This implementation evaluation was a qualitative descriptive study using thematic analysis of physician semistructured interviews. This study was guided by the domains of the PRISM [11] and STS frameworks [12]. We selected PRISM to examine the broader organizational and implementation context. We selected STS to complement PRISM by capturing AI-specific sociotechnical factors related to workflow, fit, transparency, and trust. Together, the frameworks supported exploration of interactions among the technology, its users, the clinical setting, and the broader implementation context. Both frameworks informed development of the interview guide, deductive codes in the initial codebook, and interpretation and reporting of the findings. The frameworks were applied complementarily rather than merged into a new framework. Figure 1 depicts the complementary roles of PRISM and STS, presenting the domains of each framework, and how they informed the study. This paper follows the Standards for Reporting Implementation studies (StaRI) [15].

Figure showing the complementary roles of the 2 frameworks, PRISM and STS, and how they informed the study. The figure presents the distinct and overlapping domains of the 2 frameworks and shows how both informed development of the interview guide, deductive codes in the initial codebook, and interpretation and reporting of the findings.

This study took place at a metropolitan university health system with a 646-bed quaternary care hospital and ambulatory clinics comprised of 17 primary care sites and 153 specialty clinics. This health system serves as a referral center for patients across a 33-county region spanning 65,000 square miles. AI scribe services were first implemented at our health system in July 2024 with an initial pilot program with 31 physicians over 2 months. The tool was subsequently deployed to all primary care and specialty care physicians in 4 waves from October 2024 through January 2025. Additional ambulatory physicians were onboarded based on request after January 2025.

The decision to use the AI scribe was at the discretion of the physician. Physicians who elected to use the tool were required to complete an 11-minute training video prior to activation. The training covered application workflow, regulatory and compliance considerations, and best practices. The AI scribe vendor was Abridge, which is a tool that has been evaluated in prior studies [3,6]. Physicians accessed Abridge via the Epic electronic health record mobile app. An iPhone (Apple Inc.) or iPad was required for use. Where personal devices were unavailable, a limited number of shared iPads were available in clinic. The AI scribe used synchronous ambient listening to summarize conversations and structure them in real-time within the electronic health record.

We conducted one-on-one semistructured interviews with physicians between May and September 2025. Participants had to be aged 18 years and older and with English language preference. Participants were limited to attending physicians working at an ambulatory clinic affiliated with UC Davis Health where AI scribe was available for use. Eligibility was limited to attending physicians because they were the only clinicians permitted to use AI scribe at the time of study initiation.

We recruited participants via email. We used maximum variation sampling [16] to identify attending physicians with varied levels of AI scribe use; sampling was not designed to achieve balance across demographic characteristics. We strived to include participants who used AI scribe never, minimally (1 to 10 notes per week), and more frequently (exceeding 10 notes per week). No formal a priori sample size estimate was conducted. Sampling continued until we reached thematic saturation, which was defined as the point at which additional interviews generated no new substantive codes or concepts relevant to the study aim [17]. We assessed for saturation iteratively during analytic team meetings. Thematic saturation was achieved after 22 interviews; subsequent interviews reinforced and added depth to existing findings.

The interviewer (AK) used a question guide that solicited participants’ general perceptions about the AI scribe and questions aligned with the PRISM and STS domains (guide provided as Multimedia Appendix 1). Interviews were immediately followed by asking the participant to verbally self-report their characteristics, including age, gender, ethnicity, race, clinical practice location, years of experience since training, and comfort using AI scribe. Interviews were audio recorded and sent to a third-party company for transcription. Final transcripts were reviewed for accuracy by the interviewer.

Data were analyzed iteratively using a constant comparative approach and a combination of deductive and inductive analysis. The analysis team consisted of the interviewer plus 3 additional researchers: an attending physician and medical educator who leads digital health research initiatives for the health care system (JLR), a health information technology educator who leads clinical implementation of digital health programs for the health care system (MJ), and an undergraduate student whose scholarly focus is health technology innovations (SH). We developed an initial codebook of a priori codes related to the PRISM and STS domains.

The research team members independently conducted memo writing and coding of the initial 3 transcripts using the PRISM- and STS-based a priori codes while identifying emergent topics. The team met every other week to ensure consensus on the application of codes, refine dimensions of existing codes, add new codes as needed, develop categories, and identify theoretical direction. Disagreements among the team members were resolved through discussions to understand viewpoints, deliberate, and achieve consensus. We iteratively adapted the interview guide to explore emerging topics. We resumed independent memo writing and coding followed by team meetings; this iterative process was repeated with every 2 to 5 transcripts. Most emerging content fit within existing a priori codes, whose dimensions we refined accordingly. One inductive code (trainee curriculum) was added, and previously coded transcripts were re-reviewed for relevant content. We identified linkages and patterns between the codes to develop analytic themes and subthemes. This process continued until we achieved thematic saturation, whereby data coalesced around similar themes. We mapped each subtheme to the corresponding domains of the PRISM and STS frameworks. We used ATLAS.ti to organize and store coding and data analysis [18]. Multimedia Appendix 2 contains the final codebook.

