Microsoft Dragon Copilot

Ambient speech to draft note generation is the entire artifact. The product does carry a large non generative inheritance in Dragon Medical One, the dictation engine that several Dragon Copilot capabilities still require a license for, so a buyer is acquiring two lineages under one brand. Graded A because the ambient documentation output that defines this category cannot exist without the model.

Draft and review by construction. The transparency white paper states plainly that decisions leveraging model output must not be based solely on system outputs and that the product is not a substitute for professional medical advice, diagnosis, treatment or judgment. Output arrives as a draft the clinician edits before transfer to the record. B rather than A because no acceptance rate, edit burden figure or confidence threshold is published, which is the specific disclosure this index rewards elsewhere in the category.

The most complete published documentation set located in this category. Microsoft publishes a formal AI transparency white paper alongside separate papers on privacy, security, data collection, sub processors, and intended and unsupported use cases.

The transparency paper names concrete limitations rather than gesturing at them: transcription errors from background noise and overlapping speech, omitted findings and hallucinated details in summarization especially in preview markets and languages, and a latency threshold above 250 ms degrading the real time dictation experience. Publishing a sub processor list is rare anywhere in this index.

The disclosure set is the most complete located in this category and a dedicated sub processor paper is part of it, which is the artifact this axis asks for and which almost nothing else in the index publishes. It sits alongside separate published papers on privacy, security, data collection, and intended and unsupported use cases, all delivered under the vendor's healthcare cloud, so a buyer can establish who is in the chain and what each party does rather than inferring it.

The chain is also structurally shorter than it looks. Hosting, platform and model provision sit inside one corporate group, so the fourth party question that dominates smaller vendors in this lane largely collapses into a question about a single named company with published commercial terms and a documented compliance programme. That is a real advantage of buying from an integrated supplier and it should be credited as one.

Two residuals worth raising in diligence rather than holding the grade down. The specific model family behind the summarisation step is not named in the located material, so the group boundary is disclosed while the component inside it is not, and vertical integration means the customer cannot choose or substitute any part of the chain. Confirm in contract which entities within the group process encounter content and where.

Independent evidence exists and the result is mixed, which is why this is B rather than A. In the first randomized trial of ambient scribes (Lukac et al, NEJM AI, published 26 November 2025, NCT06792890), 238 UCLA Health outpatient physicians across 14 specialties were randomized to DAX Copilot, Nabla or usual care over two months.

The Dragon Copilot arm showed a 1.7 percent reduction in time in note against control, which was not statistically significant, while the Nabla arm reached 9.5 percent and was the only arm to meet the primary endpoint. Both arms showed roughly 7 percent improvement on secondary burnout measures. The trial was funded by UCLA rather than by either vendor. Utilization was low in both arms: Dragon Copilot was used at 8,271 of 24,696 eligible visits. Graded on the existence, venue and independence of the evidence; this index does not re verify the underlying study.

Dedicated privacy and data collection white papers plus a published sub processor list, delivered under Microsoft Cloud for Healthcare. B rather than A because an explicit commitment that customer clinical audio and transcripts are not used to train foundation models was not confirmed for this specific product in this verification pass. That commitment is the decisive disclosure on this axis and is the first thing to re check on refresh.

Sold as HIPAA covered enterprise software under Microsoft's business associate agreement framework with healthcare specific contracting through Microsoft Cloud for Healthcare. Enterprise negotiated rather than a published self serve BAA, so the terms any given buyer receives are not publicly inspectable.

A product specific security white paper is published and platform compliance is published centrally through the Microsoft Service Trust Portal, which is a stronger structure than the marketing trust pages typical of this category. Held at B rather than A because named and dated product level attestations were not verified in this pass, and platform level assurance evidences the infrastructure the product runs on rather than the product itself. Refresh candidate.

Registered with the United Kingdom Medicines and Healthcare products Regulatory Agency as a Class 1 medical device, and available to physicians in the United States, Canada, the United Kingdom, Ireland, France, Germany, Austria, Belgium, the Netherlands and Switzerland.

