Do ambient AI scribes actually save time?

Sometimes, and less than advertised. One randomized trial found a 9.5 percent reduction in time-in-note for one product and no significant change for another [5]. A crossover trial found pajama time largely unaffected [7]. The more consistent benefit across studies is reduced cognitive load and burnout [5][6][7].

Does AI improve diagnostic accuracy?

Not automatically. A randomized trial found physicians given an LLM as a diagnostic aid performed no better than those using conventional resources, even though the model alone outperformed both groups [8]. A later trial found a 6.5 percentage point improvement on management reasoning tasks [9]. How you use the tool matters more than which tool you use.

Am I liable if an AI tool contributes to an error?

Your signature on the note is the legal event, and AI-generated documentation is discoverable in malpractice and regulatory proceedings [15]. Liability allocation between physician, employer, and vendor is unsettled and varies by contract and state. Ask who indemnifies what before you adopt, and have an attorney review any agreement that assigns you risk.

How do I know whether the prediction model in my EHR is any good?

Ask for local validation data. Under the ONC HTI-1 rule, certified health IT with Predictive Decision Support Interventions must disclose 31 source attributes covering training data, validation, and intended use [16]. The Epic Sepsis Model example, an AUC of 0.63 with 67 percent of sepsis cases missed in external validation, is why this question is not academic [12].

Which AI tools are FDA-cleared?

Mostly imaging and signal analysis software. FDA had authorized more than 1,000 AI-enabled devices as of January 2025 [3], and the public list is dominated by radiology [4]. Most ambient scribes and general-purpose clinical chatbots are not regulated as devices, which means the validation burden shifts to your institution and to you.

Back to blogs

AI in Clinical Medicine: A Practical Guide for Working Physicians

What the trials really show about AI in clinical medicine, which tools earn their keep, and how to verify output before it reaches the chart.

Reviewed by O. Daniel Odutola, MD, MBA, MPHNovember 3, 202616 min read

AI in clinical medicine is no longer a pilot project. Roughly 81 percent of physicians now report using AI professionally, more than double the 2023 rate [1], and the strongest evidence supports a narrow slice of that use: ambient documentation, literature retrieval, and drafting. The diagnostic and management benefits are real but smaller and more conditional than the marketing suggests, and the failure modes are quiet enough that you will not notice them unless you are looking.

That last sentence is the whole article, honestly. Everything below is the detail.

Where AI in clinical medicine actually stands in 2026

Two numbers frame the moment. The American Medical Association surveyed nearly 1,700 physicians and found 81 percent using AI in some professional capacity as of March 2026, up from roughly 38 percent in 2023 [1]. Doximity surveyed 3,151 physicians across 15 specialties and found 63 percent using AI daily in early 2026, up from 47 percent a year earlier [2].

So adoption is not the question anymore. Calibration is.

On the regulated side, the FDA had authorized more than 1,000 AI-enabled devices through established premarket pathways as of January 2025 [3], and its public list has kept growing since, with entries running into mid-2026 [4]. The overwhelming majority are imaging tools. Almost none of the software most physicians actually touch every day, the chatbot, the ambient scribe, the inbox drafter, sits inside that regulated perimeter at all. It reaches you through your employer's IT contract, not through a 510(k).

That gap matters more than any benchmark score. The tools with the most rigorous oversight are the ones you use least. The tools you use hourly arrived with a vendor demo and a training module.

This article is educational and is not legal advice. Contract terms vary by state and employer. Have any agreement reviewed by a licensed attorney before signing.

The four jobs physicians are actually handing to AI

Most confusion about clinical AI tools comes from treating one category's evidence as if it applies to all four. It does not. The verification burden is completely different depending on what the tool is doing.

