primaa’s dermatopathology AI, running inside PathPresenter’s FDA cleared clinical viewer. The case mix covered melanoma, naevus, squamous and basal cell carcinoma and benign lesions, and the companies report faster and more consistent reporting when the AI fills structured fields inside the viewer.

Interim results published in March put the gain at a 25 percent reduction in analysis time and a 6 percent improvement in diagnostic accuracy. Recorded Partially Verified.

The finding worth keeping is that the gains came from the AI being inside the viewer rather than beside it. That is the deployment shape most labs still have not achieved. The evidence is vendor and partner reported, with the population and endpoints described at summary level, so treat it as a workflow study rather than a diagnostic accuracy trial until the full results appear.

Aignostics released PathoSearch in early access, a visual search engine for pathology. From a screenshot of a region of interest on a slide it retrieves morphologically similar, diagnosed reference cases, so a pathologist can build a differential in minutes.

Search runs on embeddings from Atlas 2, its pathology foundation model co developed with Mayo Clinic, against a reference set of more than 310,000 whole slide images from over 35,000 cases across 27 organs. The first integration is live inside Techcyte’s Fusion AP viewer. PathoSearch is free during early access, by waitlist, and is labelled research use only.

Rare and complex cases are where a pathologist reaches for a textbook or a paid consult, and similarity search against a diagnosed reference set is a plausible replacement for the first of those. The research use only label is the operative constraint. It cannot be part of a signed out diagnosis today, and the Fusion AP integration is the way to try it inside the workflow rather than a clinical deployment.

Our read

Two pathology entries in one fortnight, and both put the AI inside the viewer the pathologist already uses. That is the shape that produced the measurable gain at Primaa, and it is the shape Aignostics chose for its first integration. In pathology the viewer is the record, and the same rule applies: the tool that lives where the work happens is the one that gets used.

Market notes

Veracyte acquired Convergent Genomics, adding the UroAmp urine tumor DNA test to its urology portfolio, so it can now offer genomic testing from urine, tissue and blood across bladder cancer care. A urine test can help monitor bladder cancer between cystoscopies, which is the procedure patients most want to avoid.

Andros was acquired by Medallion, and its customers will move onto Medallion’s credentialing platform. Health plans that chose Andros for its NCQA certified verification now have a new platform underneath, and the question is whether the certification carries across without a gap.

Upheal logged two changes. Practices can now invite non clinical staff such as billers and practice managers, and those roles cannot see clinical notes, treatment plans, transcripts or AI generated clinical content.

Therapy notes carry some of the strictest privacy expectations in healthcare, and giving a biller full access just to reach the invoices was a real exposure. The second change is a reports section with dashboards and charts on request, a convenience layer rather than a clinical one.

Parlance put its IT help desk assistant on the ServiceNow Store with a new identity check for the voice channel. For a password reset, the assistant reads a one time code to the caller, who confirms it in Cisco Duo or another authenticator before anything happens. Recorded Partially Verified.

That step ties the approval to that caller and that request, which closes the social engineering path that help desk voice bots opened. Parlance cites estimates that password resets drive 20 to 50 percent of health system help desk calls, which makes this either a small feature or half the queue.

Sorcero’s Fall 2026 release adds purpose built AI for field medical, medical communications and launch teams, a conversational layer called Wizard, and real time analytics views. Medical affairs works under strict rules about what it can say and to whom, so the feature that matters is how the platform separates what field medical can see from what commercial teams can see.

What the fortnight says about the category

Nineteen entries, and the pattern is movement inward. The agent moved from the phone to the chart. The scribe moved from the note to the bill. The pathology model moved into the viewer. Voice biomarkers moved into the scribe a health system already runs. Even Medicare moved, from paying for activity to paying for what the patient’s numbers did.

Every one of those moves puts the AI closer to a system of record and further from a demo. That is good for buyers, because a tool inside the record can be measured and a tool beside it can only be admired. It also changes the questions that matter. Not what the model can see, but what it is allowed to write, under whose name, and who pays when it is right.

Index Answer

Which healthcare AI vendors can actually write into the medical record?

About one in five can show it. Of the 554 vendors the AI Health Index has assessed on EHR and Interoperability Depth as of September 13, 2026, 102 grade A. That grade means a named, two way integration with a major record system, verifiable in a marketplace listing or in integration documentation, with evidence it runs in production.

Another 253 grade B. They name the systems they connect to, but the connection reads without writing, writes in one direction only, or rests on a standard with named deployments behind it.

The remaining 199, 189 at C and 10 at D, claim integration through a standard or a middleware layer without naming a system, or publish no integration evidence at all. A connector described as available on request sits in that bottom band until one exists.

The phrase integrates with covers everything from a native embedded workflow to a nightly file drop, which is why the index grades the depth of the integration rather than the claim. This fortnight showed why the distinction matters. An agent that can write a diagnosis into the chart is doing something a nightly file drop never could, and it needs an audit trail a file drop never needed.

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 marketplace listing, the list of fields the integration reads and the list it writes, and the name of a site where it runs today. A vendor graded A here can usually send all three the same day.

Full grades, the axis definition and what separates each band are on the EHR and Interoperability Depth page.