The multidisciplinary analysis team included individuals with backgrounds in public health, medical education, medicine, and health technology. A total of 2 team members (MJ and JLR) held digital health leadership roles; one (JLR) who led digital health research and practiced as an attending physician exclusively in the inpatient setting, while the other led health information technology clinical implementation. These roles could have influenced the interpretations of AI scribe implementation, likely in a more favorable manner. To mitigate this potential bias, interviews were conducted by a nonphysician who did not hold such a leadership role and had no prior working relationship with participants. Transcripts were coded independently by members of the 4-person analysis team (AK, MJ, SH, and JLR), divergent perspectives were examined, and differences were resolved through team discussion.

The University of California Davis Institutional Review Board approved this study (#2307389) with a waiver for written documentation of consent to satisfy ethical reporting standards. Participants provided verbal informed consent. The third-party company deidentified the data at the time of transcription. Interviewees did not receive compensation for participation in this study.

A total of 25 interviews were completed with physicians. The duration of each interview was approximately 30 to 45 minutes. Representation of AI scribing usage was relatively equal across groups defined by AI scribe use: never (n=7), minimally (n=9), and more frequently (n=9). Interview participant characteristics are described in Table 1. The following sections display the results of our qualitative analysis organized by theme. Overall, 5 themes and 15 subthemes emerged and are described below. Table 2 presents representative quotes supporting each subtheme, with corresponding PRISM and STS domain classification.

Characteristics of attending physicians participating in this qualitative implementation evaluation, stratified by frequency of AI scribe use. Characteristics were self-reported. Participants with prior experience using AI scribe (n=18) rated their current comfort, whereas those without prior experience (n=7) rated their anticipated comfort with using the tool.

A common issue raised by most participants was that AI scribe inadequately integrated into their existing workflows. Several emphasized that the tool’s output misaligned with their established templates. They highlighted that using AI scribe required them to make significant edits to fit their preferences. Many nonusers said their workflow for writing notes was already efficient and AI scribe would provide no benefit. It was also noted that AI scribe did not include the patient’s past medical history, and since a patient’s past medical history is often not discussed out loud during the clinic visit, the note omitted this critical information. Physicians ubiquitously shared how AI scribe was not helpful in documenting the physical exam, as that would require extensive audible explanation and would be cumbersome and potentially uncomfortable for the patient. Nonuser and low-users alike expressed concern over AI scribe generating notes that were bloated and lengthy.

In contrast, several physicians shared how they had adapted their workflows to align with the tool’s functionality. These adopters described developing workarounds and hybrid documentation strategies to generate notes. For example, a few physicians explained how they changed the way they spoke, such as saying the past medical history out loud to the AI scribe before walking into the clinic room. One physician adapted their in-room communication style, using verbal signposting (“now we’re going to talk about…”) to help the tool segment note content to result in a more organized note. Many others tailored their workflows using note templates and dot phrases to compensate for sections not populated by, or poorly generated by the AI scribe.

Nonusers frequently expressed concern and resentment about being early or experimental users of the AI scribe. They frequently framed this early implementation as asking physicians to burden the task of beta testing. Some shared reluctance to adopt a tool not yet ready for real-world use; others rejected the notion to use the tool in its current state. A few perceived use of the tool to be uncompensated labor to improve or train the product. Several of these individuals described openness to revisiting adoption later, once its performance improved substantially.

Most physician adopters described AI scribe as being user-friendly and straightforward. Many highlighted that the mandatory training video felt unnecessary. Some shared the video added little value; they learned more by hands-on “trial and error.” Some could not recall specific content from the video; others did not even recall watching it. While retainment of training seemed limited, physicians still used the AI scribe because of its ease of use and the ability to learn in real time. Of note, video content included details and tips about the tool that would have addressed or mitigated many of the expressed issues with the tool. Lack of attention to details in the training materials not only impacted usability but also influenced physicians’ awareness about policies and regulations related to AI scribe. Despite stating they had limited knowledge on policies and regulations, these same individuals still used the program.

