AI Scribes Improved Note-Taking Time but Not Productivity, Study Finds

— More frequent users reaped more benefits

The use of an artificial intelligence (AI)-driven ambient scribe in outpatient practice was associated with reductions in time spent taking notes and increased note length, but other productivity measures were largely unchanged, researchers found.

Among all users, median note-active time decreased from 21 minutes to 18.6 minutes and median note length increased from 4,854 characters to 5,187 characters after clinicians enrolled in Heidi, an ambient scribe system, according to Jennifer P. Stevens, MD, of Beth Israel Deaconess Medical Center in Boston, and colleagues.

However, no differences were seen in same-day or after-hours note closure, or in clinical productivity, they reported in Annals of Internal Medicineopens in a new tab or window.

"By reducing documentation burden, AI scribes offer a promising strategy to address a contributor to EHR [electronic health record]-related clinician burnout," the researchers wrote. "A flexible, clinician-driven, implementation strategy can support rapid deployment of AI tools. Future research should evaluate the impact of AI scribes on clinician productivity, documentation accuracy, and health equity."

Clinician burnout continues to be a problem in medical practice, with 43.2% of clinicians reporting burnout symptoms in 2024, the authors noted in their introduction. EHR-related tasks are a major contributor to this problem because they reduce time spent directly with patients and can also extend into after-work hours, with some clinicians spending an average of 3 hours on the EHR in their off-hour days. The researchers wanted to see if ambient scribes could alleviate some of this burden.

However, the study's conclusions "suggest that ambient scribes may have unintended consequences that require a broader set of research methods and questions," Kirsten Ostherr, PhD, MPH, of Rice University in Houston, wrote in an accompanying editorialopens in a new tab or window. "Although the modest reduction in note-active time aligns with previous research, other studies have found that additional time required for review and editing of the AI-generated note offsets the initial time saved."

The study also didn't capture the time that clinicians spent editing notes within Heidi before transferring them to the EHR. "This ambiguity may impact the true measure of note-active time and raises an important question about whether ambient scribes relocate documentation burden rather than reduce it," she added.

"Modest reductions in note-active time, coupled with the unexplained increase in note length, raise the possibility that ambient scribes transform patients' narratives with downstream consequences for the quality of care they receive," she said, adding that these unintended consequences signal a need for more AI regulation. "Ambient scribes are currently classified as low-risk medical devices that do not require FDA review. Yet, the long-term impact of incorporating AI-generated content into clinical notes -- potentially including fabrications, omissions, and misinterpretations -- is unknown."

For their study, researchers analyzed the introduction of Heidi into the authors' 14-hospital health system, Beth Israel Lahey Health. The system was voluntary for clinicians to join and implemented separately from the EHR so clinicians could incorporate it into their workflow as needed.

"Our primary aim was to estimate the effect of Heidi enrollment on clinician documentation burden, comparing outcomes before versus after enrollment among enrolled clinicians," the investigators explained. "Our secondary aim was to estimate differential changes across high versus low user levels with enrolled nonusers as the comparison group."

The system was launched in January 2025 and implemented in several phases; physicians, non-physician practitioners, trainees, and allied health professionals were eligible to enroll. Among 4,993 outpatient clinicians delivering direct patient care during 2025 (averaging 10 or more encounters per month), 1,140 enrolled in Heidi between January and November 2025. After exclusions for too few preperiod encounters, too few months of Heidi enrollment, or fewer than 30 encounters during their first 3 months of Heidi enrollment, 902 clinicians met study inclusion criteria. Of these, 67.0% used Heidi at least once, whereas 33.0% never used Heidi after enrollment.

Enrolled clinicians were mostly female (60.2%), attending physicians (70.6%), and in primary care (30%). Compared with nonusers, Heidi users were more frequently represented in primary care and less represented in surgical specialties. Users were slightly younger, more often female (63% vs 55%) and more represented among nurse practitioners.

Mean patient age was 61.1 years, and most notes were written for female patients (57.8%) insured through Medicare or commercial insurance (about 44% each), with visits of moderate complexity (46.0%). Most notes were authored by attending physicians (73.6%) followed by nurse practitioners (21.4%). Primary care accounted for the largest share of notes (40.6%).

Heidi users were divided into "high" users, "low" users, and "nonusers" based on their usage during their first 3 months of enrollment, and researchers found that among high users, the immediate reduction in time actively dealing with notes was significantly larger (-2.50 minutes) than both enrolled nonusers (-0.29 minutes) and low users (-0.78 minutes). Low users did not differ significantly from enrolled nonusers in that respect.

High users also experienced a significant immediate increase in note length (160.26 characters), whereas note length decreased among low users (-110.44 characters) and remained unchanged among enrolled nonusers (-10.63 characters).

Study limitations included the fact that observational design is subject to unmeasured confounding and that the investigators could not reliably distinguish between AI-scribed and clinician-authored notes in the postperiod data. "This may dampen trends in documentation metrics and bias our findings if Heidi was preferentially used for certain visit types," they noted. Researchers also could not quantify how many patients declined consent for AI scribe use, though Heidi users reported this rate to be low, nor did they analyze note content and therefore couldn't draw conclusions about AI scribe errors, missing information, readability, or potential harms.