AI Notetaking in Psychotherapy: Consent, Efficiency, and Clinical Outcomes from a Large-Scale Evaluation

AI Notetaking in Psychotherapy: Consent, Efficiency, and Clinical Outcomes from a Large-Scale Evaluation | Research Square window.SnipcartSettings = { analytics: { enabled: false } }; (function() { var accessVector = localStorage.getItem('access_vector') || ''; window.dataLayer = window.dataLayer || []; if (accessVector) { window.dataLayer.push({ user: { profile: { profileInfo: { snid: accessVector } } } }); } })(); (function(w,d,s,l,i){w[l]=w[l]||[];w[l].push({'gtm.start':new Date().getTime(),event:'gtm.js'});var f=d.getElementsByTagName(s)[0],j=d.createElement(s),dl=l!='dataLayer'?'&l='+l:'';j.async=true;j.src='https://www.googletagmanager.com/gtm.js?id='+i+dl;f.parentNode.insertBefore(j,f);})(window,document,'script','dataLayer','GTM-K279D39R'); Browse Preprints In Review Journals COVID-19 Preprints AJE Video Bytes Research Tools Research Promotion AJE Professional Editing AJE Rubriq About Preprint Platform In Review Editorial Policies Our Team Advisory Board Help Center Sign In Submit a Preprint Cite Share Download PDF Article AI Notetaking in Psychotherapy: Consent, Efficiency, and Clinical Outcomes from a Large-Scale Evaluation Emily Ward, Matt Hawrilenko, Millard Brown, Adam Chekroud This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-9589129/v1 This work is licensed under a CC BY 4.0 License Status: Posted Version 1 posted You are reading this latest preprint version Abstract Clinical documentation is a major burden, particularly in psychotherapy where notes must meet clinical, compliance, and billing requirements. Across 850,000 appointments within a nationwide digital mental health benefit, we evaluated the feasibility, adoption, efficiency, and clinical impact of AI-powered notetaking by making these tools available to over 8,700 providers and 189,000 patients. Mutual patient-provider consent was required for session recording and AI summarization, and nearly half of providers and 66.2% of patients consenting regularly. Although providers were more reticent, adoption rates increased steadily over the study period. Those providers who used AI-generated notes completed documentation 25% faster, submitted notes 8.2 hours sooner, and had reduced after-hours work compared to sessions without AI. Integration into workflows was common, with AI-generated text incorporated in over 73% of submitted notes. Importantly, continuity of care was slightly higher for adopters and clinical outcomes—including depression and anxiety symptom trajectories—were equivalent between AI adopters and non-adopters, indicating no adverse impact on patient care. These findings demonstrate that AI-powered notetaking can be scaled in psychotherapy, and its use is associated with improved provider efficiency and quality of life without impacting clinical effectiveness. Scientific community and society/Business and industry Health sciences/Health care Health sciences/Medical research Biological sciences/Psychology Social science/Psychology Figures Figure 1 Figure 2 Figure 3 Figure 4 Figure 5 Figure 6 Key Takeaways AI-notetaking can be scaled successfully in psychotherapy, easing provider burden while maintaining patient outcomes. AI-notetaking was associated with reduced documentation burden and increased efficiency: clinical notes were completed 25% faster, submitted 8.2 hours sooner and with fewer after-hours signings. The efficiency gains associated with AI-notetaking did not hurt clinical quality: clinical outcomes (continuity of care, depression/anxiety improvement) were just as strong or better as without AI. Initial provider skepticism about data handling and trust was overcome through mutual consent and transparent security measures, leading to steadily growing acceptance. Introduction Provider capacity and quality of engagement are key drivers of patient outcomes 1,2 , yet documentation burden can erode both. Clinical documentation is among the most significant drivers of clinician burnout 3 . In psychotherapy, notes require a nuanced narrative yet must also meet billing, compliance, and clinical communication requirements. These demands consume substantial provider time, slow the administrative and billing cycle, and can often spill into late-night “catch-up” hours 4 . AI-powered notetaking has emerged as a potential solution to improve provider quality of life. In this approach, therapy sessions are transcribed and summarized using Large Language Models (LLMs). Such “ambient scribes” have been employed in primary care 5–7 and other settings, demonstrating time-saving benefits and a reduction in the cognitive load associated with documentation. Large-scale implementations have shown significant reductions in time spent in notes per appointment 5 and time spent on the electronic health records (EHR) after hours, although some studies found no significant change in after-hours work 8 . Both physicians and patients have reported positive experiences, with physicians feeling more present during visits and patients perceiving that their doctor spent less time looking at a computer. However, such ambient scribes are not always directly integrated into EHR, making it difficult to track the incorporation of AI-generated text which offers a novel and objective approach to measuring practice integration. While existing research provides a strong foundation for the operational benefits of ambient AI scribes, critical gaps remain, particularly within the specialized context of psychotherapy. One study found that mental health providers had the highest adoption rates of AI notetaking (used in approximately 50% of visits) 6 , suggesting a significant need and potential benefit in this population. Most prior work has largely centered on clinician experience and workflow efficiency, but in psychotherapy, clinical notes are less structured than those in primary care, so understanding the gain (if any) to provider efficiency is important, especially given the mixed results from EHR data. Psychotherapy depends on privacy and trust: patients must feel safe enough to disclose information that makes them vulnerable, and session recordings or AI notetaking tools that reduce this willingness may pose a significant hurdle. Moreover, a potential fear is that automated notetaking may result in less provider engagement with a patient, leading to worse outcomes. Direct clinical impact of AI notetaking adoption, including continuity of care and patient symptom improvement, have not been investigated. The Current Study Session recording and downstream AI documentation tools were available in over 850,000 total appointments, across over 8,700 providers and 189,000 patients. We evaluated whether and how such AI tools would be integrated into psychotherapy at such a scale, while addressing both operational efficiency and clinical quality. Specifically, our goals were to: Measure adoption in a large, nationwide provider network where mutual patient-provider consent was required for session recording and AI-notetaking. Assess workflow impact using objective EHR-linked note completion and submission times, capturing after-hours documentation rather than self-report. Examine integration into practice through tracking of AI-generated text incorporation into clinical documentation. Evaluate clinical quality and patient outcomes , including continuity of care and symptom improvement (PHQ-9, GAD-7 trajectories) to ensure recording and AI tools had no negative impact. Methods Study design This retrospective cohort study was approved by the Yale Institutional Review Board (IRB protocol ID: 2000029276) and classified as not involving research with human subjects. That is, there was no contact with human subjects by any research team member, data were collected for another purpose (here, transcripts and clinical notes were part of routine care), and there was a data use agreement in place for a limited data set. All data were anonymized prior to analysis. Program Design Data came from an employer-sponsored digital mental health benefit (Spring Health) that includes free or low-cost access to a number of psychotherapy or medication management sessions, care navigation, and digital self-help mental wellness resources. Additional sessions were covered as an in-network benefit via the health plan. Care is coordinated through a centralized data system that integrates patient data, assessment results, and patient scheduling. Treatment was primarily delivered through video conferencing and provider notes were submitted using the benefit platform. Approximately 70% of psychotherapy sessions were billed as 60-minute appointments (CPT 90837), 10% as diagnostic evaluations (90791), and the rest a mix of 45-minute sessions (90834), couples therapy (90847) and no shows. The benefit included a digital mental health assessment tool for initial screening and outcome monitoring. For their initial screening, individuals completed a series of self-report questionnaires to identify common mental health difficulties (such as stress, anxiety, eating, or substance use issues) and completed the Patient Health Questionnaire 9-item scale (PHQ-9) 9,10 for depression. Based on their responses, they may also complete the Generalized Anxiety Disorder 7-item scale (GAD-7) 11 and/or targeted questionnaires for specific issues. (e.g., patients indicating any issues related to trauma were given the PC-PTSD questionnaire). For longitudinal outcome monitoring, participants were reminded every two to four weeks to complete follow-up questionnaires, consisting of PHQ-9 and/or GAD-7 assessments for most conditions. AI Strategy The benefit program leverages the NIST AI Risk Management Framework as a foundational element of AI governance strategy and maintains an internal AI Governance Board to oversee the responsible development, deployment, and research of AI technologies. This framework has guided the creation of internal policies, the establishment of the governance board itself, and the development of an AI security program, including risk-based testing protocols and embedding safety layers into our products. All AI tools in this study are HIPPA compliant, and implemented within a SOC 2 attested and HITRUST certified platform, leveraging existing administrative and technical safeguards to protect data in alignment with the necessary requirements. In addition, at the time of this study, the program was aligning practices