Evaluating ambient artificial intelligence documentation: effects on work efficiency, documentation burden, and patient-centered care

Background and significance

The Health Information Technology for Economic and Clinical Health (HITECH) Act accelerated the widespread adoption of electronic health records (EHRs) and emphasized their meaningful use.^1,2 Although EHRs have benefited many aspects of clinical workflow, their rapid implementation has also significantly increased physicians’ documentation burden.^3–5 This added workload is closely associated with clinician dissatisfaction and has emerged as a major contributor to burnout, which affects not only clinician well-being but also the quality and safety of patient care.^6,7

To reduce this burden, early solutions included digital scribes and voice dictation tools designed to automate routine notetaking.⁸ These systems predominantly relied on natural language processing (NLP) techniques for tasks such as clinical entity extraction, text classification, and summarization.⁹ While promising, these approaches often required manual correction and lacked the capacity to fully capture clinical nuance and context. More recently, ambient listening technologies have marked a significant advancement over traditional tools.^10,11 By integrating large language models (LLMs) and generative artificial intelligence (GenAI), these systems can passively capture patient-clinician conversations and generate tailored, context-aware clinical notes.

Many large health systems in the United States have begun piloting ambient listening tools, prompting a growing body of research evaluating their impact on documentation efficiency, clinician burnout, patient experience, and adoption factors. Studies such as simulated outpatient consultations, the STREAMLINE pilot, and implementations in dermatology clinics have reported reduced documentation time and EHR interaction, along with improvements in note quality and clinician satisfaction.^12–15 Paired surveys further demonstrated reductions in task load and burnout, as well as improvements in usability, work-life integration, and patient engagement among physicians.^16–19 Additionally, qualitative interviews revealed clinician enthusiasm for ambient AI’s potential to enhance workflow and care delivery, while identifying key areas for refinement and better integration into clinical practice.²⁰ Observational studies incorporating patient feedback have also shown general agreement with the use of ambient AI tools in the clinical setting.²¹

Building on this evidence base, our study contributes to the ongoing evaluation of ambient AI as part of a quality improvement pilot project at the University of California, Irvine (UCI) Health system in ambulatory settings. Ambient AI remains at an early stage of adoption in US healthcare, and few evaluations have combined early implementation across multiple specialties, a relatively large sample, extended duration of use, and mixed-method analysis within a single quality improvement framework. Using matched pre- and post-implementation surveys alongside Epic EHR time data, which automatically capture system usage logs, we assess changes in documentation time and length, as well as physician perceptions of ambient AI scribes’ impact on documentation burden, clinical efficiency, and care quality. We hypothesize that integrating ambient AI tools will improve work efficiency, reduce physician workload, and help mitigate burnout, offering critical insights into the scalability and sustainability of these technologies in routine clinical practice.

Materials and methods

Results

A total of 319 physicians participated in the ambient listening tool pilot. Participants included those in primary care, various specialty care disciplines, and medical trainees. For the EHR metrics analysis, only participants who met the inclusion criteria, including a minimum duration of ambient tool usage and sufficient note volume, were included. A subset of enrolled physicians completed both the pre- and post-implementation surveys and were included in the corresponding survey analysis.

Discussion

This study presents a mixed-method evaluation integrating Epic Signal metrics and survey data across multiple specialties in a real-world implementation setting. By analyzing automatically recorded EHR time-efficiency metrics, pre- and post-implementation survey data, and open-ended feedback, we found that ambient AI use was associated with reduced documentation burden, improved perceptions of clinical efficiency, and enhanced patient-centered care. Consistent findings across both self-reported experiences and system-generated metrics reinforce the potential of ambient AI to support clinician workflow and alleviate cognitive workload in everyday practice. While other recent studies have examined ambient AI tools in real-world settings, our evaluation is distinct in integrating multiple specialties, a larger and longer-duration cohort, and a mixed-method approach within a system-wide quality improvement initiative—providing a timely contribution to the emerging evidence base.

Survey results indicated meaningful decreases in mental and physical effort associated with note writing, alongside improved perceptions of system integration and training support. While overall improvements were consistent across provider types, only a few statistically significant differences emerged between primary and specialty care providers. One such difference was in perceptions of implementation and training support, with primary care providers reporting greater improvements. No other EHR system changes or general EHR training occurred during the pilot. Observed improvements in training effectiveness and system usability likely reflect broader shifts in user perception driven by satisfaction with documentation after ambient AI implementation. These changes suggest that clinicians perceived existing support structures more favorably in light of the new tool, rather than benefiting from new EHR training content or direct functional modifications to the EHR. This suggests that onboarding strategies and organizational support may have been better aligned with the workflows of primary care clinicians, highlighting the need for more tailored training approaches to maximize the impact of ambient AI tools across different specialties. Another notable difference was observed in perceptions of alerts that help prevent care-delivery mistakes. This aligns with qualitative feedback emphasizing the need for specialty-specific customization of ambient AI systems. For example, orthopedic surgeons noted differences in documentation needs between a shoulder evaluation and a spine evaluation, while psychiatry providers emphasized language that accurately reflects patient affect, thought process, and nuanced interpersonal observations. Providers in neurology, psychiatry, orthopedics, and surgery noted that current tools often fail to capture the complexity of their documentation, which remains closely tied to nuanced physical exams and specialty-specific narratives. These findings underscore the importance of developing adaptive, specialty-aware AI solutions that can flexibly accommodate diverse clinical workflows and accurately reflect domain-specific terminology and reasoning.²⁹ However, this study was conducted within a single academic health system, which may limit generalizability to other care settings. While these accuracy limitations were qualitatively identified, we did not measure the time or cognitive burden associated with proofreading and editing AI-generated content. Quantifying this burden and its downstream effects on efficiency, documentation workload, and cognitive load represents an important direction for future research. Additionally, potential survey bias and the voluntary nature of participation may have influenced the representativeness of responses. Failure to address these limitations may result in unequal adoption and benefit across specialties, ultimately constraining the broader impact and scalability of ambient AI in clinical practice.

