Microsoft Dragon Copilot

Speech recognition and a large language model turn an ambient conversation into a draft clinical note, and that ambient output depends on the model.

The generative layer reaches past the note. The transparency white paper lists encounter summaries, evidence summaries that tie diagnoses in the assessment and plan to the transcript and draft note, patient friendly after visit summaries, starter referral letters, and a freeform prompt interface for creating, editing or summarizing content.

The product also carries a large inheritance from Dragon Medical One, the dictation engine Microsoft folded in. Dictation runs on speech recognition rather than generation, so a customer acquires two lineages under one brand, and Microsoft's licensing guidance prices them together under a Physician Per User license covering front end speech, ambient and AI features.

Every output is a draft for a clinician. The transparency white paper says Dragon Copilot is designed to augment, never replace, clinical expertise, and that all generated content must be reviewed and confirmed by the clinician before it becomes part of the medical record. Clinicians remain responsible for the final record, and the product is not a substitute for professional medical advice, diagnosis, treatment or judgment.

The use case paper draws the boundary from the other side. Use as an autonomous clinical decision making tool is listed as unsupported, as is recording or generating notes for multiple patients at once. Dragon Copilot is not a system of record, so the customer carries responsibility for the reviews, approvals and filings that move content into the EHR.

Automation bias, meaning over reliance on generated text, is named among the responsible AI concerns, and the main control against it is the clinician review the product requires. Model calls run inside guardrails that ground each prompt in the note data and check the reply against responsible AI policies.

No acceptance rate, edit burden figure or confidence threshold is published.

Dragon Copilot takes in dictation, ambient conversation, typed text and natural language prompts, and returns transcriptions, draft clinical notes and generated summaries or letters. Before a prompt reaches the large language model, its context is grounded in the specific note data, with external sources and data from other sessions excluded; responsible AI checks run again on the model's reply, and prompt injection checks keep system prompts and guardrail logic out of the output. Microsoft names the concrete limits: transcription errors from background noise and overlapping speech, omitted findings and hallucinated details in summarization, particularly in preview markets and languages, and latency above 250 milliseconds degrading real time dictation.

The stack is Microsoft's own and runs on its healthcare cloud, with the Azure OpenAI API named in its statements about output. The models are trained solely on anonymized data, including selected customer audio, transcriptions, clinical documentation and clinician corrections, and Microsoft acknowledges they can underperform on dialects and accents not well represented in that data.

The model family, version and update practice behind the summarization step are not published.

The supply chain appears in a dedicated subprocessor paper, published alongside separate papers on privacy, security, data collection, and intended and unsupported use cases, all delivered under Microsoft's healthcare cloud. The privacy white paper says Microsoft uses subprocessors to process customer data on its behalf, as set out in its data protection addendum, and discloses them in the Health and Life Sciences Subprocessor list.

Most of the chain sits inside one company. Hosting, platform and model serving are Microsoft's: the transparency white paper names the Azure OpenAI API alongside Dragon Copilot and uses Azure OpenAI Service classifiers as its content filter. The question of which fourth parties touch encounter content largely becomes a question about a single named company with published commercial terms and a documented compliance program.

People are part of the chain as well. In strictly limited scenarios, the privacy white paper says, select Microsoft employees and subprocessors may review customer data under tightly managed access controls, and human annotators help build the anonymized data sets used for model improvement.

Vertical integration has a cost. A customer cannot choose or substitute any part of the chain. The specific model family or version behind summarization is not named, so the provider is known while the component is not.

The ambient engine behind Dragon Copilot has been tested in an independently funded randomized trial, and its effect on documentation time was not statistically significant.

The trial, Lukac and colleagues in NEJM AI, published 26 November 2025 under registration NCT06792890, randomized 238 UCLA Health outpatient physicians across 14 specialties to DAX Copilot, Nabla or usual care over two months, from November 2024 to January 2025. DAX Copilot is the ambient product Microsoft has since folded into Dragon Copilot. The lead author describes the study as the first randomized trial to rigorously evaluate whether AI scribes deliver on their promise.

For DAX Copilot users, the estimated change in time in note against the control group was a 1.7 percent reduction, with a 95 percent confidence interval running from a 9.4 percent reduction to a 5.9 percent increase (P=0.66). The paper reports this as no significant change. The Nabla arm reached a 9.5 percent reduction and was the only arm to meet the primary endpoint. Physicians in both scribe arms showed roughly 7 percent improvement in burnout scores against control, and physicians reported that the generated notes occasionally contained clinically significant inaccuracies.

