Commure

Ambient AI documentation is the product; output feeds autonomous coding and claims automation.

Human oversight is the product structure itself: Go (AI only), Assist (AI plus specialist review), Live (synchronous human documentation) as explicit tiers.

No model is named, and there is no model class or version, foundation model provider, architecture description, model card, update policy or published accuracy, error or benchmark figure. The technology is described only in category terms, as language models that transcribe and code, and in marketing terms as true ambient scribe technology rather than dictation, a distinction that describes the interaction model rather than the system behind it.

The transparency problem here has a structural dimension its competitors do not share, and it affects what a buyer is actually purchasing. The ambient capability reaching a customer may originate from Augmedix, from the pre acquisition Commure and Athelas line, or from a converged successor, and public material describes multiple concurrent product names including Commure Ambient AI, Commure Scribe, Augmedix Go, Augmedix Assist and Augmedix Live.

A group assembled through five acquisitions in three years presents a harder provenance question than a single product line: which engine is under the product on offer, and is it the one the published results and the KLAS score were earned on. Two capability figures are published and imply internal measurement that is not shared: multilingual conversational support across more than 60 languages, and specialty specific templates. Worth asking which underlying system the quoted product runs on, whether its evidence transfers, and for transcription accuracy by language.

For most vendors this axis asks whose model sits inside the product. For a group assembled through five acquisitions in three years the question starts one level earlier and is still unanswered: which of the company's own engines is inside the product on offer. Public material describes multiple concurrent ambient product names spanning the acquired lines and the pre acquisition ones, and nothing establishes which system produces the note a given customer would receive.

That is a supply chain disclosure failure in the most literal sense, because a buyer cannot identify the first link, let alone the ones behind it. Beyond that, nothing is named at all. No foundation model provider, model class or version, no hosting or cloud arrangement, and no sub processor list was located in two passes. The technology is described in category terms as language models that transcribe and code.

One genuine credit belongs here and it is about the boundary rather than the chain: the trust centre commits to no selling, renting or sharing of protected health information or customer data without permission, states purpose limitation, and affirms customer data ownership with export and deletion rights.

That bounds where content may go without saying who currently holds it, and the purpose limitation language stops short of stating whether protected health information is used to train models. Ask which engine the quoted product runs on, then ask the chain questions about that engine specifically.

The strongest item is third party but measures satisfaction rather than outcome, and that distinction shapes the assessment. A 2025 KLAS First Look Report scored Commure Ambient AI at 93.3, with 100 percent of respondents saying they would buy it again, based on interviews with a selected sample of customers, and credited reduced documentation burden, improved accuracy, better collection rates, lower denial rates and reduced burnout.

KLAS is a genuine independent research firm and a First Look is real third party scrutiny, but it is a customer perception study on a vendor selected sample rather than a controlled evaluation. The company also holds a Vizient contract for the Augmedix suite, awarded January 2025, which is a procurement credential from an organisation serving most United States acute care providers and academic medical centres.

Everything else located is vendor reported and undenominated: two hours of documentation time saved per physician per day, documentation time reduced by more than 80 percent, and billions of dollars in productivity savings. No peer reviewed study of the product was located.

One candour point runs against the company's own interest and deserves noting: Commure's published pricing analysis states that the only published randomised controlled trial in this category found modest and tool variable effects. Augmedix had a research presence before acquisition, so this assessment could change on a later review.

Trust center commits to no selling, renting, or sharing of PHI or customer data without permission, purpose limitation, customer data ownership, export in standard formats (FHIR, CSV, JSON), and deletion rights. Held back from A because the purpose limitation language (data processed to deliver, improve, and support products) does not explicitly state whether PHI is used to train models, the central AI era stewardship question.

Trust center states all products and workflows align with HIPAA and HITECH Act standards. A standard Business Associate Agreement is published among the legal documents on commure.com. Held back from A because HIPAA alignment is asserted rather than evidenced by a named healthcare certification, and BAA execution terms are not detailed publicly.

Public trust center at commure.com/trust-center with concrete controls: AES-256 at rest, TLS 1.2+ with perfect forward secrecy, automated key rotation, MFA, RBAC, SSO, automated vulnerability scanning, annual third party penetration tests, annual subprocessor security reviews.

Held back from A because the audit claim is ambiguously worded as audited by an independent third party (in progress / certified) with no named certification or date, and the SOC 2 Type 2 reference attaches to the data centers rather than to Commure itself.

