Understanding the Human-Computer Interaction of Ambient AI in Healthcare: A Qualitative Study of Early Provider Experiences

2.1 Study design

This was a qualitative thematic analysis embedded within a larger mixed methods evaluation of ambient AI. The objective of the qualitative component was to examine provider experiences one month after using ambient AI documentation tools and to identify emerging themes that characterize the human-computer interaction between clinicians and technology. This inductive, exploratory approach was selected to generate hypotheses for future study and to capture rich narrative insights that are not easily measured through quantitative methods. The project was reviewed and approved as a Quality Improvement (QI) initiative by the health system and was conducted in alignment with institutional QI guidelines.

2.2 Setting and participants

The study was conducted within the University of Pittsburgh Medical Center (UPMC), a large integrated healthcare and finance delivery system with facilities across Pennsylvania, New York, and Maryland. Providers working in any clinical setting—primary care or specialty—were eligible to participate, provided they completed a pre-implementation survey prior to using ambient AI technology. Eligible participants included physicians and advanced practice providers (APPs).

2.3 Ambient AI implementation context

Providers had access to either of two ambient AI solutions: Abridge or Nuance’s Dragon Ambient eXperience (DAX). The system-wide implementation included a dedicated website with training materials, as well as on-demand technical support and in-person “elbow” support to assist with onboarding. Providers were introduced to the technology via internal communication campaigns in the year leading up to the survey distribution.

2.4 Data collection

One month after implementation, providers who had completed the pre-survey were sent a follow-up survey via Qualtrics. The survey included both structured and open-ended questions. Here we focused exclusively on responses to the following free-text prompt:

Participants could respond anonymously or choose to include their contact information. A reminder email was sent one week after the initial invitation to improve response rates.

2.5 Data analysis

We used Braun and Clarke’s six-phase reflexive thematic analysis framework to analyze the free-text responses: (1) familiarization with the data, (2) generating initial codes, (3) searching for themes, (4) reviewing themes, (5) defining and naming themes, and (6) producing the report [27]. The lead investigator reviewed and coded all 103 responses line-by-line, generating a comprehensive codebook of descriptive and analytic codes. These codes were then grouped into broader axial categories to reflect emerging patterns.

To enhance the rigor and credibility of the coding process, a second physician independently coded a random subset of 12 qualitative responses (approximately 10% of the dataset) using the initial codebook. Each response was assigned a primary thematic category by both coders. Interrater reliability was assessed using Cohen’s Kappa statistic, which yielded a value of κ = 0.80, indicating substantial agreement between coders (z = 6.33, p < 0.001). Discrepancies were resolved through discussion to finalize thematic assignments and support the validity of the coding framework.

Following axial coding, the unit of analysis shifted to individual codes. Codes were systematically organized into thematic categories, with illustrative quotes selected for each. Sentiment analysis was also performed to deepen interpretation.

2.6 Reflexivity and rigor

The lead investigator is a physician with training in clinical informatics, which may influence interpretation of provider feedback in favor of technological optimism. To mitigate potential bias, an independent coder reviewed a subset of responses, and consensus coding was used to finalize theme assignments. An audit trail of codes, category definitions, and thematic evolution was maintained in Excel to ensure transparency and reproducibility.

3.1 Theme 1: Efficiency and workflow impact

Providers reported that ambient AI technology improved documentation efficiency and helped reduce after-hours charting. The ability to automatically generate clinical notes allowed clinicians to focus more on patient care during encounters and spend less time writing notes at the end of the day. Respondents noted that AI-assisted transcription helped streamline their workflow, particularly for post-visit documentation.

Providers also described a reduction in cognitive load, as they no longer had to mentally retain encounter details for later documentation. Instead, they were able to edit and refine AI-generated notes, which they felt was more efficient than starting from a blank note.

However, efficiency gains varied depending on workflow and documentation style. Some clinicians found AI-generated notes helpful only in specific scenarios and reported that any potential time savings were offset by the need for detailed proofreading and corrections.

3.2 Theme 2: Enhanced patient interaction

Another cited benefit of ambient AI technology was its ability to reduce provider distraction during patient encounters, allowing for more engaged, face-to-face communication. Feeling less tethered to the computer enabled more natural conversations and stronger rapport with patients.