Category	What it does	What the evidence supports	Main failure mode	Who checks it
Ambient documentation	Listens to the visit, drafts the note	Modest reductions in note time, consistent improvement in burnout and task load [5][6]	Confident fabrication of history you never took; omission of a key negative	You, every note, before signing
Clinical reference and search	Answers a question, cites literature	Fast retrieval; accuracy varies enormously by platform [11]	Fabricated or misattributed citations [10]	You, against the primary source
Predictive models in the EHR	Risk scores, deterioration and sepsis alerts	Highly variable; strong local validation is the exception, not the rule [12]	Silent performance decay; alert fatigue	Your institution, ideally with published local validation
Drafting and administrative work	Portal replies, prior auth letters, patient instructions	Widely used, 30 percent of physicians report it [1]	Tone and content drift; medicolegal exposure of a draft you signed without reading	You, before it leaves the chart

Two categories put you personally on the hook for every output. One puts your institution on the hook and gives you almost no visibility. One is low risk right up until you sign something you did not read.

If you are trying to sort out where a specific product falls, the distinction between passive listening tools and interactive ones is worth understanding on its own, and we cover it in detail in ambient AI scribes versus clinical copilots.

See how your offer compares

Compare real pay data before you sign anything

Physicians' Copilot Clinical AI

What the trial evidence says, and what it does not

Here is where the honest reading gets uncomfortable for everyone, vendors and skeptics alike.

Ambient scribes: real, modest, and vendor-specific. The first randomized trial of ambient AI scribes, published in NEJM AI in November 2025, assigned 238 outpatient physicians across 14 specialties to one of two commercial scribes or usual care [5]. One product cut time-in-note by 9.5 percent versus control. The other showed no significant change. Both improved burnout, task load, and work exhaustion scores. Clinically significant inaccuracies were reported as occurring "occasionally" with both.

Read that again. Two mature products, same trial, same institution, and only one moved the primary outcome. The category works. A given product may not.

A multicenter quality improvement study of 263 ambulatory clinicians found burnout dropped from 51.9 percent to 38.8 percent after 30 days with an ambient scribe, with about 0.9 fewer hours of after-hours documentation [6]. That study was uncontrolled and included a vendor author, which is worth knowing when you weigh it.

A randomized crossover trial of 160 clinicians at an academic center found both tools improved satisfaction and reduced burnout scores, but pajama time was "largely unaffected" [7]. So the wellbeing signal is more consistent than the time signal. That is a real finding, not a disappointing one: feeling less cognitively taxed at the end of clinic has value even if the clock does not move much.

Diagnostic reasoning: the result nobody quotes correctly. In a 2024 randomized trial, 50 physicians given access to an LLM as a diagnostic aid scored no better on structured diagnostic reasoning than physicians with conventional resources, a 2 percentage point difference that was not significant [8]. The part people skip: the LLM working alone scored 16 percentage points higher than the physicians using conventional resources.

The tool was better than the doctors. The doctors plus the tool were not better than the doctors. That is a finding about human-AI interaction, not about model capability, and it should make you curious about your own habits rather than smug.

A follow-up trial in Nature Medicine found that for management reasoning, which is messier and more like actual practice, physicians using GPT-4 did outperform conventional resources by 6.5 percentage points [9]. Small, real, and it took them about two extra minutes per case.

How to verify AI output before it reaches the chart

This is the section to keep. Using AI in medicine safely is a workflow problem, not an attitude problem.

Start with what fabrication actually looks like. A 2026 cross-sectional evaluation asked three leading models for peer-reviewed neurocritical care references under retrieval-disabled conditions: 55.0 percent of 300 citations contained an inaccuracy, and 28.3 percent were completely fabricated, with per-model rates ranging from 23 percent to 73 percent [10]. A separate 2026 analysis in spine surgery found general-purpose models produced fabricated citations at rates up to 26.2 percent, and got substantially worse on patient-facing prompts, while a PubMed-indexed clinical platform (OpenEvidence) achieved perfect citation accuracy in that sample [11].

The lesson is not "AI lies." The lesson is that retrieval architecture is the single biggest predictor of whether a citation is real. A model answering from memory invents. A model searching a real index does not have to.

