Is DAX Copilot a scribe or a copilot?

At its core, DAX is an ambient scribe. Microsoft folded it into Dragon Copilot in 2025, which adds dictation, medical information searches, and task automation such as orders and referral letters [3]. So the product does both jobs, and you should review the documentation and the answers with different habits.

Are ambient AI scribes considered clinical decision support?

A tool that only drafts documentation is not recommending a diagnosis or treatment, so it is generally treated as a documentation tool. Once it suggests orders, diagnoses, or doses, it is doing decision support, and the FDA's 2026 guidance expects the clinician to be able to review the basis for those recommendations independently [14].

Do I still need to read every AI-drafted note?

Yes. In a 2026 UC Davis pilot, 5.3 percent of reviewed notes contained errors rated as posing serious or imminent risk if not corrected, and omissions were the most common error type [9]. Your signature makes the note yours, whoever drafted it.

Can a clinical copilot replace UpToDate?

Some physicians use AI answer engines alongside or in place of traditional references, and some scribes now embed UpToDate directly [4]. Citation accuracy varies widely by tool, from perfect in one 2026 comparison to fabrication rates above 25 percent in general-purpose models [12]. Whichever you use, check the source before changing management.

Should residents use AI scribes and copilots?

Residents can learn a lot from a copilot that shows its sources, but writing notes is also how clinical reasoning gets built. The AMA found 88 percent of physicians have at least some concern about AI-related skill loss [1]. Check your program's policy first, and be deliberate about which parts of the work you hand off.

Back to blogs

Ambient Scribe vs Clinical Copilot: The Real Difference

An ambient scribe writes your note. A clinical copilot answers your questions. Here is how they differ in risk, review burden, and regulation.

Reviewed by O. Daniel Odutola, MD, MBA, MPHNovember 5, 202613 min read

An ambient AI scribe listens to the visit and drafts your note. A clinical copilot answers the questions you ask it, from dosing to differentials, and ideally shows you where the answer came from. That is the core of the ambient scribe vs clinical copilot question: one tool records what already happened, the other tries to shape what happens next, and that difference changes how much you have to check its work.

The confusing part is that vendors now sell both in the same box, sometimes under the same name. So the useful question is not which brand you have. It is which job the software is doing at a given moment, and what can go wrong when it does that job badly.

Key takeaways

A scribe's job is documentation. A copilot's job is answering clinical questions. The failure modes are different, so the review habits have to be different too.
Scribe errors are mostly omissions and invented details in the note. Copilot errors are wrong or unsupported answers, including fabricated citations.
The product lines are merging. Microsoft, Abridge, OpenEvidence, and Epic all now blend listening, drafting, and answering in one workflow.
Once a tool starts suggesting orders or diagnoses, it has crossed from documentation into decision support, and your review standard should rise with it.
Match the tool to your real bottleneck. If the pain is pajama time, start with a scribe. If the pain is the question you cannot answer between rooms, start with a copilot.

Why this distinction suddenly matters

Most of us did not pick a category on purpose. Someone in IT turned on a pilot, a colleague mentioned a free app, and now there are two or three AI tools touching your day.

Adoption is no longer the fringe. In the AMA's March 2026 survey, 81 percent of physicians reported using AI professionally, more than double the 2023 figure [1]. The top uses split right along the line this article is about: 39 percent used it to summarize research and standards of care, while 28 percent used it to document visit notes, billing codes, or charts [1].

Doximity's 2026 report, drawn from 3,151 physicians across 15 specialties, found the same split. Literature search for clinical reference was the most common use at 35 percent, and ambient scribes came second at 29 percent [2]. Daily AI use climbed to 63 percent from 47 percent a year earlier, and 71 percent named accuracy and reliability as their top concern [2].

That last number is the reason categories matter. "Is it accurate?" means something different for a note than for an answer.

Ambient scribe vs clinical copilot: the short version

Here is the side-by-side most comparison pages skip. It sorts the tools by what they do and how they fail, not by logo.