with the ISO/IEC 42001 standard and to pursue formal certification and build a comprehensive program grounded in both NIST and ISO guidance. Participants Participants were employees and dependents eligible for the benefit from March 1, 2025, and January 31, 2026. Patients were located in the U.S. and 18 years or older. When analyzing consent and adoption, we included anyone who had started treatment within the study period. To measure clinical outcomes, additional criteria were applied: patients were included in the analyses if they had a baseline assessment within 30 days prior to starting treatment, if they screened positive for depression and/or anxiety (PHQ-9>=10 or GAD-7>=10) and if they had an additional assessment after treatment began. Participants must have started therapy during the study period and remained eligible for the benefit during the study period. Assessments taken up to the end of the study period were included (note that patients starting care later in the study period had less opportunity for improvement). Patients did not receive monetary reward. Providers Providers included licensed clinical psychologists, mental health counselors, licensed clinical social workers, psychiatrists, and psychiatric nurse practitioners. Approximately 42% of providers were LMHC, 40% were LICSW, 13% were LMFT, and 5% were clinical psychologists. Providers had 13.3 years of licensed experience and their average tenure in the benefit network was 25.9 months. Recording Eligibility & Consent Procedures Session recording and AI-powered notetaking were only available for specific patients engaging in tele-therapy sessions: the patient’s employer or health plan must have opted into session recording, and the patient could not be located in Vermont. If a patient or session did not meet eligibility criteria, no consent prompt appeared. Because audio recording was required for AI-powered notetaking, both patients and providers were prompted to give explicit consent to recording at the beginning of each session (Figure 1). They were also given information about the downstream processes that would use the session recordings: patients were informed that their session would be transcribed, that their provider would receive detailed notes about the session, and that they—the patient—would receive a summary and list of takeaways (Figure 1, right). Providers were told that the session would be recorded and could be used to speed their own clinical notes. They were also informed that the audio recording would be deleted immediately after the session was summarized (Figure 1, left). Thus, consent included session recording and downstream summarization, etc. Both parties were provided with links to the terms, privacy policy, and help center for any other questions. Either party could pause or stop the recording at any point. Providers did not receive any incentives to record sessions or use AI-generated notes. Intervention Transcription: An overview of the data processing pipeline is shown in Figure 2. Sessions were recorded using an open-source cloud platform for collecting audio data. At the end of a session, the audio file was uploaded to an AWS S3 bucket. The upload triggered the core backend to link the recording to session data and to begin transcription. The audio file was transcribed (using Amazon Transcribe), and the final transcript was used to produce downstream AI processes, stripped of all personally identifiable information (PII) and protected health information (PHI), and then stored for quality assurance . The audio file was deleted and the raw transcripts were available to providers for 72 hours then deleted. Only fully de-identified transcripts collected within this routine care pipeline were available for further research, and the data was accessible only to authorized staff through a documented approval process. Furthermore, these transcripts were not used to train or fine tune new AI models. Summarization: Session summaries were generated using an LLM and a standardized prompting framework for transcript-based therapy documentation. The prompt constrained the model to produce de-identified, professionally worded summaries grounded only in information explicitly present in the session transcript. It further instructed the model to omit irrelevant procedural details, such as technical difficulties, and to avoid redundancy across sections. Safeguards: Summaries were limited to behavioral observations, diagnostic terminology, and absolute language that was directly supported by transcript content. Editing workflow: AI-generated transcripts and notes were made available to providers during clinical documentation. The documentation included two required fields (Figure 3): “Presenting Problem,” capturing the client’s broader clinical concerns and therapeutic context and a “Session Focus” section, comprising current status and changes since the previous session, therapeutic themes and interventions, and safety/risk content. Thus, the AI-powered summaries were automatically generated for each of these two sections and presented to providers alongside the text-field for their clinical notes. Providers could incorporate these AI-generated notes directly via a button in the platform that allowed them to copy-to-clipboard and whenever the button was clocked, the platform logged the action. However, providers could also copy-and-paste the text without using the platform copy button (using the usual keyboard shortcuts). In this case, the platform could not track whether the AI-powered notes were integrated into clinical notes. Thus, the copy-to-clipboard button action may be a conservative measure of integration into workflows. Providers could further review and edit the notes with their own observations before completing and submitting the note, but the extent to which AI-generated text was used minimally, partially, or substantially was not tracked. Measures Note completion time corresponded to time elapsed (in minutes) from when the clinical note page was opened in the provider’s browser until it was closed. Time to note submission measured the time (in hours) from the end of the appointment until the note was submitted. Continuity of Care was the percent of patients who attended 3 or more sessions within 45 days with the same provider. Depression symptoms were measured with the PHQ-9 (range, 0–27), consisting of nine items assessing the frequency of a range of depression symptoms. Anxiety symptoms were measured with the GAD-7 (range, 0–21), consisting of seven items assessing the frequency of a range of anxiety symptoms. Depression and anxiety symptoms were modeled separately. The primary outcome was change in clinical symptoms over time (log-days in treatment, measured from the start of therapy [t = 0], estimated at 1-week post-treatment, corresponding to the average duration of care across all patients plus one week [t = 45 days (average duration) + 7 days]. Statistical analyses We used 3-level mixed-effects regression models to estimate clinical outcomes, with random intercepts to account for clustering of repeated observations (level-1) within participants (level-2) and providers (level-3). An identity link was used for continuous outcomes (depression and anxiety symptoms) and a logit link for categorical outcomes. Symptom change, modeled using log-days since the start of treatment, was the main effect of interest. To determine whether AI-notetaking use was associated with symptom change, AI use by log-days was included as an interaction term. Marginal effects were estimated using the delta method 12 . All statistical tests were 2-sided with statistical significance set at α levels of .05. All analyses were conducted in R, version 2024.04.2 13 . Results Adoption and Utilization Session recording and downstream AI tools were available for 853,605 total appointments, corresponding to a total of 8,724 providers and 189,530 patients, and we first measured adoption among this study sample. Overall receptivity to session recording grew steadily, although patients, compared to providers, were more likely to consent to session recordings. Across the study period, 66.2% of patients (n=125,437) consented regularly (i.e., for over half their sessions), with 80.9% (n=153,383) trying session recording at least once. Patient-consent was stable across the study period ( b =-0.01, p = 0.293; Figure 4), with a clear interaction between consentee [patient vs. provider] on consent rates ( b =0.08, p <0.001), where providers’ consent rate started lower but increased faster. Although less receptive than patients, providers demonstrated steady adoption across the study period. At one year since rollout, 47.4% of providers (n=4,139) consented regularly (i.e., for over half their sessions) (Figure 4), with 81.3% (n=7,094) consented to having their sessions recorded at least once. The average by-provider consent rate increased steadily over time ( b =0.07, p <0.001), from 46.2% to 66.7% by the end of the study period. Consent rates for session recording and AI tools varied by provider demographics. Male providers ( n =1,180) consented at a higher rate than female providers ( n =7,323; 51.6% vs. 48.5%; t (1592.5)=2.4, p =0.016), and both had substantially higher rates than non-binary providers ( n =53; 30.1%, p s < 0.001). Across licensures, those with LICSW ( n =3,474), LMFT ( n =1,073), and LMHC ( n =3,661) consented at approximately the same rate (46.7% to 49.9%), whereas clinical psychologists ( n =361) consented much less (37.6%, p s <0.001). Those with longer tenure (by months employed) were less likely to consent than newer providers ( b = -0.003, p <0.001), but there was no effect of years since licensure ( p = 0.362). Given that both patient and provider consent were required to record the session, only 36.6% (n=312,193) of all sessions were recorded during the study period. When recording was declined, it was declined by providers 46.9% of the time, by patients 22.4% of the time, and by both 30.0% of the time. Neither patient or provider were required to provide rationale for declining to have the session recorded. However, across the study period, mutual consent rates continued to increase ( b =0.08, p <0.001), and by the end, 45.0% of sessions were recorded per month. Intake sessions were more likely to be recorded than follow-up sessions (40.4% vs 36.6%, X 2 (1)=505.6, p <0.001). Efficiency and Workflow Across all sessions (recorded or not recorded), 