Here is the protocol I would hand a resident on day one:

Decide the stakes before you ask. If the answer changes management, it needs a primary source. If you are drafting a patient instruction about hydration, it does not. Sort every query into one bucket or the other before you type it.
Use a tool that cites, and use one that searches. Prefer a platform that retrieves from an indexed corpus over a general chatbot answering from memory. This one change eliminates most fabrication risk.
Open the citation. Not the abstract snippet the tool shows you. The actual paper. If the tool cannot produce a working DOI or PMID, treat the claim as unsourced.
Check that the study population is yours. A fabricated citation is obvious once you look. A real citation that studied a different population is the error that survives review and ends up in your note.
Read the AI draft as if a resident wrote it at 2 a.m. You already have this skill. Ambient notes fail the same way tired trainees do: plausible history that was never obtained, pertinent negatives that were never asked, a plan that drifts from what you said.
Sign nothing you did not read. Your signature is the legal event. The tool has no license.

If you want the prompting side of this in depth, we walk through question construction, source constraints, and how to make a model say "I do not know" in getting reliable clinical answers from AI. The taxonomy of when fabrication happens and how to spot it is covered in AI hallucinations in medicine.

See how Physicians' Copilot fits a safe AI workflow. The point is not to add another tab. It is to make the verification step fast enough that you actually do it.

Three failure modes that should worry you more than hallucinations

Hallucinations get the headlines because they are easy to demonstrate. These three are harder to see and, in my view, more dangerous.

Automation bias. When a tool is right most of the time, your threshold for questioning it drops. The 2024 diagnostic reasoning trial is a clean illustration: physicians had access to a system that outperformed them, and their performance did not improve [8]. Anchoring on a machine's differential is still anchoring. The countermeasure is boring and effective: form your own impression first, then ask the tool, then reconcile the difference explicitly.

Models that never worked as well as advertised. The Epic Sepsis Model was deployed at hundreds of US hospitals. An external validation at Michigan Medicine covering 38,455 hospitalizations found an AUC of 0.63, missed 67 percent of patients with sepsis, and generated alerts on 18 percent of all hospitalizations [12]. That is a tool that was simultaneously insensitive and exhausting. It was not a fringe product. It shipped inside the EHR most of us use.

Ask your informatics team one question about any predictive model firing in your workflow: has it been validated on our patients, and what were the numbers? If nobody can answer, that is the answer.

Bias that predates the model. A widely used population health algorithm affecting millions of patients was found to systematically under-refer Black patients, because it predicted healthcare costs as a proxy for illness, and less money is spent on Black patients for the same disease burden [13]. Correcting the target variable would have raised the share of Black patients flagged for extra help from 17.7 percent to 46.5 percent. No malice, no bad data in the narrow sense. Just a proxy that encoded an inequity and then scaled it.

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

Three practical rules cover most of the risk.

Protected health information only goes into tools your organization has a business associate agreement with. Consumer chatbot accounts do not qualify. A vendor-specific enterprise agreement should address model training restrictions, subcontractors, audit rights, and post-termination data deletion [15]. If you cannot name the BAA, assume there is not one.

Recording a patient encounter is a separate legal question from documenting it. Health law commentators have flagged litigation in California and Illinois alleging that health systems deployed ambient scribing without adequate patient consent, implicating state wiretapping and confidentiality statutes [15]. Consent requirements vary by state, and a checkbox buried in a registration packet is not obviously the same thing as informed consent to being recorded. This is worth a real conversation with your compliance office, not an assumption. We go deeper on the consent and accuracy questions in are AI medical scribes safe.

Assume every AI-generated draft is discoverable. Documentation produced with AI assistance can be scrutinized in malpractice claims, privacy actions, and regulatory investigations [15]. The draft is not privileged because a machine wrote it.

One piece of structure does exist. Under the ONC HTI-1 final rule, certified health IT that includes Predictive Decision Support Interventions must disclose 31 source attributes about each one, including training data, validation, and intended use, with developer compliance required by December 31, 2024 [16]. You are entitled to that information about the models running in your EHR. Very few physicians ask for it. It is sitting there.