Utilization shapes what the result means. DAX Copilot was used in 33.5 percent of the 24,696 visits in its arm, about a third, so the trial measured adoption as well as effect. Funding came from the UCLA Department of Medicine, the UCLA Faculty Practice Group and the National Institutes of Health, not from either vendor.

Microsoft has not said what it takes from the UCLA result or what has changed in the product since.

Encounter audio streams to the cloud, including the voices of the patient, the care team and anyone else taking part, and Microsoft's privacy white paper sets out what happens to it next.

Selected customer data, such as audio recordings, transcriptions, clinical documentation and clinician corrections, moves to a dedicated research environment where it is anonymized within 90 days of capture. Microsoft says the process meets the HIPAA deidentification standard, the GDPR standard for anonymization and Canadian guidance under PIPEDA and Quebec Law 25. The resulting data is used to improve Dragon Copilot's models, which Microsoft says are trained solely on anonymized data. Human annotators help build those data sets, and in strictly limited scenarios select Microsoft employees and subprocessors may review customer data under tightly managed access.

Patients have a route out of the operational data. If a patient revokes consent, the customer can file a support ticket to purge any data linked to that patient from the operational and research environments, though anonymized data created before the revocation is not purged. Retention and deletion otherwise follow Microsoft's data protection addendum, and consent to recording follows the customer's own policy.

All of this appears in dedicated privacy and data collection white papers and a published subprocessor list, delivered under Microsoft Cloud for Healthcare. No customer opt out from the use of its data for model improvement is described.

Dragon Copilot is sold as HIPAA covered enterprise software under Microsoft's business associate agreement framework, with healthcare specific contracting through Microsoft Cloud for Healthcare. The privacy white paper lists HIPAA among the laws under which Microsoft commits to protect personal information, alongside the EU and UK GDPR and Canada's PIPEDA.

The framework carries through to data handling. Customer data is retained, deleted and purged as set out in Microsoft's data protection addendum, with purging done through Microsoft Purview Data Lifecycle Management. Data drawn off for model improvement is anonymized by a process Microsoft says meets the HIPAA deidentification standard.

The agreement itself is negotiated within enterprise contracts rather than offered as a published self serve BAA, so the terms a given health system receives cannot be inspected publicly.

Dragon Copilot has its own security white paper, and compliance for the platform beneath it runs through the Microsoft Service Trust Portal, which opens without a request and serves current audit reports.

Those reports cover Azure and the wider Microsoft cloud rather than the product. No dated audit report scoped to Dragon Copilot itself, naming ambient capture, the summarization pipeline or the clinical information extension, is published.

In Great Britain, Dragon Copilot for physicians is UKCA marked as a Class I medical device. In every other jurisdiction Microsoft states that the product is not designed, intended or made available as a medical device and is not a substitute for professional medical advice, diagnosis, treatment or judgment.

The electronic instructions for use set out the intended purpose: to assist healthcare professionals with documentation automation and workflow optimization, using AI to streamline the capture of patient information for review by qualified professionals. It is not intended to make decisions or recommendations for the diagnosis, prognosis, monitoring or treatment of individual patients, and Microsoft claims no direct benefit to individual patients. The instructions name a UK responsible person, Emergo Consulting (UK) Limited, and direct serious incidents to the manufacturer and the national competent authority.

The registration covers one edition and one territory, Great Britain rather than the whole United Kingdom. The radiologist experience is unavailable in the UK, nursing capability is in private preview in the United States only, and the clinical information extension, the part with coding suggestions and evidence summaries, carries its own transparency document. Class I is self certified by the manufacturer rather than assessed by a notified body, and brings post market surveillance and change control obligations; which features the declaration covers, and whether the extension sits inside it, is not said. UK deployments also carry NHS England's Digital Technology Assessment Criteria, ISO 27001 and DCB0129 clinical risk management.

The responsible AI risks of a speech product are named in the transparency white paper as Microsoft's own findings. The paper lists over reliance on autogenerated text, which it calls automation bias, privacy of voice data, and fairness when models underperform on dialects or accents. Its limitations section adds that transcription errors can come from accented speech not well represented in training data. Accent and dialect performance is a recurring weakness of ambient documentation.

Controls are described around the model. Prompts are grounded in the specific note data, content filters built on Azure OpenAI Service classifiers block violent, hateful, self harm and sexual content, and responsible AI checks run on the model's reply as well as on the prompt. The use case paper rules out use in a language the clinician lacks the competence to work in, and states that Dragon Copilot is not a translation tool.