Dimension	Ambient AI scribe	Clinical copilot
Core job	Turns the visit conversation into a draft note	Answers clinical questions you type or speak
Input	Audio of the encounter, sometimes chart context	Your question, sometimes patient context and the literature
Output	HPI, exam, assessment and plan, patient instructions	Answers, summaries, differentials, dosing, references
Timing	During and right after the visit	Whenever you get stuck, before or during a decision
Typical failure	Omissions, invented details, wrong speaker	Wrong or unsupported answer, fabricated or misread citations
Who catches the error	You, reading the note before signing	You, checking the answer against a source
What it replaces	Typing, dictation, a human scribe	A search, a reference lookup, a curbside consult
Regulatory lens	Generally a documentation tool	Can fall under FDA clinical decision support rules
Best measure of value	Minutes saved and after-hours charting	Right answers, fast, with a checkable source

The last row is the one to hold vendors to. A scribe that saves time but slips errors into the chart is not a win. A copilot that answers fast but hides its sources is not either.

See how your offer compares

Compare real pay data before you sign anything

Physicians' Copilot Clinical AI

What an ambient AI scribe actually does

An ambient scribe records the conversation, transcribes it, and uses a language model to write a structured note. You review, edit, and sign. The better ones let you set templates and a house style, and some push the draft straight into the EHR.

The market filled up fast. Doximity began offering a free scribe to verified U.S. physicians, NPs, PAs, and medical students in July 2025 [7]. Epic released its own AI Charting tool in February 2026, part of a broader assistant it calls Art, which drafts notes from the visit and can suggest orders based on the conversation [6].

The evidence on time savings is real but modest, and it varies by product. In the first randomized trial of ambient scribes, 238 outpatient physicians at UCLA were assigned to DAX Copilot, Nabla, or usual care. Nabla users cut time-in-note by 9.5 percent versus control, while DAX users showed no significant change [8]. Both groups reported improvements in burnout and task load, though the authors called those secondary findings in need of confirmation [8].

Health system leaders seem to be buying it for burnout rather than the balance sheet. The Peterson Health Technology Institute's March 2025 review found scribes potentially effective at reducing documentation time and cognitive load, with gaps in evidence on productivity and financial performance [10].

How scribes fail

Scribes fail quietly. The note looks finished, which is exactly the problem.

A UC Davis pilot published in April 2026 is the most useful look at this so far. Thirty-one physicians produced 7,545 AI-drafted notes, and physicians reviewed 356 of them in detail. Omissions were the most common error (18 percent of reviewed notes), followed by hallucinations (11.5 percent) and accidental inclusions (9.3 percent) [9]. Most errors were mild to moderate, but 5.3 percent of notes had errors rated as posing serious or imminent risk if left uncorrected [9].

Two other numbers from that study stuck with me. Across the notes with editing data, a median of just 9 percent of the AI's words were changed, and 14.9 percent of notes were left entirely unedited [9]. Editing ranged from 1.9 percent to 69.3 percent of words across physicians [9]. Some of us are reading these notes line by line. Some of us are not.

For a deeper look at those risks, see our breakdown of whether AI medical scribes are safe.

What a clinical copilot actually does

A clinical copilot is the tool you ask. "Renal dosing for this antibiotic at a CrCl of 25?" "What else causes this pattern of LFTs?" "What does the latest guideline say about stopping this drug before surgery?" It pulls from the literature or a curated knowledge base and hands you an answer, ideally with the references attached.

This is the lane that point-of-care references have occupied for years, now with a conversational front end. If you want the wider map of that space, our guide to point-of-care tools in 2026 walks through the options, old and new.

How copilots fail

In principle, copilot errors are easier to spot, because the answer sits right in front of you. The trouble is that a wrong answer sounds exactly as confident as a right one. The risk is an answer you act on, or a reference that does not exist.

Citation quality varies enormously by tool. In a June 2026 comparison on spine surgery questions, OpenEvidence produced perfectly accurate citations, while general-purpose models fabricated up to 26.2 percent of their references, and every general model did worse on patient-level prompts [12]. A separate August 2026 study of three general models in neurocritical care, with web retrieval turned off, found 55.0 percent of 300 citations contained an error and 28.3 percent were completely fabricated [13].

There is a second, less obvious failure: the tool can be right and still not help you. In a 2024 randomized trial, 50 physicians given access to an LLM scored 2 percentage points higher on diagnostic reasoning than those using conventional resources, a difference that was not significant [11]. The LLM working alone scored 16 points higher than the conventional group [11]. Having a good copilot and using it well turn out to be separate skills.

See where Physicians' Copilot fits. We built Physicians' Copilot for the question-answering side of this line, with the evidence visible so you can check it yourself.