Epic usage data further demonstrated significant reductions in documentation time. These results are consistent with existing studies and further validate the effectiveness of ambient AI in a larger, multispecialty sample.^13,14,30 Documentation length, which slightly increased post-implementation, is an important complement to time metrics. While this trend may reflect additional work or information captured, further research is needed to determine whether increased length contributes to note completeness and clinical clarity. Besides, several limitations warrant consideration. First, because participation was voluntary, our sample may reflect physicians more receptive to AI adoption, which could bias perceptions toward more favorable views of the tool. Second, time spent on notes may not accurately represent active documentation time, as delays related to system performance or clinicians leaving notes open while multitasking may inflate these metrics. Ambient AI use was not mandatory for all encounters; even among physicians with access, usage varied depending on personal preference and clinical context. Notably, post-implementation Epic data captured all encounters, regardless of whether the ambient tool was used, which may have diluted the observed effects. Third, no statistically significant change was detected in same-day encounter closure rates. This outcome may be attributed to the constrained analytic sample, limited to physicians with at least 40 weekly encounters within a single institution, and to the high baseline same-day closure rate of approximately 70%, which reduced the potential for improvement. It may also reflect that certain workflow indicators, such as encounter closure, are less sensitive to documentation tools like ambient AI and more dependent on external operational factors, including clinic schedules and staffing patterns.²⁹ While Epic Signal metrics showed modest time reductions, the self-reported percentage of encounters documented during or immediately after visits did not change significantly, suggesting that efficiency gains may not have shifted the timing of documentation or may reflect limitations in self-reported recall. Additionally, some clinicians may not yet be fully familiar with or trained to leverage the full capabilities of the ambient scribe tool, instead using it in its default configuration without customization or optimization, which may have limited its observed impact.

The findings of this study highlight both the potential benefits and current limitations of ambient AI tools in clinical practice. Beyond improvements in documentation burden and efficiency, thematic analysis revealed strong interest in expanding the role of ambient AI to include tasks such as inbox management, billing, chart summarization, and clinical decision-making.^31,32 Participants also emphasized the need for greater personalization, both at the individual clinician level and across specialties, to ensure the tool aligns with varying documentation styles and workflow demands.

These insights suggest that future research should focus on refining ambient AI systems to better support diverse clinical workflows, enable flexible customization, improve integration with historical medical records, and expand core functionalities to address broader administrative burdens. As ambient AI becomes more deeply integrated into clinical environments, developers must also account for ethical, legal, and regulatory considerations.^33–37 While these safeguards mitigated immediate liability concerns during the pilot, uncertainty around responsibility for AI-generated inaccuracies and transparency of ongoing model updates remains relevant as the technology scales. Ensuring transparency, safeguarding patient privacy, and clarifying liability for AI-generated documentation will be essential to building clinician trust and enabling responsible adoption.³⁸ In parallel, investigating patient perspectives through surveys or interviews will be important to assess perceived communication quality, trust, and acceptance of AI-assisted documentation. Long-term evaluations will be essential to determine whether the benefits of ambient AI, such as improved efficiency and reduced burden, are sustained over time and whether they lead to measurable improvements in clinical outcomes, provider satisfaction, and burnout.

Conclusion

This quality improvement study at UCI Health provides real-world evidence supporting the potential of ambient AI tools to reduce documentation burden, enhance workflow efficiency, and improve patient experiences. By analyzing pre- and post-implementation surveys and EHR usage metrics, our findings demonstrate that ambient AI effectively supports physician productivity and addresses key contributors to burnout. Nevertheless, several areas for further improvement were identified, including specialty-specific customization, integration with historical patient data, and enhanced accuracy of AI-generated content. Future implementation efforts should emphasize thoughtful scalability, adaptability to diverse clinician needs, and careful consideration of clinical, ethical, and legal requirements. Ongoing evaluation and iterative optimization will be crucial for building clinician trust, promoting equity in AI-supported care, and fully leveraging ambient technologies to facilitate patient-centered healthcare.

Author contributions

Yawen Guo (Conceptualization, Data curation, Formal analysis, Investigation, Methodology, Project administration, Validation, Writing—original draft, Writing—review & editing), Jiayuan Wang (Data curation, Methodology, Writing—review & editing), Di Hu (Conceptualization, Methodology, Writing—review & editing), Steven Tam (Conceptualization, Project administration, Supervision, Writing—review & editing), Charles Gilman (Project administration, Resources, Writing—review & editing), Emilie Chow (Supervision, Writing—review & editing), Danielle Perret (Supervision, Writing—review & editing), and Deepti Pandita (Supervision, Writing—review & editing), and Kai Zheng(Conceptualization, Supervision, Writing—review & editing)

Funding

This research received no specific grant from any funding agency in the public, commercial or not-for-profit sectors.

Conflicts of interest

The authors have no competing interests to declare.

Data availability

De-identified survey data supporting the findings of this study are available from the corresponding author upon reasonable request.