Maturity is labeled as well. Nursing capability is marked as private preview and generally available in the United States only, rather than presented as shipping.

A measurement of that risk by language, accent or dialect is not published.

Where the system fails is written down in concrete terms. The transparency white paper lists transcription errors from background noise, overlapping speech or accented speech not well represented in training data. It lists summarization inaccuracies such as omitted findings or hallucinated details, especially in preview markets or languages. And it sets a latency threshold, above 250 milliseconds, beyond which real time dictation degrades. A clinician reviewing a draft knows where Microsoft itself expects errors to appear.

The same paper disclaims responsibility for output. It states that the accuracy and reliability of information provided by Dragon Copilot or the Azure OpenAI API may vary and are not guaranteed, and that Microsoft disclaims any liability for damages resulting from use of or reliance on that information. Clinicians remain responsible for the final record.

Commercially, Dragon Copilot is sold under enterprise agreements with a documented compliance program and published commercial terms, so a health system negotiates with a defined counterparty rather than accepting consumer style terms.

The named limitations show where the risk sits without moving it. No clinical performance guarantee, remediation commitment or indemnity specific to clinical output is published, and nothing says whether the master terms carry an AI indemnity that reaches clinical output rather than intellectual property claims. Patients have no route to contest a note that misdescribes their encounter and no relationship with Microsoft through which to pursue one.

Epic is the deepest integration. The product is embedded natively there, with Android and iOS parity and an Epic nursing app path for flowsheet capture, and the web application launches through SMART on FHIR R4. A partner Developer Kit lets a record system vendor embed the whole workflow, so recording is captured and the summary delivered without clinical content leaving the EHR. TruBridge ships that natively.

The developer kit changes the data path, not only the user interface. Where a scribe reaches the chart through a copy step or a browser extension, content leaves the record system and comes back. With the embedded kit it can stay inside, which removes a class of exposure rather than mitigating it.

The EHR remains the system of record. Microsoft states that Dragon Copilot is not an electronic health record or a medical records repository, and the customer is responsible for exporting and permanently storing Dragon Copilot data in the patient's record and for the reviews, approvals and filings that put it there.

Depth outside Epic varies considerably: reach across many record systems and equivalent function inside each are different claims, and only the first is shown.

Delivery is a hosted service on Microsoft infrastructure, reached through web, desktop, mobile and EHR embedded access paths, with a published subprocessor list. The partner Developer Kit changes where the workflow runs rather than offering the customer an alternative deployment, so there is effectively one deployment option.

Residency is published by customer location. Clinician availability spans ten countries, and the privacy white paper's table says where each one's data rests. US customers' data rests in the US, Canadian customers' in Canada and UK customers' in the UK. Customers in France, Ireland and Belgium have data at rest in France, within the EU, and customers in Germany, Austria, the Netherlands and Switzerland have it in Germany, within the EU. Dragon Copilot is an EU Data Boundary Service that stores and processes customer data within the EU Data Boundary, and nursing features, available only in the US, keep their data in that region.

Two parts of the data path are not mapped. Processing location is stated for the EU Data Boundary and not for the other regions, and the location of the research environment where selected customer data is anonymized within 90 days is not given. A Swiss customer's recordings rest in Germany rather than in Switzerland.

Microsoft sells Dragon Copilot through its direct sales team and through Cloud Solution Provider partners, and quotes each deal individually, with a customer's existing Microsoft agreements bearing on the price. Its licensing guidance describes a Physician Per User license that equips a clinician with all Dragon Copilot features, including front end speech, ambient and AI features, for one monthly fee without additional charges. athenahealth announced in December 2025 that Dragon Copilot would be embedded in athenaOne in the first half of 2026 as one of the ambient models its clinicians can choose.

No rate card or per user price is published, so a health system learns the cost in a sales conversation.

Physicians, nurses and radiologists each get a purpose built experience, rather than one clinician product stretched across roles. Physicians get specialty specific note templates, multilingual capture of conversations with several participants, and generated clinical letters, including consultation letters to the general practitioner, letters to patients and referral letters, along with patient friendly after visit summaries. Nursing capture populates flowsheet observations from the conversation with the patient or from the clinician's own spoken account.

Availability narrows by role and market. Nursing capability is in private preview and generally available in the United States only, and the radiologist experience is not available in the UK.

Language support follows the product documentation. Only languages listed there are supported for speech to text and ambient summarization in a given locale, and multilingual mode must be enabled by an organization administrator before clinicians can use it. Microsoft rules out use in a language the clinician lacks the competence to work in.