This improvement in clinical presence provided a renewed sense of satisfaction in being able to focus fully on their patients.

Moreover, respondents highlighted that ambient AI also improved how well their documentation captured the true scope of their clinical conversations, especially regarding patient education.

However, not all providers viewed the technology as enhancing patient interaction. Some raised concerns about awkwardness in verbalizing exam findings aloud, misgendering, and difficulty interpreting diverse accents—all of which could disrupt rapport. Others noted that some patients were reluctant to consent to AI use, which introduced tension into the encounter.

Overall, while ambient AI was seen as a promising tool for enhancing patient-centered care, its success depended heavily on provider adaptation, documentation accuracy, and patient receptivity.

3.3 Theme 3: Documentation quality and editing burden

While efficiency benefits of ambient AI were acknowledged, concerns about documentation accuracy and the need for extensive editing were expressed. AI-generated notes often contained errors, misinterpretations, or inconsistencies, requiring substantial proofreading and correction before finalizing.

Ambient AI appeared more useful for routine follow-ups than for complex or new encounters. In more nuanced clinical scenarios, AI struggled to accurately capture reasoning or new diagnoses.

A recurring concern was the misrepresentation of clinical details, including incorrect summaries, attribution errors, and misgendering of patients. Pronoun inaccuracies and awkward phrasing were cited as frequent editing burdens.

This need for constant oversight begged the question whether ambient AI truly saved time. A trade-off appeared between the reduced cognitive burden and increased editing workload—particularly when AI-generated notes required significant revision to meet clinical standards.

Despite these concerns, providers expressed optimism that with ongoing refinement, improved contextual understanding, and customization, ambient AI tools could become more accurate and reduce editing time over the long term.

3.4 Theme 4: Learning curve and workflow adaptation

Providers described a varied learning curve in adopting ambient AI documentation tools. While some clinicians adapted quickly and found the technology intuitive, others encountered challenges related to workflow integration, speech pattern adjustment, and limited training resources.

Respondents illustrated a trial-and-error process during which they learned to modify their phrasing, slow down their speech, or restructure their encounters to optimize AI-generated notes. Some clinicians developed personalized strategies to support more accurate transcriptions, such as outlining the assessment and plan point by point or emphasizing key exam findings aloud.

Others highlighted limitations in current functionality, particularly around physical exam capture, complex multi-problem visits, and team-based care scenarios like teaching rounds or procedural encounters.

Providers also noted that while AI integration was not yet seamless, they expected continued improvement and were motivated to seek further training to maximize its utility.

Despite initial challenges, respondents expressed that AI documentation improved with regular use, leading to time savings, reduced stress, and increased satisfaction.

Overall, this theme highlights that successful implementation depends not only on the capabilities of the technology, but also on sufficient training, time for adaptation, and alignment with clinical workflows.

3.5 Theme 5: Specialty-specific performance and suitability

The effectiveness of ambient AI documentation varied across clinical specialties, with some areas benefiting more than others. Psychiatry, oncology, rheumatology, and other specialty care settings encountered limitations in the ability to capture the full depth and nuance of clinical encounters.

In specialties that rely heavily on structured templates, longitudinal assessments, or standardized clinical terminology, AI-generated notes required additional editing due to overgeneralization, factual inaccuracies, or missing clinical reasoning.

In contrast, some specialties with higher patient volumes and more standardized workflows, such as urgent care, reported that ambient AI improved efficiency and reduced after-hours charting.

Specialty-specific adaptations could improve AI usability. Suggestions included tailoring the AI to specialty-specific templates, enhancing terminology recognition, and training models to capture multi-problem visits or longitudinal care plans more effectively.

Overall, while ambient AI showed promise across multiple specialties, its performance was highly dependent on the clinical context and documentation needs. Providers in specialized fields emphasized the need for further customization, accuracy, and refinement to realize the full benefits of the technology.

3.6 Theme 6: Provider skepticism and trust in AI

While providers found ambient AI documentation beneficial, skepticism was expressed regarding its accuracy, reliability, and long-term role in clinical care. Concerns included over-reliance on AI-generated notes, potential for critical omissions, and errors that could undermine trust in the documentation process.

Concerns were raised about the trustworthiness of AI-generated content, noting that statements were sometimes misattributed or that the AI misunderstood the speaker, leading to inaccuracies in clinical documentation.