95% of clinical notes took under 160 minutes to complete (i.e., from document open to close), with the median time being 11 minutes. To analyze the impact of AI-powered notetaking on efficiency and workflow, notes taking over 160 minutes were excluded (5%; 42,376 notes). In addition, because the AI-notes were not generated until after a session finished, notes that were submitted before the end of a session were also excluded (7%; 53,387 notes). This resulted in 285,839 sessions that were recorded with AI-notaking available and 465,756 sessions that were not recorded (and thus downstream AI processes were not used), and notes from these sessions were used in subsequent analysis. A primary hypothesis was that AI-powered notetaking would improve providers’ documentation and clinical note submission. For sessions with AI-powered notetaking, the median time for a provider to complete their clinical note was 7 minutes [IQR 2-38] compared to 11 minutes [IQR 4-48] for sessions where AI-notetaking was not used. Overall, AI-notetaking was associated with faster note completion ( b =-6.90, p <0.001), saving providers 25% of the time completing clinical notes. As is apparent from Figure 5, cumulative distribution analyses revealed that the time saving benefit was most apparent for notes taking under 60 minutes to finish, which likely corresponds to those providers who aimed to finish their note before the start of the next session. Integration of AI-notetaking into workflows was common: providers incorporated AI-generated notes directly into their own clinical notes in at least 73.6% of sessions, as tracked through click-to-copy actions (note that this number—and the subsequent analyses—may be conservative since the platform did not track when providers used keyboard copy-and-paste functions). Providers who incorporated AI-generated notes spent 31% (about 8.5 minutes) less time on documentation ( b =-8.5, p <0.001; SI Table 1), and having access to session recordings without using AI notes modestly improved time on documentation ( b =-2.7, p < 0.001). AI-notetaking was also associated with a decrease in the time-to-submit notes: when AI-notes were incorporated, the note was submitted 8.2 hours faster ( b =-8.2, 95% CI, -8.80 to -7.56, p <0.001; Figure 6, SI Table 2) than notes for with appointments without session recording, and 4.5 hours faster when the session recorded, but AI-notes were not incorporated ( b =-4.5, 95% CI -5.13 to -3.96, p <0.001). Overall, 50.5% of providers who used AI-notetaking submitted their note within the same day of the appointment compared to 44.5% of those who did not, and 84.8% submitted within 72 hours, compared to 80.4% who did not. As an additional quality of life measure, AI-notetaking also reduced the proportion of notes signed “after hours” (e.g., from 8pm to 6am). Excluding therapy appointments that took place after hours, notes had a 7.5% probability of being submitted after hours when sessions were not recorded (OR, 0.081, 95% CI 0.077 – 0.085), whereas this probability dropped by about 17.6% when providers copied summaries (SI Table 3), indicating reduced late-night “catch-up” work. Clinical Quality and Outcomes A concern of any automated notetaking is that providers will engage in less case formulation and their patients will see worse outcomes. We found that AI-powered notetaking accelerated provider documentation, but did it come at a cost to clinical quality, both in terms of continuity of care and symptom improvement? To simplify the pattern of AI-notetaking usage, patient-provider pairs were categorized as AI-notetaking adopters if over 50% of their sessions were recorded, otherwise they were categorized as non-adopters. In total there were 183,635 patient-provider pairs 1 : 66,657 adopter pairs, 116,978 non-adopter pairs. AI-notetaking adopter providers had higher continuity of care with their patients (i.e., more patients attending 3 or more sessions within 45 days with the provider) than those who did not adopt AI-notetaking (41.5% among adopters vs. 40.0% among non-adopters, X 2 (1)=20.27, p <0.001). The acuity level (measured by patients’ baseline assessment scores) did not differ between adopters and non-adopters for either depression or anxiety, suggesting that AI tool use was not just used for lower-acuity encounters. Among patients with higher acuity (baseline PHQ-9 >= 10 or GAD7 >= 10), symptoms significantly improved over the study period ( b =-1.47 pts/log-day for depression; -1.34 pts/log-day for anxiety; S1 Table 4). This corresponded to a 6.3-point decrease in depression symptoms and 5.8-point decrease in anxiety symptoms at 1-week post-treatment after 68 days in care (average treatment duration within the study period). The improvement trajectory for those patients who worked with providers using AI-notetaking was consistent with those who did not (i.e., no interaction between AI-notetaking adoption and time in treatment; p s>0.265), suggesting that these new tools do not negatively impact clinical outcomes. Discussion In this large-scale evaluation of AI-powered notetaking in a behavioral health benefit program, we found three central findings: (1) high and growing adoption of session recording and downstream AI tool use under a mutual patient-provider consent framework, (2) reductions in documentation burden and after-hours work, and (3) no detectable adverse impact on clinical continuity or symptom improvement. These findings support the feasibility of large-scale implementation of AI-assisted documentation in psychotherapy settings while maintaining core elements of care. Documentation burden is among the strongest contributors to clinician burnout 3 , 4 , and ambient AI scribes in primary care have shown time-saving benefits 5 – 7 . However, psychotherapy presents unique challenges: notes are narrative, clinically nuanced, and tightly linked to therapeutic formulation. Thus, the 25% reduction in documentation time and 8.2-hour acceleration in note submission we observed in this study are particularly meaningful. The benefits were most pronounced among providers completing notes within an hour of session end, suggesting that AI tools may support real-time or near-real-time documentation practices rather than simply shortening delayed administrative work. Reduced after-hours documentation further suggests improvements in provider quality of life. In this study, providers who used AI-powered notetaking were approximately 17% less likely to submit notes after-hours. Such late-night “catch-up” work is strongly associated with burnout and work-life imbalance. 4 By decreasing after-hours signings, AI-notetaking may help preserve clinician capacity—an increasingly critical concern amid national behavioral health workforce shortages. Furthermore, linking utilization patterns to objective measures of time and cost savings is key for demonstrating return on investment and helping providers learn what they may gain by using the tools. The primary hurdles in implementing session recordings and AI-powered notetaking were human-centered rather than technological. Similar to other surveys of mental health professionals 14 , 15 , therapists’ attitudes toward AI were ambivalent and context-dependent (see Supplemental Analysis Preliminary AI Attitudes among Providers for qualitative summary of these concerns). Prior to roll-out, the single biggest barrier was therapists’ deep skepticism about data handling. Even though the platform leveraged existing administrative and technical safeguards to maintain confidentiality and protect data, such as deleting session audio recordings and raw transcripts after summarization and de-identifying sessions for longer-term storage, providers were more reticent to record sessions than patients. Two major concerns with sessions recording and AI-assisted documentation were that recording may undermine the trust and openness that therapy relies on, and that access to AI-assisted documentation may make providers less present or professional. In behavioral health and other therapeutic contexts, adoption depends on both building patient trust and provider comfort. Our mutual consent framework provided a clear, structured process addressing this concern, ensuring that no recording took place without by-session buy-in from both provider and patient and that both parties knew that their session would be summarized after it concluded. The consent procedure was simple and repeated (Fig. 1 shows one example of a prompt at the start of each session) rather than a one-time consent process. Users—both patients and providers—had quick access to clear terms of service and privacy policies to review the safeguards and ensure usage aligns with HIPAA and state-level requirements. Across the study period, provider consent rates increased steadily and patient consent rates remained stable, consistent with increased comfort with session recording over time. To address the second major concern—that providers may be less present, resulting in a negative clinical impact—we measured continuity of care and symptom improvement to ensure that the implementation of session recording and AI-notetaking did not hurt patient outcomes. The acuity level of patients was equivalent in encounters with and without AI notaking, indicating the tool was not used for just “easy” or just “hard” cases. Importantly, we observed no differences in depression and anxiety symptom trajectories between adopters and non-adopters of AI-notetaking, and continuity of care was slightly higher for adopters. Symptom improvement was clinically meaningful and equivalent across groups. These findings align with other, much smaller studies showing that AI-assisted documentation does not inherently diminish therapeutic effectiveness—and may actually improve it 16 . While we did not directly measure therapeutic alliance, the absence of adverse clinical outcomes provides reassuring evidence that structured, consent-based AI integration can coexist with high-quality psychotherapy. Limitations This study is subject to important limitations inherent to its observational design. In particular, results may be confounded by systematic differences between providers who choose to adopt the session recording and AI notetaking and those who do not, including variation in motivation, documentation practices, typing proficiency, technological comfort, openness to workflow change, work schedules, and baseline efficiency. Similar selection