Medical-legal risk emerged as a barrier to use for some physicians. A few shared concerns that inaccuracies in a signed note could create liability. One physician noted that these medical-legal risks could be enhanced from patients perceiving the encounter to be “impersonal” because the physician had to speak into their device to use the tool. At the same time, one participant shared that physicians’ worries about medical-legal consequences likely exceed patients’ concerns. Regardless, physicians responded to their medical-legal concerns by being very vigilant about proofreading and editing practices. This perceived need for vigilant review of notes, however, limited adoption.

Physicians shared varying levels of trust in using AI scribe, and AI in general, ranging from complete skepticism to near-complete trust. Most articulated a “trust but verify” approach, explaining that AI scribe captured details of the clinical encounter thoroughly and accurately, but that it was the physician’s responsibility to review, detect, and correct inaccuracies or “hallucinations.” While most users spoke about the importance of professional accountability, others acknowledged that their trust level was likely higher than it should be. Furthermore, a few physicians demonstrated implicit trust. For example, one physician assumed the tool accurately captured encounters conducted in languages other than English, including languages they did not speak and thus could not independently verify.

Several participants described mistrust as a barrier to adoption. They shared how the potential for inaccurate documentation necessitated time-intensive review of the generated notes, which they perceived diminished potential efficiency gains. One physician, however, noted that their trust increased over time through use and experience with the tool, resulting in a deeper understanding of the tool’s limitations and the implementation of strategies to mitigate them, such as hybrid documentation approaches.

Beyond individual trust, a few physicians expressed how AI scribe led to their mistrust in other care team members. Multiple physicians recounted examples of reading notes from other physicians that had inaccuracies that were overlooked and not corrected. One physician shared an example of receiving a note with recommendations in the assessment and plan that were outside the signing physician’s scope of practice. Specifically, a primary care physician included recommendations that would typically require surgical subspecialist evaluation and management, suggesting they were AI-generated and not appropriately edited. Similarly, another expressed worries that AI scribes used by medical students may increase the risk of inaccurate information, as supervising physicians might assume the information was accurately obtained when signing the student’s AI-generated note. Some physicians raised concerns about a “telephone effect,” whereby information inaccurately transcribed could be propagated in the electronic health record. These concerns intersect with the perceived medical-legal risk described in Theme 2.

All interviewees expressed that use of AI in health care will become ubiquitous. Multiple nonadopters stated that they will likely have to eventually use AI for clinical care, including AI scribes. The perceived main driver of widespread integration was efficiency in health care delivery. Many framed this topic of efficiency as an organizational pressure to improve productivity and reduce cost. Some contrasted AI scribes with human scribes, sharing that training human scribes required substantial time and effort and may therefore be more costly compared with AI-supported documentation. While most spoke about AI scribe improving efficiency by reducing documentation burden, a few physicians questioned its value. Those with this divergent view explained that AI scribe slowed their workflow and reduced their efficiency. Others questioned whether efficiency gains offset high licensing cost. Aside from cost, another driver that was less commonly raised related to AI scribe reducing after-hours documentation, which was perceived to align with health system goals to address burnout.

Several physicians indicated a desire for AI scribe to expand its capabilities into other clinical workflows. Expressed interests included AI-assisted patient scheduling, medication refills, interpretation of laboratory and exam results, insurance authorizations, and responses to patient portal messages. This desire for expansion of AI applications beyond scribing was expressed both by AI scribe adopters and nonadopters.

However, physicians expressed mixed views regarding expanding AI scribe use to trainees. Some thought trainees should not be exposed to these tools, most commonly citing the value of writing clinical notes as a critical experiential learning activity that supports development of clinical reasoning skills. While most interviewees believed AI scribes would diminish learning opportunities, there was great variability in how these concerns applied across training levels. In general, apprehension lessened as the trainee stage advanced, yet the threshold for introducing these tools ranged widely.

In contrast, a minority supported early exposure. These individuals thought it was important for medical students to become familiar with AI scribes early in their education due to the inevitable rising presence of AI in health care. A few physicians described potential advantages, suggesting that AI-generated notes could show trainees how to organize written information, be a source of clinical knowledge and clinical reasoning, and allow more time for teaching by reducing note-writing burden during encounters.

Physicians described several mechanisms through which AI scribe was simultaneously perceived as both advantageous and disadvantageous. Productivity was one of the most discussed topics. Some high users perceived AI scribe as advantageous because it reduced time spent writing notes, allowing them to see more patients and thus increase revenue generation. However, a few physicians worried that improved efficiency would lead to pressure to see more patients within the same timeframe, resulting in further burnout. This concern was further amplified by the perception that integration of AI scribes into clinical care is inevitable, as described in Theme 4. This difference in framing—increasing capacity versus intensifying workload—influenced physician adoption. As noted previously, others questioned or disagreed with the idea that AI scribe improved efficiency, noting the need to do extensive edits to AI-generated notes.