On the billing side, CPT Appendix S gives a taxonomy for AI-enabled services, sorting software into assistive, augmentative, and autonomous categories based on how much interpretive work the algorithm does, with the CPT Editorial Panel strengthening those definitions in May 2026 [14]. It is the vocabulary your coders and your contract will use, so it is worth knowing which bucket a tool falls into before you agree to be measured on it.

This article is educational and is not legal advice. Contract terms vary by state and employer. Have any agreement reviewed by a licensed attorney before signing.

A checklist before you adopt any clinical AI tool

Run this before the pilot, not after.

Is there a signed BAA covering this tool, and can someone show it to me?
Does it retrieve from a real, citable index, or answer from memory?
Has it been validated on a population that resembles mine, and are the numbers published?
What is the documented failure mode, and how would I detect it in my own workflow?
Who is accountable if its output causes harm: me, my employer, or the vendor?
What is the patient consent process, and does it satisfy my state's requirements?
How long is audio or input data retained, and who else can access it?
Can I turn it off without losing access to the rest of my workflow?

If you are still deciding which category of tool you need in the first place, start with our overview of point-of-care tools for finding answers fast and the head-to-head comparison in UpToDate versus OpenEvidence versus the new AI tools. For the specific traps of using a general model in real time during a patient encounter, see how to use AI at the point of care without getting burned.

Where this is going, and what it asks of you

The uncomfortable finding from the diagnostic reasoning trial has not been replicated away: the bottleneck is increasingly us, not the model [8]. That is not a reason to defer to software. It is a reason to get deliberate about how we interrogate it, the same way we got deliberate about reading our own imaging rather than trusting a preliminary read.

Physicians in the AMA survey were clear about the conditions. 88 percent want robust validation of safety and efficacy, 86 percent prioritize data privacy, 85 percent want to be consulted on adoption decisions, and 88 percent worry about skill loss in training physicians [1]. Those are not luddite positions. They are the terms of a reasonable deal, and most of us have not been offered one yet.

Ask for the validation data. Read the note before you sign it. Keep your own differential.

See how Physicians' Copilot fits a safe AI workflow. Built for the way you already verify, not around it.

Key takeaways

81 percent of physicians use AI professionally and 63 percent use it daily, so the question is no longer whether to adopt but how to verify [1][2].
Ambient scribes are the best-evidenced use case, with one randomized trial showing a 9.5 percent reduction in note time for one product and no significant change for another, alongside consistent burnout improvement [5][6].
Fabricated citations are common in general-purpose models (28.3 percent completely fabricated in one 2026 evaluation), and retrieval-based clinical platforms perform far better [10][11].
Automation bias, unvalidated EHR prediction models, and proxy-driven algorithmic bias are quieter risks than hallucinations and harder to catch [8][12][13].
Protected health information belongs only in tools covered by a business associate agreement, and consent to record is a separate legal question from consent to treat [15].