biases may occur at the session and patient level: although acuity levels were equivalent in cases when the tool was used versus not, providers may have preferentially used the tool with more cooperative patients or in routine contexts that are easier to document. Finally, the analysis does not clearly distinguish the extent of AI-generated text usage (e.g., minimal use versus substantial reliance) and may underestimate the use of AI in documentation since the platform only tracked click-to-copy action within the platform and not any third-party tools the providers may use. Conclusions As consumer AI becomes more accessible, providers may be tempted to adopt unsanctioned tools like chatGPT or other purported “clinical” AI, which risks undermining patient consent and data security. Embedding AI notetaking capabilities within a therapeutic platform helps maintain auditable data trails, enforce risk management framework, and ensure consistent oversight. Here, we showed that session recording and AI notetaking, implemented with a clear consent procedure, were associated with reduced documentation burden and improved provider quality of life, without differences in clinical outcomes for patients. Declarations Acknowledgement The authors would like to thank Dan Harrah and Jane Huston for their feedback on this manuscript and their help with data curation. This study received no funding. Data Availability The data in this study comprises personal health information that is considered highly sensitive and confidential. Thus, these data are available upon request for researchers based upon compliance with legal, regulatory, confidentiality, and patient privacy requirements. Please contact the Yale Human Research Protection Program with requests ( [email protected] ) Competing Interests Dr Ward, Dr Hawrilenko, Dr Brown, and Dr Chekroud reported being employed by and holding equity in Spring Care Inc outside the submitted work. In addition, Dr Chekroud reported being the lead inventor on 3 patent submissions relating to treatment for major depressive disorder (US Patent and Trademark Office number Y0087.70116US00 and provisional application numbers 62/491 660 and 62/629 041) outside the submitted work. Finally, Dr Chekroud reported holding equity in Carbon Health Technologies Inc, Wheel Health Inc, Parallel Technologies Inc, Healthie Inc, and UnitedHealthcare; receiving consulting fees from Fortress; and providing unpaid advisory services to health care technology startups outside the submitted work. No other disclosures were reported. Author Contributions CRediT: EJW: Conceptualization, Data curation, Formal analysis, Investigation, Methodology, Project administration, Resources, Software, Validation, Visualization, Writing – original draft, Writing – review & editing; MH: Conceptualization, Data curation, Methodology, Project administration, Resources, Supervision, Writing – review & editing; MB: Conceptualization, Resources, Supervision, Writing – review & editing; AMC: Conceptualization, Supervision, Writing – review & editing. References Stubbe DE. The Therapeutic Alliance: The Fundamental Element of Psychotherapy. Focus J Life Long Learn Psychiatry . 2018;16(4):402. doi:10.1176/appi.focus.20180022 Baier AL, Kline AC, Feeny NC. Therapeutic alliance as a mediator of change: A systematic review and evaluation of research. Clin Psychol Rev . 2020;82:101921. doi:10.1016/j.cpr.2020.101921 Gesner E, Dykes PC, Zhang L, Gazarian P. Documentation Burden in Nursing and Its Role in Clinician Burnout Syndrome. Appl Clin Inform . 2022;13(5):983-990. doi:10.1055/s-0042-1757157 AMIA Survey Underscores Impact of Excessive Documentation Burden | AMIA - American Medical Informatics Association. Accessed September 30, 2025. https://amia.org/news-publications/amia-survey-underscores-impact-excessive-documentation-burden Tierney AA, Gayre G, Hoberman B, et al. Ambient Artificial Intelligence Scribes to Alleviate the Burden of Clinical Documentation. NEJM Catal Innov Care Deliv . Published online February 21, 2024. doi:10.1056/CAT.23.0404 Tierney AA, Gayre G, Hoberman B, et al. Ambient Artificial Intelligence Scribes: Learnings after 1 Year and over 2.5 Million Uses. NEJM Catal . 2025;6(5):CAT.25.0040. doi:10.1056/CAT.25.0040 Stults CD, Deng S, Martinez MC, et al. Evaluation of an Ambient Artificial Intelligence Documentation Platform for Clinicians. JAMA Netw Open . 2025;8(5):e258614. doi:10.1001/jamanetworkopen.2025.8614 Does AI-Powered Clinical Documentation Enhance Clinician Efficiency? A Longitudinal Study | NEJM AI. Accessed September 24, 2025. https://ai.nejm.org/doi/full/10.1056/AIoa2400659 Kroenke K, Spitzer RL, Williams JBW, Löwe B. The Patient Health Questionnaire Somatic, Anxiety, and Depressive Symptom Scales: a systematic review. Gen Hosp Psychiatry . 2010;32(4):345-359. doi:10.1016/j.genhosppsych.2010.03.006 Löwe B, Unützer J, Callahan CM, Perkins AJ, Kroenke K. Monitoring depression treatment outcomes with the patient health questionnaire-9. Med Care . 2004;42(12):1194-1201. doi:10.1097/00005650-200412000-00006 Spitzer RL, Kroenke K, Williams JBW, Löwe B. A Brief Measure for Assessing Generalized Anxiety Disorder: The GAD-7. Arch Intern Med . 2006;166(10):1092-1097. doi:10.1001/archinte.166.10.1092 Oehlert GW. A Note on the Delta Method. Am Stat . 1992;46(1):27-29. doi:10.1080/00031305.1992.10475842 R Core Team. R: A Language and Environment for Statistical Computing . R Foundation for Statistical Computing; 2022. https://www.R-project.org/ Cross S, Bell I, Nicholas J, et al. Use of AI in Mental Health Care: Community and Mental Health Professionals Survey. JMIR Ment Health . 2024;11(1):e60589. doi:10.2196/60589 Hipgrave L, Goldie J, Dennis S, Coleman A. Balancing risks and benefits: Clinicians’ perspectives on the use of generative AI chatbots in mental healthcare. Front Digit Health . 2025;7:1606291. doi:10.3389/fdgth.2025.1606291 Sadeh-Sharvit S, Camp TD, Horton SE, et al. Effects of an Artificial Intelligence Platform for Behavioral Interventions on Depression and Anxiety Symptoms: Randomized Clinical Trial. J Med Internet Res . 2023;25(1):e46781. doi:10.2196/46781 Footnotes Pairs needed to have 45 days or more within the study period after the first session to be included. Additional Declarations Competing interest reported. Dr Ward, Dr Hawrilenko, Dr Brown, and Dr Chekroud reported being employed by and holding equity in Spring Care Inc outside the submitted work. In addition, Dr Chekroud reported being the lead inventor on 3 patent submissions relating to treatment for major depressive disorder (US Patent and Trademark Office number Y0087.70116US00 and provisional application numbers 62/491 660 and 62/629 041) outside the submitted work. Finally, Dr Chekroud reported holding equity in Carbon Health Technologies Inc, Wheel Health Inc, Parallel Technologies Inc, Healthie Inc, and UnitedHealthcare; receiving consulting fees from Fortress; and providing unpaid advisory services to health care technology startups outside the submitted work. No other disclosures were reported. Supplementary Files AINotetakingScientificReportsInitialSubmissionSI.pdf Cite Share Download PDF Status: Posted Version 1 posted You are reading this latest preprint version Research Square lets you share your work early, gain feedback from the community, and start making changes to your manuscript prior to peer review in a journal. As a division of Research Square Company, we’re committed to making research communication faster, fairer, and more useful. We do this by developing innovative software and high quality services for the global research community. Our growing team is made up of researchers and industry professionals working together to solve the most critical problems facing scientific publishing. Also discoverable on Platform About Our Team In Review Editorial Policies Advisory Board Help Center Resources Author Services Accessibility API Access RSS feed Manage Cookie Preferences © Research Square 2026 | ISSN 2693-5015 (online) Privacy Policy Terms of Service Do Not Sell My Personal Information {"props":{"pageProps":{"initialData":{"identity":"rs-9589129","acceptedTermsAndConditions":true,"allowDirectSubmit":false,"archivedVersions":[],"articleType":"Article","associatedPublications":[],"authors":[{"id":633253537,"identity":"cd8ea079-5869-4995-8c2f-d1b90e85be88","order_by":0,"name":"Emily Ward","email":"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAAA6klEQVRIiWNgGAWjYLACngogwQxGDDIMEkRpOYPQwsNDlBbeNghNnBZ+ieRnH97Oq8szb+c9/Lqgwo7HXrqB8cPHHNxaJGekGc+cu+1wscxhvjTrGWeSeXhkDjBLztyGW4vB7QRjZt5tBxJnMPOYGfO2MQMdlsAGFMGtxf52+mdm3jl1UC3/6glrMZDOAdrSwAzSYvyYt+EwYS0S998UM845drhYAmgLM8+x4zw8NxKb8fqFv+f4ZoY3NXV5EvxnjD/z1FTLsc9IPvjhIx4tMJAAxGzQGGFsIKweqoX5A1FKR8EoGAWjYMQBAGUTRZPfG0ehAAAAAElFTkSuQmCC","orcid":"","institution":"Spring Health","correspondingAuthor":true,"prefix":"","firstName":"Emily","middleName":"","lastName":"Ward","suffix":""},{"id":633253538,"identity":"fb8950a9-bc68-41a4-959d-e91da4253e04","order_by":1,"name":"Matt Hawrilenko","email":"","orcid":"","institution":"Spring Health","correspondingAuthor":false,"prefix":"","firstName":"Matt","middleName":"","lastName":"Hawrilenko","suffix":""},{"id":633253539,"identity":"4340c66a-025a-44ee-806e-7d5e076d8752","order_by":2,"name":"Millard Brown","email":"","orcid":"","institution":"Spring Health","correspondingAuthor":false,"prefix":"","firstName":"Millard","middleName":"","lastName":"Brown","suffix":""},{"id":633253540,"identity":"23e30617-3b49-4b7b-b426-d46b1375a988","order_by":3,"name":"Adam Chekroud","email":"","orcid":"","institution":"Yale University","correspondingAuthor":false,"prefix":"","firstName":"Adam","middleName":"","lastName":"Chekroud","suffix":""}],"badges":[],"createdAt":"2026-05-01 20:38:11","currentVersionCode":1,"declarations":"","doi":"10.21203/rs.3.rs-9589129/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-9589129/v1","draftVersion":[],"editorialEvents":[],"editorialNote":"","failedWorkflow":false,"files":[{"id":108840792,"identity":"72bf8b1f-7034-47b8-8bf1-79c428bffc07","added_by":"auto","created_at":"2026-05-09 00:56:14","extension":"png","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":57317,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eThe desktop consent interface. \u003c/strong\u003eThis shows the interface to obtain mutual consent from providers (left) and patients (right) . The session was only recorded if both agreed. During the session, the bottom toolbar displayed the recording status and either party could opt-out at any time.