Physicians also had contrasting perceptions about the impacts of AI scribe use on patient-provider communication. Some expressed how the tool allowed them to be more attentive and make more eye contact during patient encounters. Specifically, they noted how standard encounters required them to constantly look at the computer screen, whereas AI scribe removed that need to continuously type. Others reported improved conversational flow and more time spent in the room as a result of AI scribe. Conversely, other physicians shared concerns that speaking aloud to a device in front of the patient could be perceived as impersonal or distracting. One physician noted that using AI scribe required them to speak differently, causing the clinical interactions to be clunky and less patient-centered.

Finally, physicians shared contrasting ways in which AI scribe impacted their cognitive load during clinical encounters. Many highlighted a positive impact, sharing how editing the AI-generated note required less cognitive effort than independently composing a note. Some physicians noted that the tool reduced the need to remember details or type information in real time, which made their experience in the patient room feel mentally less taxing or stressful. However, a few nonusers commented on the value of note-writing, especially the care plan (ie, assessment and plan), sharing how that process helped them to reflect and synthesize information. They raised concerns that reliance on AI could negatively impact how physicians process clinical information, ultimately worsening the quality of patient care. This concept overlaps with the concern that emerged about trainee learning described in Theme 4.

Our qualitative implementation evaluation of an AI scribe for physicians in the ambulatory setting identified multilevel contextual factors influencing successful implementation, including factors not previously reported. The 5 themes that emerged underscore the complex interplay between workflow integration, tool usability, trust in AI, and provider- and system-level drivers. Using PRISM and STS frameworks, our findings increased understanding of multilevel influences that can be leveraged or addressed to optimize outcomes of interest. As highlighted in this study, what defines a relevant and meaningful outcome may differ across users, particularly with respect to efficiency, which was variably experienced as increased capacity to see more patients—viewed as a benefit by some and a downside by others.

Prior qualitative research has examined AI scribes in the ambulatory setting, including an implementation evaluation guided by RE-AIM/PRISM [7] and a qualitative content analysis characterizing edits to AI-generated notes [19]. Findings from these studies align with our current findings, including facilitators and barriers around topics of AI-generated note construction, accuracy, workload, and physician-patient interactions. Other quantitative and qualitative studies have similarly identified workflow integration, usability, documentation quality, and trust as influences on AI scribe adoption and use [9,10,20,21]. While these prior studies have characterized physician perceptions and experiences with AI scribe, including note quality and physician edits to AI-generated documentation, our study extends the literature by shifting the focus from individual user experiences to the broader organizational and sociotechnical context in which AI scribe implementation occurs.

Our study contributed new knowledge by combining an implementation and sociotechnical approach to explore how physician experiences map to key PRISM domains—particularly organizational characteristics—and STS domains—particularly transparency and trust. While Shah et al [7] similarly applied PRISM, their sample disproportionately represented high users and did not incorporate the STS framework. By more evenly sampling physicians across usage levels, including nonuse, our study captured a broader range of experiences and provided additional insight into contextual factors impeding adoption and implementation. Beyond individual user experiences, our findings highlighted the interplay between perceived health system priorities and physician experiences, beliefs, and values during implementation. Our findings suggest that the perceived organizational emphasis on productivity and cost not only shapes AI scribe implementation but also shapes how physicians interpret external pressures to adopt these tools, contributing to a sense of inevitability that their individual agency in AI scribe adoption decisions may decline over time. Within this context, reluctance to serve as early adopters underscores a misalignment between organizational momentum for implementation and physician readiness, with some physicians explicitly resisting the perceived role as “beta testers” responsible for using a suboptimal product to contribute data to refine the technology. This resistance appeared to be driven, in part, by skepticism about whether health system motivations to expand AI scribe use were truly aligned with physician well-being. Although the tool is framed as a strategy to reduce documentation burden and thus burnout, physicians expressed mistrust in the downstream outcomes, wondering if efficiency gains would instead be leveraged to increase patient volume—a mechanism that could worsen burnout.