No FDA clearance, none claimed and none required for ambient documentation. But this record sits at the extreme end of the scope expansion pattern seen across this category, and it differs in kind rather than degree. Commure is not a scribe company that added adjacent features; it is a healthcare AI group in which the scribe is one acquired component.

The assembled suite spans ambient documentation through the acquisition of Augmedix in 2024, revenue cycle management through the 2023 Athelas merger, patient navigation through Memora Health, an electronic health record extension platform with clinical and billing workflows, and a staff duress system.

The sharpest statement is in the company's own acquisition announcement, and the verb matters: language models that transcribe appointments, autonomously code them, and support back office operations for billing teams. Autonomous coding is a materially stronger claim than the coding suggestions offered elsewhere in this category, because a suggestion presumes a human selects it and autonomy presumes one does not.

Reported outcomes reinforce the direction, with customers crediting improved collection rates and decreased denial rates, which are revenue cycle results attributed to a documentation product. A system that drafts the note, codes it autonomously and operates the billing workflow has closed the loop from clinical encounter to claim. Worth establishing whether autonomous is literal or marketing, and where human review sits in the coding path specifically.

No AI governance artefact was located: no responsible AI statement, no bias or fairness position, no demographic, language or speaker group performance analysis, no error taxonomy, no model monitoring description and no published evaluation output. Most vendors in this category publish nothing on how their models behave across different people, so this is the norm rather than an outlier, but two things make the absence more pointed on this record. The first is scale.

More than 130 health systems, a stated reach across 250,000 providers, a co development partnership with HCA Healthcare covering emergency departments and hospitalist operations, and a Vizient contract reaching most United States acute care providers. A governance gap at that footprint touches more patients than any competitor's. The second is the multilingual claim.

Support across more than 60 languages is marketed as a differentiator, and no performance data by language, accent or dialect accompanies it. Speech recognition accuracy varies measurably across exactly those dimensions, and the affected populations overlap with those already underserved. A vendor marketing language breadth as a reason to buy has both the strongest obligation and the best data to show that breadth is matched by performance. The autonomous coding claim raises the stake further, since coding errors distribute unevenly across payer mix and patient population.

Two passes located no accuracy or error figure, no published limitations, no warranty, indemnity or remediation commitment. The distinctive problem here is not the absence of a number but the instability of what any published evidence refers to, and it is a representation problem a buyer should take seriously.

Performance results and third party recognition attached to this group were earned by particular products at particular times, and the group now spans several acquired lines with overlapping names for the ambient capability. A claim earned by one engine is not a representation about a successor engine unless the vendor says so explicitly, and where the product names overlap a buyer may not notice that a substitution has occurred at all.

The practical effect is that the evidence a purchaser relies on and the system they receive may not be the same thing, which is a harder position to contract around than a simple lack of published data. The remedy is procedural and belongs in diligence.

Establish in writing which underlying system the quoted product runs on, require that any performance evidence cited during the sale be attributed to that specific system, and ask what the vendor commits to if the engine behind the product changes during the term.

Augmedix Go documented across more than 50 EHRs including Epic, Oracle Health, and MEDITECH, ambulatory and acute.

US data residency stated: US based infrastructure in SOC 2 Type 2 compliant data centers. Cloud delivered platform integrated with 60+ EHRs. Held back from A because no dedicated tenancy, VPC, or on premises options are published and residency detail stops at country level.

Commure Scribe is sold on custom pricing for medium and large group practices, with no published rate for the tier a health system actually buys. Third party review material reports an entry tier in the region of 29 to 59 dollars per user per month, but those figures come from competing scribe vendors and are not treated as vendor disclosure here. What makes this record distinctive is an uncomfortable asymmetry worth stating plainly.

Commure publishes a detailed pricing analysis of its own category in which it names competitors' specific per user rates and lists its own as custom. That analysis quotes named monthly figures for rival products, sets out tier bands across the market, cites federal data on after hours documentation, and even acknowledges that the only published randomised controlled trial in the category found modest and tool variable effects.

It is genuinely useful buyer education, and the candour about the evidence base deserves credit. But a vendor that will publish every price in the market except its own has made the one omission that matters to the reader of that page. The feature description is more forthcoming than the price: the tier is stated to include write back electronic health record integration, custom AI workflows, live onboarding and return on investment analytics. That last item is the asymmetry in its purest form, since the product ships a tool for calculating return on an investment whose cost is not disclosed.

Dozens of specialties documented across ambulatory and acute settings.