REFERENCES

American Medical Association. "More than 80% of physicians use AI professionally: AMA survey" (reporting the 2026 Physician Survey on Augmented Intelligence, n = approximately 1,700). American Medical Association. Published March 12, 2026. https://www.ama-assn.org/practice-management/digital-health/more-80-physicians-use-ai-professionally-ama-survey. Accessed September 15, 2026.
Doximity. 2026 State of AI in Medicine Report (n = 3,151 physicians across 15 specialties), as reported by HIT Consultant. Published March 18, 2026. https://www.doximity.com/reports/state-of-ai-medicine-report/2026. Accessed September 15, 2026.
US Food and Drug Administration. "FDA Issues Comprehensive Draft Guidance for Developers of Artificial Intelligence-Enabled Medical Devices." FDA press announcement. Published January 6, 2025. https://www.fda.gov/news-events/press-announcements/fda-issues-comprehensive-draft-guidance-developers-artificial-intelligence-enabled-medical-devices. Accessed September 15, 2026.
US Food and Drug Administration. "Artificial Intelligence-Enabled Medical Devices" (searchable authorization list). FDA. List entries current through mid-2026. https://www.fda.gov/medical-devices/software-medical-device-samd/artificial-intelligence-enabled-medical-devices. Accessed September 15, 2026.
Lukac PJ, Turner W, Vangala S, et al. "Ambient AI Scribes in Clinical Practice: A Randomized Trial." NEJM AI. 2025;2(12). Published November 26, 2025. https://doi.org/10.1056/aioa2501000. Accessed September 15, 2026.
Olson KD, Meeker D, Troup M, et al. "Use of Ambient AI Scribes to Reduce Administrative Burden and Professional Burnout." JAMA Network Open. 2025;8(10):e2534976. Published October 1, 2025. https://doi.org/10.1001/jamanetworkopen.2025.34976. Accessed September 15, 2026.
Chowdhury A, Casey M, Wilson J, et al. "Comparing ambient scribes: a randomized crossover clinical trial addressing ambient scribe technologies' impact on physician burnout." Journal of the American Medical Informatics Association. 2026;33(5):990-999. Published May 1, 2026. https://doi.org/10.1093/jamia/ocag018. Accessed September 15, 2026.
Goh E, Gallo R, Hom J, et al. "Large Language Model Influence on Diagnostic Reasoning: A Randomized Clinical Trial." JAMA Network Open. 2024;7(10):e2440969. Published October 28, 2024. https://doi.org/10.1001/jamanetworkopen.2024.40969. Accessed September 15, 2026.
Goh E, Gallo RJ, Strong E, et al. "GPT-4 assistance for improvement of physician performance on patient care tasks: a randomized controlled trial." Nature Medicine. 2025;31(4):1233-1238. Published February 5, 2025. https://doi.org/10.1038/s41591-024-03456-y. Accessed September 15, 2026.
Seifi A, Seyfi A. "Hallucination Rate of Peer-Reviewed Citations Generated by Large Language Models in Neurocritical Care." Critical Care Explorations. 2026;8(9):e1474. Published August 25, 2026. https://doi.org/10.1097/CCE.0000000000001474. Accessed September 15, 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. 2026;7(3):e000244. Published June 2, 2026. https://doi.org/10.1227/neuprac.0000000000000244. Accessed September 15, 2026.
Wong A, Otles E, Donnelly JP, et al. "External Validation of a Widely Implemented Proprietary Sepsis Prediction Model in Hospitalized Patients." JAMA Internal Medicine. 2021;181(8):1065-1070. Published August 1, 2021. https://doi.org/10.1001/jamainternmed.2021.2626. Accessed September 15, 2026.
Obermeyer Z, Powers B, Vogeli C, Mullainathan S. "Dissecting racial bias in an algorithm used to manage the health of populations." Science. 2019;366(6464):447-453. Published October 25, 2019. https://doi.org/10.1126/science.aax2342. Accessed September 15, 2026.
American Medical Association. "CPT Appendix S: Taxonomy for artificial intelligence in medical services and procedures" (Appendix S introduced 2021; definitions strengthened at the May 2026 CPT Editorial Panel meeting). American Medical Association. https://www.ama-assn.org/practice-management/cpt/cpt-appendix-s-taxonomy-artificial-intelligence-medical-services-procedures. Accessed September 15, 2026.
Ummer-Hashim S. "Ambient AI Scribes: Efficiency Gains vs Emerging Privacy and Cybersecurity Risks." American Bar Association, Health Law Section. Published 2026. https://www.americanbar.org/groups/health_law/news/2026/ambient-ai-scribes-privacy-cybersecurity/. Accessed September 15, 2026.
Office of the National Coordinator for Health Information Technology (now ASTP/ONC). "HTI-1 Decision Support Interventions (DSI) Fact Sheet" (31 source attributes for Predictive DSIs; developer compliance by December 31, 2024). HealthIT.gov. Published December 2023. https://www.healthit.gov/sites/default/files/page/2023-12/HTI-1_DSI_fact%20sheet_508.pdf. Accessed September 15, 2026.

FAQs

Not with identifiable patient information, unless your organization has a business associate agreement covering that specific product and configuration [15]. For de-identified clinical questions, general models are usable but produce fabricated citations at meaningful rates [10][11], so anything that changes management needs verification against a primary source.