\u003c/p\u003e","description":"","filename":"image1.png","url":"https://assets-eu.researchsquare.com/files/rs-9589129/v1/69003735eb4b6231f8ef5e9c.png"},{"id":108840793,"identity":"e8de3bad-17b5-45e6-a61c-d06cebc28d8c","added_by":"auto","created_at":"2026-05-09 00:56:14","extension":"png","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":70590,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eOverview of session recording, transcription, and downstream AI processes. \u003c/strong\u003eAfter the therapy session concluded on the telehealth platform, the audio recording was temporarily stored in an S3 bucket. This triggered a transcription job sequence that sent the recording to be transcribed, then generated downstream AI products, such as AI-notes, summaries, and takeaways, which were delivered to providers and patients on the platform. The session transcript was de-identified and stored in another S3 bucket; the audio recording and raw transcript were deleted after 72 hours.\u003c/p\u003e","description":"","filename":"image2.png","url":"https://assets-eu.researchsquare.com/files/rs-9589129/v1/f947bc8f20153ce5c4b444a1.png"},{"id":108840795,"identity":"ceadd424-877e-4c7e-9d10-6e2f8da5b4d8","added_by":"auto","created_at":"2026-05-09 00:56:15","extension":"png","order_by":3,"title":"Figure 3","display":"","copyAsset":false,"role":"figure","size":483397,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eAn example of AI-powered session notes.\u003c/strong\u003e AI-powered notes were generated automatically after a therapy session and summarized for providers. These summaries were presented alongside the text-box for clinical notes and providers could click-to-copy the AI-generated text directly into their workflow.\u003c/p\u003e","description":"","filename":"image3.png","url":"https://assets-eu.researchsquare.com/files/rs-9589129/v1/8310dae32e6f34a3673ee993.png"},{"id":108840796,"identity":"4a6c0be1-d41b-4d82-828f-d7a896012b65","added_by":"auto","created_at":"2026-05-09 00:56:15","extension":"png","order_by":4,"title":"Figure 4","display":"","copyAsset":false,"role":"figure","size":928701,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eConsent rates for patients and providers.\u003c/strong\u003e Consent rates among all patients [red] with access to the session recording feature and consent rates among the provider network for those who ever consented [dark blue], the average by-provider consent rate [blue] and by those who consent regularly (\u0026gt;50% of their sessions) [light blue].\u003c/p\u003e","description":"","filename":"image4.png","url":"https://assets-eu.researchsquare.com/files/rs-9589129/v1/a7ad6a3992c91b528c932bc5.png"},{"id":108840797,"identity":"ca4475cf-9fef-415d-aeac-27b5717fa293","added_by":"auto","created_at":"2026-05-09 00:56:15","extension":"png","order_by":5,"title":"Figure 5","display":"","copyAsset":false,"role":"figure","size":596913,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eNote completion times. \u003c/strong\u003eThe cumulative distribution (% of notes) based on how long to complete the clinical note (in minutes). When sessions were recorded with AI-notetaking (cyan), notes were completed faster, especially for those finished within an hour, compared to those associated with sessions where AI features were available, but not used (pink).\u003c/p\u003e","description":"","filename":"image5.png","url":"https://assets-eu.researchsquare.com/files/rs-9589129/v1/0016a885cd706e1088aabacd.png"},{"id":108840794,"identity":"2b4d4b1c-9ceb-4090-8c1c-414706f11908","added_by":"auto","created_at":"2026-05-09 00:56:15","extension":"png","order_by":6,"title":"Figure 6","display":"","copyAsset":false,"role":"figure","size":844958,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eTime to sign and submit a clinical note. \u003c/strong\u003eSubmission times (hours) associated with appointments where there was no session recording, session recording but AI-notes were not incorporated into the clinical note, or session recording and AI-notes were incorporated.\u003c/p\u003e","description":"","filename":"image6.png","url":"https://assets-eu.researchsquare.com/files/rs-9589129/v1/d6c696d8c6d28599c5842771.png"},{"id":109067531,"identity":"118990d6-9266-498c-b518-04f0ad74a4f9","added_by":"auto","created_at":"2026-05-12 09:55:24","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":2833034,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-9589129/v1/2b40c149-ccc9-49aa-bf8a-92337650542c.pdf"},{"id":108840791,"identity":"68f6ece3-b980-4866-89e5-040b85b5ec6e","added_by":"auto","created_at":"2026-05-09 00:56:14","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"supplement","size":151291,"visible":true,"origin":"","legend":"","description":"","filename":"AINotetakingScientificReportsInitialSubmissionSI.pdf","url":"https://assets-eu.researchsquare.com/files/rs-9589129/v1/857491b4e5d23c7441b89320.pdf"}],"financialInterests":"Competing interest reported. Dr Ward, Dr Hawrilenko, Dr Brown, and Dr Chekroud reported being employed by and holding equity in Spring Care Inc outside the submitted work. In addition, Dr Chekroud reported being the lead inventor on 3 patent submissions relating to treatment for major depressive disorder (US Patent and Trademark Office number Y0087.70116US00 and provisional application numbers 62/491 660 and 62/629 041) outside the submitted work. Finally, Dr Chekroud reported holding equity in Carbon Health Technologies Inc, Wheel Health Inc, Parallel Technologies Inc, Healthie Inc, and UnitedHealthcare; receiving consulting fees from Fortress; and providing unpaid advisory services to health care technology startups outside the submitted work. No other disclosures were reported.","formattedTitle":"AI Notetaking in Psychotherapy: Consent, Efficiency, and Clinical Outcomes from a Large-Scale Evaluation","fulltext":[{"header":"Key Takeaways","content":"\u003cul\u003e\n \u003cli\u003eAI-notetaking can be scaled successfully in psychotherapy, easing provider burden while maintaining patient outcomes.\u0026nbsp;\u003c/li\u003e\n \u003cli\u003eAI-notetaking was associated with reduced documentation burden and increased efficiency: clinical notes were completed 25% faster, submitted 8.2 hours sooner and with\u003cstrong\u003e\u0026nbsp;\u003c/strong\u003efewer after-hours signings.\u0026nbsp;\u003c/li\u003e\n \u003cli\u003eThe efficiency gains associated with AI-notetaking did not hurt clinical quality: clinical outcomes (continuity of care, depression/anxiety improvement) were just as strong or better as without AI.\u0026nbsp;\u003c/li\u003e\n \u003cli\u003eInitial provider skepticism about data handling and trust was overcome through mutual consent and transparent security measures, leading to steadily growing acceptance.\u003c/li\u003e\n\u003c/ul\u003e"},{"header":"Introduction","content":"\u003cp\u003eProvider capacity and quality of engagement are key drivers of patient outcomes\u003csup\u003e1,2\u003c/sup\u003e, yet documentation burden can erode both. Clinical documentation is among the most significant drivers of clinician burnout\u003csup\u003e3\u003c/sup\u003e. In psychotherapy, notes require a nuanced narrative yet must also meet billing, compliance, and clinical communication requirements. These demands consume substantial provider time, slow the administrative and billing cycle, and can often spill into late-night \u0026ldquo;catch-up\u0026rdquo; hours\u003csup\u003e4\u003c/sup\u003e.\u003c/p\u003e\n\u003cp\u003eAI-powered notetaking has emerged as a potential solution to improve provider quality of life. In this approach, therapy sessions are transcribed and summarized using Large Language Models (LLMs). Such \u0026ldquo;ambient scribes\u0026rdquo; have been employed in primary care\u003csup\u003e5\u0026ndash;7\u003c/sup\u003e and other settings, demonstrating time-saving benefits and a reduction in the cognitive load associated with documentation. Large-scale implementations have shown significant reductions in time spent in notes per appointment\u003csup\u003e5\u003c/sup\u003e and time spent on the electronic health records (EHR) after hours, although some studies found no significant change in after-hours work\u003csup\u003e8\u003c/sup\u003e. Both physicians and patients have reported positive experiences, with physicians feeling more present during visits and patients perceiving that their doctor spent less time looking at a computer. However, such ambient scribes are not always directly integrated into EHR, making it difficult to track the incorporation of AI-generated text which offers a novel and objective approach to measuring practice integration.\u003c/p\u003e\n\u003cp\u003eWhile existing research provides a strong foundation for the operational benefits of ambient AI scribes, critical gaps remain, particularly within the specialized context of psychotherapy. One study found that mental health providers had the highest adoption rates of AI notetaking (used in approximately 50% of visits)\u003csup\u003e6\u003c/sup\u003e, suggesting a significant need and potential benefit in this population. Most prior work has largely centered on clinician experience and workflow efficiency, but in psychotherapy, clinical notes are less structured than those in primary care, so understanding the gain (if any) to provider efficiency is important, especially given the mixed results from EHR data.\u003c/p\u003e\n\u003cp\u003ePsychotherapy depends on privacy and trust: patients must feel safe enough to disclose information that makes them vulnerable, and session recordings or AI notetaking tools that reduce this willingness may pose a significant hurdle. Moreover, a potential fear is that automated notetaking may result in less provider engagement with a patient, leading to worse outcomes. Direct clinical impact of AI notetaking adoption, including continuity of care and patient symptom improvement, have not been investigated.\u003c/p\u003e\n\u003ch2\u003e\u003cstrong\u003eThe Current Study \u003c/strong\u003e\u003c/h2\u003e\n\u003cp\u003eSession recording and downstream AI documentation tools were available in over 850,000 total appointments, across over 8,700 providers and 189,000 patients. We evaluated whether and how such AI tools would be integrated into psychotherapy at such a scale, while addressing both operational efficiency and clinical quality. Specifically, our goals were to:\u003c/p\u003e\n\u003col start=\"1\" type=\"1\"\u003e\n\u003cli\u003e\u003cem\u003eMeasure adoption\u003c/em\u003e in a large, nationwide provider network where mutual patient-provider consent was required for session recording and AI-notetaking.