That makes this record the clearest illustration in the index of a point the ambient scribe category keeps raising. The same software is a registered medical device in one market and an unregulated documentation tool in another. In the United States no device pathway attaches, and Microsoft states expressly that Dragon Copilot is not designed or intended as a substitute for professional medical advice, diagnosis, treatment or judgment, and that it does not warrant the product is sufficient for any medical purpose. In the United Kingdom the regulator treats ambient documentation as software as a medical device and the product carries a registration. Elsewhere in this index that contrast has to be drawn between two different vendors. Here it sits inside one product.

Read the registration accurately. Class 1 is the lowest risk classification, and under the current United Kingdom framework conformity at that class is declared by the manufacturer rather than assessed by a notified body. So this is a declaration Microsoft has made and registered, not a regulator review of the software. It is still materially more than the United States only peers graded C in this category carry, and it brings post market surveillance and change control obligations with it.

Also on record for the United Kingdom deployment: NHS England Digital Technology Assessment Criteria, ISO 27001 and DCB0129 clinical risk management. Nurse and radiologist availability is narrower than physician availability, so scope the market and the clinical role together rather than assuming the registration covers every workflow sold.

The strongest governance disclosure located in this category, and it comes from the incumbent rather than a challenger. The transparency white paper names automation bias, privacy of voice data and fairness as its responsible AI concerns, and states outright that models can underperform on dialects and accents not well represented in training data. That is the exact failure mode this index flags across speech products, named by the vendor rather than by a critic. Microsoft also discloses maturity honestly, marking nursing capability as private preview and generally available in the United States only.

The published transparency documentation does something almost nothing else in this lane does: it names concrete failure modes rather than gesturing at limitations in general. Transcription errors from background noise and overlapping speech, omitted findings and hallucinated details in summarisation with preview markets and languages called out as worse, and a stated latency threshold above 250 milliseconds beyond which the real time experience degrades.

That is a falsifiable published account of where the system breaks, and it is what gives a customer something to hold the vendor to when it breaks there. Admitting hallucinated details in a clinical summarisation product, in writing, is a more useful disclosure to a buyer than any accuracy percentage, because it tells the review process what to look for.

The contracting posture supports the grade as well: this is sold under enterprise agreements with a documented compliance programme and published commercial terms, so a health system negotiates against a defined counterparty rather than clicking through consumer style terms. Held below the top grade because no clinical performance guarantee, remediation commitment or output specific indemnity was located for this product.

Named limitations tell a buyer where the risk sits; they do not move it. The patient position is unchanged from the rest of the category, with no route to contest a note that misdescribes their encounter and no relationship with the vendor through which to seek one.

Two follow ups for diligence: whether the master terms carry any artificial intelligence specific indemnity that reaches clinical output rather than intellectual property claims, and what the vendor commits to when a documented failure mode occurs in production.

Embedded natively in Epic with Android and iOS parity and an Epic nursing app path for flowsheet capture, plus SMART on FHIR R4 launch for the web application. A partner Developer Kit lets an EHR vendor embed the whole workflow so recording is captured and the summary delivered without clinical content leaving the EHR, which TruBridge ships natively. Reach across leading EHRs is the widest in the category. Two caveats a buyer should price in: depth outside Epic varies considerably, and several capabilities require a separate Dragon Medical One license.

Cloud delivered on Microsoft infrastructure with web, desktop, mobile and EHR embedded access paths, and a published sub processor list which is more than most vendors here disclose. B rather than A because the specific data residency map for this product was not verified in this pass.

No published rate card. Microsoft does not disclose pricing publicly and buyers quote through Microsoft direct sales or a Cloud Solution Provider partner. Third party and reseller figures cluster between roughly 150 and 830 US dollars per provider per month depending on volume, contract term and existing Microsoft agreements, and a Physician Practice offer aimed at practices of 5 to 50 clinicians launched through the CSP channel on 1 February 2026. All such figures are partner or third party reported and are recorded in the pricing record as estimates.

The broadest role coverage in the category. Distinct purpose built experiences for physicians, nurses and radiologists rather than one clinician product stretched across roles, plus specialty specific note templates, multilingual multi party capture, and generated clinical letters covering consultation letters to the general practitioner, letters to patients and referral letters. Vendor stated caveats: nursing capability is private preview and United States only, and multilingual mode must be enabled by an organisation administrator before clinicians can access it.