AI Scribes

What an AI scribe actually is

An AI scribe is a clinician-facing application that captures the patient encounter as audio, transcribes and structures it, and writes a draft note back into the EHR. Modern ambient scribes go beyond transcription: they synthesize SOAP-format notes, suggest ICD-10 / E/M / HCC codes, draft after-visit summaries and referral letters, and — increasingly — pull labs, vitals, and prior visit context forward to ground the draft. The clinician edits, signs, and the note is filed.

The category's name is misleading in two ways. First, "scribe" undersells the scope — most products are now closer to clinical assistants that touch documentation, coding, and patient communication on the same audio input. Second, "ambient" is doing a lot of work: the clinician still has to start the recording, edit the draft, and sign before anything reaches the chart. None of this is autonomous, and the published evidence is unambiguous that "signature is review" is not safe.

The four workflows that produce most of the value

What the published evidence actually shows

The evidence base in 2026 is thicker than it was eighteen months ago, and uneven across vendors. The headline numbers a buyer should know — and the caveats that go with each:

• checkUCLA NEJM AI RCT (2025) — Three-arm pragmatic randomized trial of Nabla, Microsoft DAX Copilot, and usual care. 238 outpatient physicians across 14 specialties. Nabla physicians cut time-on-notes by 41 seconds per note (4:30 → 3:49) — a statistically significant 9.5% drop. DAX showed a smaller, non-significant drop. Both arms reported ~7% improvement in burnout. NCT06792890.
• checkMass General Brigham JAMA study (2026) — Across five academic medical centers, ambient scribes reduced total EHR time by 13.4 minutes/day and documentation time by 16.0 minutes/day. Scribe usage was associated with 0.49 additional visits per week per clinician. Modest, durable, defensible.
• checkMulti-system burnout QI study (PMC, 2025) — 263 clinicians across six health systems. Burnout dropped from 51.9% to 38.8% at 30 days. ~74% reduction in odds of burnout. Significant improvements in cognitive task load, after-hours documentation, and focused attention on patients.
• checkSingle-system data points — Emory Healthcare reported a 30.7% increase in documentation-related well-being prevalence. Mass General Brigham reported a 21.2% reduction in burnout prevalence at 84 days. Cooper University Healthcare clinicians saved ~4.15 minutes per patient — about an hour daily.
• checkThe Permanente Medical Group — ~15,791 physician hours saved across the system after ambient-scribe rollout, the most-cited single-system number in the category.
• closeCaveat: STAT News (April 2026) — A large multi-site analysis found that ambient-scribe time savings are real but modest, and adoption is inconsistent — clinicians who use the tool for fewer than ten encounters often do not see durable gains. Pilots that report large numbers usually exclude low-utilization clinicians.

Where ambient scribes go wrong

The error rates are small in percentage terms and patterned in a way buyers can plan around — but the failure modes are real and the published evidence is clear that "signature is not review."

• closeHallucinations: ~1.47% of sentences. The npj Digital Medicine framework analysis (12,999 clinician-annotated sentences across 18 model configurations) reported a 1.47% hallucination rate and a 3.45% omission rate in AI-generated clinical notes. 44% of hallucinations were classified as major — capable of changing diagnosis or management if uncorrected.
• closeOmissions cluster in the parts that matter most. 55% of major omissions occurred in the "current issues" section, 35% in past medical / family / social history, 10% in the assessment and plan. Exactly where a missed item becomes a clinical safety event.
• closePronoun and attribution errors. The UCLA RCT reported one mild patient-safety event during the trial window. Most ambient-scribe failure modes outside hallucination are pronoun swaps, mis-attributed quotes, and dropped negations ("denies chest pain" rendered as "chest pain").
• closeAutomation bias is a documented risk. The 2025 npj Digital Medicine editorial "Beyond human ears" is direct about it: when clinicians are asked to review a draft, they miss errors more often than they realize. A pre-signed pre-filled note is harder to scrutinize than a blank one. Designing the workflow with this assumption is non-optional.
• closeEquity is unsolved. Accuracy degrades for non-English consultations and for patient populations underrepresented in training data. Vendors that ship strong multilingual coverage (Nabla, Suki) are not exempt — they are merely better instrumented than monolingual competitors.

The evaluation rubric that survives the demo

Vendors will offer to run a pilot with their own measurement tooling. Don't take that offer at face value. The rubric your team should fix in writing before the pilot starts:

For a published reference: the SCRIBE evaluation framework (npj Digital Medicine, 2025) and PDQI-9 are the most cited validated rubrics for AI-generated note quality. Mayo Clinic Proceedings: Digital Health published a 2025 study using simulated ambulatory encounters as the safety-evaluation methodology — useful for stress-testing under conditions that vendor demos cannot reproduce.

Cloud commercial vendors — the shortlist

The five commercial scribes that consistently make U.S. healthcare shortlists. All five are cloud-only — none currently offers a customer-tenanted on-prem deployment. Profiles, comparison, and decision short-circuits live in the vendors hub.

When the on-prem stack is the right answer instead

Cloud commercial scribes are the right answer for most U.S. health systems whose security program already operates under signed BAAs. The architectural choice every cloud vendor on this page makes — vendor-controlled inference, customer audio leaving the network — does not work for every buyer. The patterns where the on-prem stack is the cleaner answer:

If one or more of the constraints above is binding, the right next step is not a different commercial vendor — it is a different architecture. The WalledCare on-prem reference stack is the case for that path: hardware footprint, model choices, FHIR-grounded RAG pattern, and the 2026 GPU budget cheatsheet.

Adjacent categories

Pick the right pilot, not the right demo

The shortest path to a defensible AI-scribe decision is to lock the evaluation rubric before the demos start, fix the constraints in writing, and run a real pilot — same workflow, same metrics, same patient population — across the cloud and on-prem options that survive the constraint screen. WalledCare's pilot process is built around this comparison.

send Request a WalledCare pilot arrow_back Back to directory