\u003c/li\u003e\n\u003cli\u003e\u003cem\u003eAssess workflow impact\u003c/em\u003e using objective EHR-linked note completion and submission times, capturing after-hours documentation rather than self-report.\u003c/li\u003e\n\u003cli\u003e\u003cem\u003eExamine integration into practice\u003c/em\u003e through tracking of AI-generated text incorporation into clinical documentation.\u003c/li\u003e\n\u003cli\u003e\u003cem\u003eEvaluate clinical quality and patient outcomes\u003c/em\u003e, including continuity of care and symptom improvement (PHQ-9, GAD-7 trajectories) to ensure recording and AI tools had no negative impact.\u003c/li\u003e\n\u003c/ol\u003e"},{"header":"Methods","content":"\u003ch3\u003e\u003cstrong\u003eStudy design\u003c/strong\u003e\u003c/h3\u003e\n\u003cp\u003eThis retrospective cohort study was approved by the Yale Institutional Review Board (IRB protocol ID: 2000029276) and classified as not involving research with human subjects. That is, there was no contact with human subjects by any research team member, data were collected for another purpose (here, transcripts and clinical notes were part of routine care), and there was a data use agreement in place for a limited data set. All data were anonymized prior to analysis.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eProgram Design\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eData came from an employer-sponsored digital mental health benefit (Spring Health) that includes free or low-cost access to a number of psychotherapy or medication management sessions, care navigation, and digital self-help mental wellness resources. Additional sessions were covered as an in-network benefit via the health plan. Care is coordinated through a centralized data system that integrates patient data, assessment results, and patient scheduling. Treatment was primarily delivered through video conferencing and provider notes were submitted using the benefit platform. Approximately 70% of psychotherapy sessions were billed as 60-minute appointments (CPT 90837), 10% as diagnostic evaluations (90791), and the rest a mix of 45-minute sessions (90834), couples therapy (90847) and no shows.\u003c/p\u003e\n\u003cp\u003eThe benefit included a digital mental health assessment tool for initial screening and outcome monitoring. For their initial screening, individuals completed a series of self-report questionnaires to identify common mental health difficulties (such as stress, anxiety, eating, or substance use issues) and completed the Patient Health Questionnaire 9-item scale (PHQ-9) \u003csup\u003e9,10\u003c/sup\u003e for depression. Based on their responses, they may also complete the Generalized Anxiety Disorder 7-item scale (GAD-7) \u003csup\u003e11\u003c/sup\u003e and/or targeted questionnaires for specific issues. (e.g., patients indicating any issues related to trauma were given the PC-PTSD questionnaire). For longitudinal outcome monitoring, participants were reminded every two to four weeks to complete follow-up questionnaires, consisting of PHQ-9 and/or GAD-7 assessments for most conditions.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eAI Strategy\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe benefit program leverages the NIST AI Risk Management Framework as a foundational element of AI governance strategy and maintains an internal AI Governance Board to oversee the responsible development, deployment, and research of AI technologies. This framework has guided the creation of internal policies, the establishment of the governance board itself, and the development of an AI security program, including risk-based testing protocols and embedding safety layers into our products. All AI tools in this study are HIPPA compliant, and implemented within a SOC 2 attested and HITRUST certified platform, leveraging existing administrative and technical safeguards to protect data in alignment with the necessary requirements. In addition, at the time of this study, the program was aligning practices with the ISO/IEC 42001 standard and to pursue formal certification and build a comprehensive program grounded in both NIST and ISO guidance.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eParticipants\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eParticipants were employees and dependents eligible for the benefit from March 1, 2025, and January 31, 2026. Patients were located in the U.S. and 18 years or older. When analyzing consent and adoption, we included anyone who had started treatment within the study period. To measure clinical outcomes, additional criteria were applied: patients were included in the analyses if they had a baseline assessment within 30 days prior to starting treatment, if they screened positive for depression and/or anxiety (PHQ-9\u0026gt;=10 or GAD-7\u0026gt;=10) and if they had an additional assessment after treatment began. Participants must have started therapy during the study period and remained eligible for the benefit during the study period. Assessments taken up to the end of the study period were included (note that patients starting care later in the study period had less opportunity for improvement). Patients did not receive monetary reward.\u003c/p\u003e\n\u003ch3\u003e\u003cstrong\u003eProviders\u003c/strong\u003e\u003c/h3\u003e\n\u003cp\u003eProviders included licensed clinical psychologists, mental health counselors, licensed clinical social workers, psychiatrists, and psychiatric nurse practitioners. Approximately 42% of providers were LMHC, 40% were LICSW, 13% were LMFT, and 5% were clinical psychologists. Providers had 13.3 years of licensed experience and their average tenure in the benefit network was 25.9 months.\u003c/p\u003e\n\u003ch3\u003e\u003cstrong\u003eRecording Eligibility \u0026amp; Consent Procedures\u003c/strong\u003e\u003c/h3\u003e\n\u003cp\u003eSession recording and AI-powered notetaking were only available for specific patients engaging in tele-therapy sessions: the patient\u0026rsquo;s employer or health plan must have opted into session recording, and the patient could not be located in Vermont. If a patient or session did not meet eligibility criteria, no consent prompt appeared.\u003c/p\u003e\n\u003cp\u003eBecause audio recording was required for AI-powered notetaking, both patients and providers were prompted to give explicit consent to recording at the beginning of each session (Figure 1). They were also given information about the downstream processes that would use the session recordings: patients were informed that their session would be transcribed, that their provider would receive detailed notes about the session, and that they\u0026mdash;the patient\u0026mdash;would receive a summary and list of takeaways (Figure 1, right). Providers were told that the session would be recorded and could be used to speed their own clinical notes. They were also informed that the audio recording would be deleted immediately after the session was summarized (Figure 1, left). Thus, consent included session recording and downstream summarization, etc. Both parties were provided with links to the terms, privacy policy, and help center for any other questions. Either party could pause or stop the recording at any point. Providers did not receive any incentives to record sessions or use AI-generated notes.\u003c/p\u003e\n\u003ch3\u003e\u003cstrong\u003eIntervention\u003c/strong\u003e\u003c/h3\u003e\n\u003cp\u003e\u003cem\u003eTranscription:\u0026nbsp;\u003c/em\u003eAn overview of the data processing pipeline is shown in Figure 2. Sessions were recorded using an open-source cloud platform for collecting audio data. At the end of a session, the audio file was uploaded to an AWS S3 bucket. The upload triggered the core backend to link the recording to session data and to begin transcription. The audio file was transcribed (using Amazon Transcribe), and the final transcript was used to produce downstream AI processes, stripped of all personally identifiable information (PII) and protected health information (PHI), and then stored for quality assurance .\u003c/p\u003e\n\u003cp\u003eThe audio file was deleted and the raw transcripts were available to providers for 72 hours then deleted. Only fully de-identified transcripts collected within this routine care pipeline were available for further research, and the data was accessible only to authorized staff through a documented approval process. Furthermore, these transcripts were not used to train or fine tune new AI models.\u003c/p\u003e\n\u003cp\u003e\u003cem\u003eSummarization:\u0026nbsp;\u003c/em\u003eSession summaries were generated using an LLM and a standardized prompting framework for transcript-based therapy documentation. The prompt constrained the model to produce de-identified, professionally worded summaries grounded only in information explicitly present in the session transcript. It further instructed the model to omit irrelevant procedural details, such as technical difficulties, and to avoid redundancy across sections.\u003c/p\u003e\n\u003cp\u003e\u003cem\u003eSafeguards:\u0026nbsp;\u003c/em\u003eSummaries were limited to behavioral observations, diagnostic terminology, and absolute language that was directly supported by transcript content.\u003c/p\u003e\n\u003cp\u003e\u003cem\u003eEditing workflow:\u0026nbsp;\u003c/em\u003eAI-generated transcripts and notes were made available to providers during clinical documentation. The documentation included two required fields (Figure 3): \u0026ldquo;Presenting Problem,\u0026rdquo; capturing the client\u0026rsquo;s broader clinical concerns and therapeutic context and a \u0026ldquo;Session Focus\u0026rdquo; section, comprising current status and changes since the previous session, therapeutic themes and interventions, and safety/risk content. Thus, the AI-powered summaries were automatically generated for each of these two sections and presented to providers alongside the text-field for their clinical notes. Providers could incorporate these AI-generated notes directly via a button in the platform that allowed them to copy-to-clipboard and whenever the button was clocked, the platform logged the action. However, providers could also copy-and-paste the text without using the platform copy button (using the usual keyboard shortcuts). In this case, the platform could not track whether the AI-powered notes were integrated into clinical notes. Thus, the copy-to-clipboard button action may be a conservative measure of integration into workflows. Providers could further review and edit the notes with their own observations before completing and submitting the note, but the extent to which AI-generated text was used minimally, partially, or substantially was not tracked.