The lines are blurring on purpose

If you only remember one thing from this article, make it this: the product names no longer tell you the category.

Microsoft's Dragon Copilot, announced in March 2025, merged DAX ambient listening with Dragon Medical One dictation and added medical information searches from trusted content sources plus automation of orders, referral letters, and after-visit summaries [3]. So the product called "Copilot" is, at its core, a scribe that grew a search box.

It runs the other way too. Abridge, a scribe company, integrated Wolters Kluwer's UpToDate into its note workflow, generally available at the end of March 2026, so clinicians can ask patient-specific questions in the same place the note gets drafted [4]. It has since signed agreements to bring in NEJM and JAMA Network content [4]. OpenEvidence, which started as an answer engine, now offers Visits, which transcribes encounters and surfaces guidance while you chart [5].

Epic's chief medical officer put the direction plainly when AI Charting launched: "We're just going to keep adding more besides just notes. So it's going to be orders, it's going to be diagnoses" [15].

None of this is bad. A scribe that knows the guideline could flag the missed vaccine. But it does mean a single tool can be doing two jobs with two different risk profiles, sometimes in the same minute.

Scribe vs decision support: where the regulatory line sits

This is the part most comparison pages leave out, and it is the part that should change your behavior.

A tool that only turns a conversation into a note is documenting. A tool that recommends a diagnosis, an order, or a dose is supporting a clinical decision. The FDA's revised clinical decision support guidance, issued January 6, 2026, lays out four criteria that software must meet to stay outside device regulation [14]. The fourth is the one that matters most at the bedside: the software must let the clinician independently review the basis for its recommendation, so the clinician is not relying primarily on the tool [14].

The 2026 revision also added enforcement discretion for tools that give a single recommendation when only one is clinically appropriate, and it tied the time-critical limitation to that same fourth criterion [14].

In plain terms, the regulators expect you to be able to see why the software said what it said. When your scribe starts proposing orders, it is doing decision support, whatever the marketing calls it. Hold it to the same standard you would hold a copilot: show me the basis, or I do not act on it.

For the bigger picture on how these rules and tools fit together, see our pillar guide to AI in clinical medicine.

Which one do you actually need?

Start with your worst hour of the day, not the demo.

Name the bottleneck. If you are finishing notes at 10 p.m., the problem is documentation. If you are stuck between rooms on a dosing or workup question, the problem is information.
Pick the tool for that job. Documentation pain points to a scribe. Information pain points to a copilot. Our guide to speeding up clinical documentation covers the non-AI fixes worth trying alongside a scribe.
Check what your system already licenses. Many health systems now have an ambient tool built into the EHR. Using the sanctioned one keeps patient audio inside a signed agreement.
Test it on your hardest visits. A scribe that handles a straightforward med refill says little about a 40-minute visit with an interpreter and a family member talking over each other.
Decide your review rule before you start. For a scribe, that means reading every assessment and plan, every medication, and every negative you did not say. For a copilot, it means opening at least one cited source before you act on an answer that changes management.
Re-check after a month. Track your own after-hours charting time and the errors you catch. If you are catching fewer errors over time, ask whether the tool got better or you started skimming.

Many physicians will end up using both. That is fine, as long as you know which hat the software is wearing when it hands you something.

The bottom line

A scribe saves your evenings. A copilot saves your uncertainty. Both can hurt a patient if you stop reading what they produce, and the newest products do both jobs at once, so the category on the invoice is less important than the job in front of you.

No AI tool replaces clinical judgment. Verify all outputs against primary sources and your own assessment.

See where Physicians' Copilot fits. If your gap is the question you cannot answer between rooms, that is the job we built for.