Trust again emerged as a construct reflecting both physicians’ perceptions of the technology itself and concerns about other physicians’ overconfidence in AI-generated documentation. Publications have previously described the influence of trust on adoption of AI scribes, noting that trust functions both as a barrier to adoption when lacking and as a facilitator of adoption when present [5,10]. Research examining the mechanisms by which trust influences implementation and impact, however, is limited. Regarding implementation, one qualitative study presented findings on clinician perceptions of AI scribing implementation, describing a concept of “safe with caution” to emphasize the importance of critically reviewing AI outputs rather than accepting them blindly [22]. This concept is very similar to the “trust but verify” approach that emerged in our study, which balanced perceived efficiency benefits with professional accountability. Participants shared that without this accountability—without verification of the AI-generated documentation—notes sometimes contained inaccurate information or content outside the physicians’ scope of practice. This over-trust without verification consequently impacted physicians’ trust in other care team members. Prior studies have reported AI scribing errors, including omissions, additions, and incorrect facts [8,9,23]. Our findings additionally raise the possibility that AI-generated documentation may include content outside a clinician’s scope of practice.

Prior research has identified workflow integration as a key determinant of AI scribe adoption; our findings extend this work by suggesting that integration is not simply about alignment with existing workflows, but rather about physician adaptability and flexibility in modifying those workflows. Physicians with high adoption described actively adapting their documentation templates to integrate with the tool’s output. A commonly used approach was development of hybrid documentation strategies that combined AI-generated and manually authored content. This approach allowed physicians to leverage the efficiency of automated documentation while maintaining control over elements of the note, including clinically relevant details, accuracy, and care plans within their scope of practice. In contrast, some were unwilling to modify established workflows, and others appeared unaware of alternative approaches. Thus, dissemination of successful use strategies (eg, stating out loud the patient’s past medical history, pausing use of the tool during the physical exam, or verbal signposting) may represent a modifiable implementation target with potentially high impact. These findings align with prior research, which has identified AI scribing adoption barriers to include mismatches between clinicians’ documentation styles and the tools’ inflexible formatting and lack of customization [21]. While the body of literature on implementation of AI scribes in health care is rapidly growing [5], clinicians and health care systems lack effective strategies to overcome these workflow integration barriers. Our study helps address this gap by identifying specific adaptive strategies used by physicians.

Our study findings also extend prior research by providing a deeper understanding around the tensions between perceived benefits and risks of AI scribes around the physician-patient relationship. Consistent with our findings, prior research suggests AI scribes result in physicians and allied health professionals spending more time interacting with the patient and less time on the computer [24], improving the therapeutic alliance with more eye contact, attention, and engagement [20]. However, also consistent with our findings, concerns remain about unintended harms of eroding patient trust and loss of empathy [25]. Our study builds on this literature by elucidating perceived mechanisms underlying these perceptions. Physicians in our study described how AI scribe shaped their behaviors, such as “clunky” speech or using their smartphone in the patient room, which in turn influenced how natural or patient-centered the interaction felt. These findings suggest that how AI scribes influence the physician-patient relationship is possibly mediated not only by time allocation, but also by altered communication dynamics. As such, optimizing implementation of AI scribes requires attention to these interactional mechanisms.

This study is not without limitations. First, it was conducted at a single university-based health system, which may limit generalizability to other settings. Additionally, because AI scribe products vary in workflow integration, output, and accuracy, findings from our study examining Abridge may not be transferable to other AI scribe products. We sought to capture varied perspectives through maximum variation sampling targeting AI scribe use; additional research is needed that examines implementation across diverse healthcare settings, including community-based practices. Second, maximum variation sampling focused on AI scribe utilization rather than demographic characteristics; gender and ethnicity were unevenly distributed across utilization groups. These imbalances may have influenced our findings. Future research should examine associations between demographic characteristics and AI scribe use and experiences. Third, because trainees currently do not use AI scribe within our health system, interview participants were limited to attending physicians. Perspectives of trainees and other clinical staff were not included. Consequently, findings regarding trainee use reflect attending physicians’ perceptions rather than direct experiences. As AI scribe use expands to trainees, future implementation evaluations among this population are needed. Fourth, member checking was not conducted after the transcripts were synthesized. Fifth, the self-reported experiences may be subject to recall or social desirability bias. However, the use of thematic saturation and maximum variation sampling for AI scribe utilization levels strengthened the trustworthiness of our findings. Finally, although sampling continued until thematic saturation was achieved, less common perspectives may have been missed.

In summary, implementation of AI scribes in ambulatory care is shaped by a complex set of sociotechnical and contextual factors. Physicians’ adoption of AI scribe was influenced by workflow alignment, trust, and preexisting beliefs about intentions and downstream consequences. Physicians varied in whether they perceived the tool’s effects on efficiency, workload, and patient-centered communication to be beneficial or detrimental. These findings illustrate how clinical context affects the implementation of AI scribe innovations.