\u003c/p\u003e\n\u003ch4\u003e\u003cstrong\u003eMeasures\u003c/strong\u003e\u003c/h4\u003e\n\u003cp\u003e\u003cem\u003eNote completion time\u0026nbsp;\u003c/em\u003ecorresponded to time elapsed (in minutes) from when the clinical note page was opened in the provider\u0026rsquo;s browser until it was closed. \u003cem\u003eTime to note submission\u003c/em\u003e measured the time (in hours) from the end of the appointment until the note was submitted. \u003cem\u003eContinuity of Care\u003c/em\u003e was the percent of patients who attended 3 or more sessions within 45 days with the same provider. \u003cem\u003eDepression symptoms\u003c/em\u003e were measured with the PHQ-9 (range, 0\u0026ndash;27), consisting of nine items assessing the frequency of a range of depression symptoms. \u003cem\u003eAnxiety symptoms\u003c/em\u003e were measured with the GAD-7 (range, 0\u0026ndash;21), consisting of seven items assessing the frequency of a range of anxiety symptoms. Depression and anxiety symptoms were modeled separately.\u003c/p\u003e\n\u003cp\u003eThe primary outcome was change in clinical symptoms over \u003cem\u003etime\u003c/em\u003e (log-days in treatment, measured from the start of therapy [t = 0], estimated at 1-week post-treatment, corresponding to the average duration of care across all patients plus one week [t = 45 days (average duration) + 7 days].\u003c/p\u003e\n\u003ch4\u003e\u003cstrong\u003eStatistical analyses\u003c/strong\u003e\u003c/h4\u003e\n\u003cp\u003eWe used 3-level mixed-effects regression models to estimate clinical outcomes, with random intercepts to account for clustering of repeated observations (level-1) within participants (level-2) and providers (level-3). An identity link was used for continuous outcomes (depression and anxiety symptoms) and a logit link for categorical outcomes. Symptom change, modeled using log-days since the start of treatment, was the main effect of interest. To determine whether AI-notetaking use was associated with symptom change, AI use by log-days was included as an interaction term. Marginal effects were estimated using the delta method\u003csup\u003e12\u003c/sup\u003e. All statistical tests were 2-sided with statistical significance set at \u0026alpha; levels of .05. All analyses were conducted in R, version 2024.04.2\u003csup\u003e13\u003c/sup\u003e.\u003c/p\u003e"},{"header":"Results","content":"\u003ch4\u003e\u003cstrong\u003eAdoption and Utilization\u003c/strong\u003e\u003c/h4\u003e\n\u003cp\u003eSession recording and downstream AI tools were available for 853,605 total appointments, corresponding to a total of 8,724 providers and 189,530 patients, and we first measured adoption among this study sample.\u003c/p\u003e\n\u003cp\u003eOverall receptivity to session recording grew steadily, although patients, compared to providers, were more likely to consent to session recordings. Across the study period, 66.2% of patients (n=125,437) consented regularly (i.e., for over half their sessions), with 80.9% (n=153,383) trying session recording at least once. Patient-consent was stable across the study period (\u003cem\u003eb\u003c/em\u003e=-0.01, \u003cem\u003ep\u003c/em\u003e= 0.293; Figure 4), with a clear interaction between consentee [patient vs. provider] on consent rates (\u003cem\u003eb\u003c/em\u003e=0.08, \u003cem\u003ep\u003c/em\u003e\u0026lt;0.001), where providers\u0026rsquo; consent rate started lower but increased faster.\u003c/p\u003e\n\u003cp\u003eAlthough less receptive than patients, providers demonstrated steady adoption across the study period. At one year since rollout, 47.4% of providers (n=4,139) consented regularly (i.e., for over half their sessions) (Figure 4), with 81.3% (n=7,094) consented to having their sessions recorded at least once. The average by-provider consent rate increased steadily over time (\u003cem\u003eb\u003c/em\u003e=0.07, \u003cem\u003ep\u003c/em\u003e\u0026lt;0.001), from 46.2% to 66.7% by the end of the study period.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eConsent rates for session recording and AI tools varied by provider demographics. Male providers (\u003cem\u003en\u003c/em\u003e=1,180) consented at a higher rate than female providers (\u003cem\u003en\u003c/em\u003e=7,323; 51.6% vs. 48.5%; \u003cem\u003et\u003c/em\u003e(1592.5)=2.4, \u003cem\u003ep\u003c/em\u003e=0.016), and both had substantially higher rates than non-binary providers (\u003cem\u003en\u003c/em\u003e=53; 30.1%, \u003cem\u003ep\u003c/em\u003es \u0026lt; 0.001). Across licensures, those with LICSW (\u003cem\u003en\u003c/em\u003e=3,474), LMFT (\u003cem\u003en\u003c/em\u003e=1,073), and LMHC (\u003cem\u003en\u003c/em\u003e=3,661) consented at approximately the same rate (46.7% to 49.9%), whereas clinical psychologists (\u003cem\u003en\u003c/em\u003e=361) consented much less (37.6%, \u003cem\u003ep\u003c/em\u003es \u0026lt;0.001). Those with longer tenure (by months employed) were less likely to consent than newer providers (\u003cem\u003eb\u003c/em\u003e= -0.003, \u003cem\u003ep\u003c/em\u003e\u0026lt;0.001), but there was no effect of years since licensure (\u003cem\u003ep\u003c/em\u003e= 0.362).\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eGiven that both patient and provider consent were required to record the session, only 36.6% (n=312,193) of all sessions were recorded during the study period. When recording was declined, it was declined by providers 46.9% of the time, by patients 22.4% of the time, and by both 30.0% of the time. Neither patient or provider were required to provide rationale for declining to have the session recorded. However, across the study period, mutual consent rates continued to increase (\u003cem\u003eb\u003c/em\u003e=0.08, \u003cem\u003ep\u003c/em\u003e\u0026lt;0.001), and by the end, 45.0% of sessions were recorded per month. Intake sessions were more likely to be recorded than follow-up sessions (40.4% vs 36.6%, \u003cem\u003eX\u003csup\u003e2\u003c/sup\u003e\u003c/em\u003e(1)=505.6, \u003cem\u003ep\u003c/em\u003e\u0026lt;0.001).\u003c/p\u003e\n\u003ch2\u003e\u003cstrong\u003eEfficiency and Workflow\u003c/strong\u003e\u0026nbsp;\u003c/h2\u003e\n\u003cp\u003eAcross all sessions (recorded or not recorded), 95% of clinical notes took under 160 minutes to complete (i.e., from document open to close), with the median time being 11 minutes. To analyze the impact of AI-powered notetaking on efficiency and workflow, notes taking over 160 minutes were excluded (5%; 42,376 notes). In addition, because the AI-notes were not generated until after a session finished, notes that were submitted before the end of a session were also excluded (7%; 53,387 notes). This resulted in 285,839 sessions that were recorded with AI-notaking available and 465,756 sessions that were not recorded (and thus downstream AI processes were not used), and notes from these sessions were used in subsequent analysis.\u003c/p\u003e\n\u003cp\u003eA primary hypothesis was that AI-powered notetaking would improve providers\u0026rsquo; documentation and clinical note submission. For sessions with AI-powered notetaking, the median time for a provider to complete their clinical note was 7 minutes [IQR 2-38] compared to 11 minutes [IQR 4-48] for sessions where AI-notetaking was not used. Overall, AI-notetaking was associated with faster note completion (\u003cem\u003eb\u003c/em\u003e=-6.90, \u003cem\u003ep\u003c/em\u003e\u0026lt;0.001), saving providers 25% of the time completing clinical notes. As is apparent from Figure 5, cumulative distribution analyses revealed that the time saving benefit was most apparent for notes taking under 60 minutes to finish, which likely corresponds to those providers who aimed to finish their note before the start of the next session.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eIntegration of AI-notetaking into workflows was common: providers incorporated AI-generated notes directly into their own clinical notes in at least 73.6% of sessions, as tracked through click-to-copy actions (note that this number\u0026mdash;and the subsequent analyses\u0026mdash;may be conservative since the platform did not track when providers used keyboard copy-and-paste functions). Providers who incorporated AI-generated notes spent 31% (about 8.5 minutes) less time on documentation (\u003cem\u003eb\u003c/em\u003e=-8.5, \u003cem\u003ep\u003c/em\u003e\u0026lt;0.001; SI Table 1), and having access to session recordings without using AI notes modestly improved time on documentation (\u003cem\u003eb\u003c/em\u003e=-2.7, \u003cem\u003ep\u003c/em\u003e\u0026lt; 0.001).\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eAI-notetaking was also associated with a decrease in the time-to-submit notes: when AI-notes were incorporated, the note was submitted 8.2 hours faster (\u003cem\u003eb\u003c/em\u003e=-8.2, 95% CI, -8.80 to -7.56, \u003cem\u003ep\u003c/em\u003e\u0026lt;0.001; Figure 6, SI Table 2) than notes for with appointments without session recording, and 4.5 hours faster when the session recorded, but AI-notes were not incorporated (\u003cem\u003eb\u003c/em\u003e=-4.5, 95% CI -5.13 to -3.96, \u003cem\u003ep\u003c/em\u003e\u0026lt;0.001). Overall, 50.5% of providers who used AI-notetaking submitted their note within the same day of the appointment compared to 44.5% of those who did not, and 84.8% submitted within 72 hours, compared to 80.4% who did not.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eAs an additional quality of life measure, AI-notetaking also reduced the proportion of notes signed \u0026ldquo;after hours\u0026rdquo; (e.g., from 8pm to 6am). Excluding therapy appointments that took place after hours, notes had a 7.5% probability of being submitted after hours when sessions were not recorded (OR, 0.081, 95% CI 0.077 \u0026ndash; 0.085), whereas this probability dropped by about 17.6% when providers copied summaries (SI Table 3), indicating reduced late-night \u0026ldquo;catch-up\u0026rdquo; work.