REFERENCES

American Medical Association. "More than 80% of physicians use AI professionally: AMA survey." AMA, March 12, 2026. https://www.ama-assn.org/practice-management/digital-health/more-80-physicians-use-ai-professionally-ama-survey. Accessed September 22, 2026.
HIT Consultant. "Doximity Releases 2026 Report on Physician AI Adoption Across 15 Medical Specialties" (Doximity 2026 State of AI in Medicine Report, n = 3,151). HIT Consultant, March 18, 2026. https://hitconsultant.net/2026/03/18/doximity-2026-ai-medicine-report-physician-adoption-pajama-time/. Accessed September 22, 2026.
Microsoft. "Microsoft Dragon Copilot provides the healthcare industry's first unified voice AI assistant." Microsoft Source, March 3, 2025. https://news.microsoft.com/source/2025/03/03/microsoft-dragon-copilot-provides-the-healthcare-industrys-first-unified-voice-ai-assistant-that-enables-clinicians-to-streamline-clinical-documentation-surface-information-and-automate-task/. Accessed September 22, 2026.
Fierce Healthcare. "Abridge expands clinical decision support solution with UpToDate partnership, new NEJM, JAMA content tie-ups." Fierce Healthcare, April 15, 2026. https://www.fiercehealthcare.com/ai-and-machine-learning/abridge-expands-clinical-decision-support-capabilities-uptodate-partnership. Accessed September 22, 2026.
OpenEvidence. "Visits Overview." OpenEvidence User Guide, undated product page. https://www.openevidence.com/user-guide/visits-overview. Accessed September 22, 2026.
Healthcare Dive. "Epic rolls out AI charting tool as scribe market heats up." Healthcare Dive, February 5, 2026. https://www.healthcaredive.com/news/epic-rolls-out-ai-charting-art-notetaking-documentation-scribe/811462/. Accessed September 22, 2026.
Fierce Healthcare. "Doximity jumps into AI scribe market, offering a free tool for doctors." Fierce Healthcare, July 24, 2025. https://www.fiercehealthcare.com/ai-and-machine-learning/doximity-jumps-ai-scribe-market-offering-free-tool-doctors. Accessed September 22, 2026.
Lukac PJ, Turner W, Vangala S, et al. "Ambient AI Scribes in Clinical Practice: A Randomized Trial." NEJM AI, November 26, 2025;2(12). doi:10.1056/AIoa2501000. https://doi.org/10.1056/aioa2501000. Accessed September 22, 2026.
Taylor SL, Jost M, MacDonald S, et al. "Quality of Clinical Notes Created by Ambient Listening Generative AI: Pragmatic Prospective Pilot Study." JMIR Medical Informatics, April 17, 2026;14:e86474. doi:10.2196/86474. https://doi.org/10.2196/86474. Accessed September 22, 2026.
Peterson Health Technology Institute. "Leading Health Systems: AI-Powered Scribes Alleviate Clinician Burnout; Financial Impact Unclear." PHTI, March 25, 2025. https://phti.org/announcement/ai-scribes-reduce-clinician-burnout/. Accessed September 22, 2026.
Goh E, Gallo R, Hom J, et al. "Large Language Model Influence on Diagnostic Reasoning: A Randomized Clinical Trial." JAMA Network Open, October 1, 2024;7(10):e2440969. doi:10.1001/jamanetworkopen.2024.40969. https://doi.org/10.1001/jamanetworkopen.2024.40969. Accessed September 22, 2026.
McLaughlin ND, Srinivas AN, Lowe ZF, et al. "Large Language Model Hallucinations in Spine Surgery: A Comparative Analysis of Clinician vs Patient-Level Prompts." Neurosurgery Practice, June 2, 2026;7(3):e000244. doi:10.1227/neuprac.0000000000000244. https://doi.org/10.1227/neuprac.0000000000000244. Accessed September 22, 2026.
Seifi A, Seyfi A. "Hallucination Rate of Peer-Reviewed Citations Generated by Large Language Models in Neurocritical Care." Critical Care Explorations, August 25, 2026;8(9):e1474. doi:10.1097/CCE.0000000000001474. https://doi.org/10.1097/CCE.0000000000001474. Accessed September 22, 2026.
Covington and Burling LLP. "5 Key Takeaways from FDA's Revised Clinical Decision Support (CDS) Software Guidance" (FDA guidance issued January 6, 2026). Covington, January 2026. https://www.cov.com/en/news-and-insights/insights/2026/01/5-key-takeaways-from-fdas-revised-clinical-decision-support-cds-software-guidance. Accessed September 22, 2026.
STAT. "Epic launches AI Charting, potentially scrambling the ambient scribe market." STAT, February 4, 2026. https://www.statnews.com/2026/02/04/epic-ai-charting-ambient-scribe-abridge-microsoft/. Accessed September 22, 2026.

FAQs

An AI scribe listens to the patient encounter and writes a draft note for you to review and sign. A clinical copilot answers clinical questions, such as dosing, differentials, or guideline recommendations, usually with references. The scribe documents the decision you already made, while the copilot tries to inform the next one.