Unease about how patients perceive AI in clinical visits was voiced. While most patients accepted its use, others were hesitant or declined due to privacy concerns or lack of clarity around data handling and storage.

Despite these reservations, conditional optimism was expressed. Trust in ambient AI could improve with ongoing system refinement, better transparency, and increased user control over how notes are structured and edited.

Overall, provider skepticism stemmed from concerns about accuracy, trust, patient consent, and documentation integrity. While they acknowledged the promise of ambient AI, they emphasized the need for ongoing improvements in reliability, user agency, and clear communication with patients to support widespread adoption.

To synthesize the themes described above, we developed a summary table highlighting reported strengths, weaknesses, and design suggestions derived from provider feedback (Table 1). This table provides a practical, high-level overview of the key insights and considerations for improving ambient AI integration in clinical practice.

4.1 Implications for practice, design, and policy

This study offers practical guidance for health systems and AI developers integrating ambient documentation tools into clinical care (Figure 2). First, implementation should be context-aware: ambient AI performs best in settings with standardized workflows and may require specialty-specific adaptations in fields with more complex documentation needs.

Second, robust training and behavioral support are critical. Successful use depends not only on technical integration but also on helping clinicians adapt their communication and workflows. Onboarding should include visit-type-specific examples, speech guidance, and ongoing troubleshooting resources.

Third, developer attention to usability, transparency, and customization is key. Tools that allow for real-time editing, personalized templates, and feedback mechanisms may improve clinician trust and long-term engagement.

Finally, equity must be prioritized. Several providers reported performance issues with non-English speakers and diverse accents, echoing broader concerns about bias in voice recognition. Proactive testing across populations and safeguards against disparities in documentation quality are essential. Collectively, these findings underscore that successful adoption of ambient AI hinges on human-centered design, equitable implementation, and alignment with real-world clinical practice.

4.2 Strengths and limitations

This study offers several strengths. It draws on a large sample of 103 provider responses across multiple specialties within a real-world, large integrated health system. The use of open-ended survey questions allowed for spontaneous, unfiltered feedback, providing rich insight into the nuanced experiences of frontline users. Our thematic analysis captured both anticipated and emergent themes, contributing to a deeper understanding of human-computer interaction in the context of ambient AI documentation.

However the study has several limitations. First, it was conducted within a single health system, potentially limiting generalizability to other organizational or geographic contexts. Second, the responses reflect experiences only one month after implementation, and thus may not capture long-term adaptation, satisfaction, or disengagement. Third, the qualitative data were drawn from provider self-report via free-text responses, which may underrepresent more extreme views or nuanced use cases that might be better captured through interviews. Finally, our analysis focused exclusively on provider perspectives; future work should incorporate patient viewpoints, especially given the growing role of AI in clinical encounters.

4.3 Future directions and research needs

This study highlights several areas for future inquiry. First, given that our data reflects provider experiences only one month after implementation, longitudinal research is essential to understand how perceptions evolve as clinicians gain proficiency and as ambient AI systems mature. Future work from our team will include analysis of qualitative responses collected at 3-month and 6-month follow-up surveys, allowing for a deeper, time-based exploration of provider adaptation, satisfaction, and perceived value. Tracking engagement and documentation quality across these time points will help determine whether initial challenges resolve, persist, or shift over time.

Second, more research is needed to explore the patient perspective on ambient AI, particularly through qualitative inquiry. While our study focused on provider experiences, understanding how patients perceive ambient documentation tools—especially in terms of trust, comfort, and communication quality—is critical to evaluating their broader impact on the clinical relationship. To date, only limited work has examined this dimension. One recent study by Owens et al. [29]. found that patients generally responded favorably to ambient voice technology, reporting a more personable and focused encounter; however, this study relied on quantitative satisfaction metrics and limited open-label feedback, without deeper qualitative analysis of patient attitudes or concerns.

Future work should incorporate interviews, focus groups, or open-ended survey responses from patients to better capture the relational and ethical implications of AI-mediated documentation. This is particularly important in diverse populations where issues of privacy, consent, and data transparency may shape a patient’s willingness to engage. As ambient AI becomes more embedded in care, a holistic understanding of its influence on both clinician and patient experience will be essential for responsible implementation.