\u003c/p\u003e\n\u003ch2\u003e\u003cstrong\u003eClinical Quality and Outcomes\u003c/strong\u003e\u003c/h2\u003e\n\u003cp\u003eA concern of any automated notetaking is that providers will engage in less case formulation and their patients will see worse outcomes. We found that AI-powered notetaking accelerated provider documentation, but did it come at a cost to clinical quality, both in terms of continuity of care and symptom improvement? To simplify the pattern of AI-notetaking usage, patient-provider pairs were categorized as AI-notetaking adopters if over 50% of their sessions were recorded, otherwise they were categorized as non-adopters. In total there were 183,635 patient-provider pairs\u003ca href=\"#_ftn1\" name=\"_ftnref1\" title=\"\"\u003e\u003c/a\u003e\u003csup\u003e1\u003c/sup\u003e: 66,657 adopter pairs, 116,978 non-adopter pairs.\u003c/p\u003e\n\u003cp\u003eAI-notetaking adopter providers had higher continuity of care with their patients (i.e., more patients attending 3 or more sessions within 45 days with the provider) than those who did not adopt AI-notetaking (41.5% among adopters vs. 40.0% among non-adopters, \u003cem\u003eX\u003csup\u003e2\u003c/sup\u003e\u003c/em\u003e(1)=20.27, \u003cem\u003ep\u003c/em\u003e\u0026lt;0.001). \u0026nbsp;The acuity level (measured by patients\u0026rsquo; baseline assessment scores) did not differ between adopters and non-adopters for either depression or anxiety, suggesting that AI tool use was not just used for lower-acuity encounters.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eAmong patients with higher acuity (baseline PHQ-9 \u0026gt;= 10 or GAD7 \u0026gt;= 10), symptoms significantly improved over the study period (\u003cem\u003eb\u003c/em\u003e=-1.47 pts/log-day for depression; -1.34 pts/log-day for anxiety; S1 Table 4). This corresponded to a 6.3-point decrease in depression symptoms and 5.8-point decrease in anxiety symptoms at 1-week post-treatment after 68 days in care (average treatment duration within the study period). The improvement trajectory for those patients who worked with providers using AI-notetaking was consistent with those who did not (i.e., no interaction between AI-notetaking adoption and time in treatment; \u003cem\u003ep\u003c/em\u003es\u0026gt;0.265), suggesting that these new tools do not negatively impact clinical outcomes.\u003c/p\u003e"},{"header":"Discussion","content":"\u003cp\u003eIn this large-scale evaluation of AI-powered notetaking in a behavioral health benefit program, we found three central findings: (1) high and growing adoption of session recording and downstream AI tool use under a mutual patient-provider consent framework, (2) reductions in documentation burden and after-hours work, and (3) no detectable adverse impact on clinical continuity or symptom improvement. These findings support the feasibility of large-scale implementation of AI-assisted documentation in psychotherapy settings while maintaining core elements of care.\u003c/p\u003e \u003cp\u003eDocumentation burden is among the strongest contributors to clinician burnout\u003csup\u003e\u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e3\u003c/span\u003e,\u003cspan citationid=\"CR4\" class=\"CitationRef\"\u003e4\u003c/span\u003e\u003c/sup\u003e, and ambient AI scribes in primary care have shown time-saving benefits\u003csup\u003e\u003cspan additionalcitationids=\"CR6\" citationid=\"CR5\" class=\"CitationRef\"\u003e5\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR7\" class=\"CitationRef\"\u003e7\u003c/span\u003e\u003c/sup\u003e. However, psychotherapy presents unique challenges: notes are narrative, clinically nuanced, and tightly linked to therapeutic formulation. Thus, the 25% reduction in documentation time and 8.2-hour acceleration in note submission we observed in this study are particularly meaningful. The benefits were most pronounced among providers completing notes within an hour of session end, suggesting that AI tools may support real-time or near-real-time documentation practices rather than simply shortening delayed administrative work.\u003c/p\u003e \u003cp\u003eReduced after-hours documentation further suggests improvements in provider quality of life. In this study, providers who used AI-powered notetaking were approximately 17% less likely to submit notes after-hours. Such late-night \u0026ldquo;catch-up\u0026rdquo; work is strongly associated with burnout and work-life imbalance.\u003csup\u003e\u003cspan citationid=\"CR4\" class=\"CitationRef\"\u003e4\u003c/span\u003e\u003c/sup\u003e By decreasing after-hours signings, AI-notetaking may help preserve clinician capacity\u0026mdash;an increasingly critical concern amid national behavioral health workforce shortages. Furthermore, linking utilization patterns to objective measures of time and cost savings is key for demonstrating return on investment and helping providers learn what they may gain by using the tools.\u003c/p\u003e \u003cp\u003eThe primary hurdles in implementing session recordings and AI-powered notetaking were human-centered rather than technological. Similar to other surveys of mental health professionals\u003csup\u003e\u003cspan citationid=\"CR14\" class=\"CitationRef\"\u003e14\u003c/span\u003e,\u003cspan citationid=\"CR15\" class=\"CitationRef\"\u003e15\u003c/span\u003e\u003c/sup\u003e, therapists\u0026rsquo; attitudes toward AI were ambivalent and context-dependent (see \u003cem\u003eSupplemental Analysis Preliminary AI Attitudes among Providers\u003c/em\u003e for qualitative summary of these concerns). Prior to roll-out, the single biggest barrier was therapists\u0026rsquo; deep skepticism about data handling. Even though the platform leveraged existing administrative and technical safeguards to maintain confidentiality and protect data, such as deleting session audio recordings and raw transcripts after summarization and de-identifying sessions for longer-term storage, providers were more reticent to record sessions than patients.\u003c/p\u003e \u003cp\u003eTwo major concerns with sessions recording and AI-assisted documentation were that recording may undermine the trust and openness that therapy relies on, and that access to AI-assisted documentation may make providers less present or professional. In behavioral health and other therapeutic contexts, adoption depends on both building patient trust and provider comfort. Our mutual consent framework provided a clear, structured process addressing this concern, ensuring that no recording took place without by-session buy-in from both provider and patient and that both parties knew that their session would be summarized after it concluded. The consent procedure was simple and repeated (Fig.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e1\u003c/span\u003e shows one example of a prompt at the start of each session) rather than a one-time consent process. Users\u0026mdash;both patients and providers\u0026mdash;had quick access to clear terms of service and privacy policies to review the safeguards and ensure usage aligns with HIPAA and state-level requirements. Across the study period, provider consent rates increased steadily and patient consent rates remained stable, consistent with increased comfort with session recording over time.\u003c/p\u003e \u003cp\u003eTo address the second major concern\u0026mdash;that providers may be less present, resulting in a negative clinical impact\u0026mdash;we measured continuity of care and symptom improvement to ensure that the implementation of session recording and AI-notetaking did not hurt patient outcomes. The acuity level of patients was equivalent in encounters with and without AI notaking, indicating the tool was not used for just \u0026ldquo;easy\u0026rdquo; or just \u0026ldquo;hard\u0026rdquo; cases. Importantly, we observed no differences in depression and anxiety symptom trajectories between adopters and non-adopters of AI-notetaking, and continuity of care was slightly higher for adopters. Symptom improvement was clinically meaningful and equivalent across groups. These findings align with other, much smaller studies showing that AI-assisted documentation does not inherently diminish therapeutic effectiveness\u0026mdash;and may actually improve it\u003csup\u003e16\u003c/sup\u003e. While we did not directly measure therapeutic alliance, the absence of adverse clinical outcomes provides reassuring evidence that structured, consent-based AI integration can coexist with high-quality psychotherapy.\u003c/p\u003e \u003cdiv id=\"Sec18\" class=\"Section2\"\u003e \u003ch2\u003eLimitations\u003c/h2\u003e \u003cp\u003eThis study is subject to important limitations inherent to its observational design. In particular, results may be confounded by systematic differences between providers who choose to adopt the session recording and AI notetaking and those who do not, including variation in motivation, documentation practices, typing proficiency, technological comfort, openness to workflow change, work schedules, and baseline efficiency. Similar selection biases may occur at the session and patient level: although acuity levels were equivalent in cases when the tool was used versus not, providers may have preferentially used the tool with more cooperative patients or in routine contexts that are easier to document. Finally, the analysis does not clearly distinguish the extent of AI-generated text usage (e.g., minimal use versus substantial reliance) and may underestimate the use of AI in documentation since the platform only tracked click-to-copy action within the platform and not any third-party tools the providers may use.\u003c/p\u003e \u003c/div\u003e"},{"header":"Conclusions","content":"\u003cp\u003eAs consumer AI becomes more accessible, providers may be tempted to adopt unsanctioned tools like chatGPT or other purported \u0026ldquo;clinical\u0026rdquo; AI, which risks undermining patient consent and data security. Embedding AI notetaking capabilities within a therapeutic platform helps maintain auditable data trails, enforce risk management framework, and ensure consistent oversight. Here, we showed that session recording and AI notetaking, implemented with a clear consent procedure, were associated with reduced documentation burden and improved provider quality of life, without differences in clinical outcomes for patients.\u003c/p\u003e"},{"header":"Declarations","content":"\u003ch2\u003e\u003cstrong\u003eAcknowledgement\u003c/strong\u003e\u003c/h2\u003e\n\u003cp\u003eThe authors would like to thank Dan Harrah and Jane Huston for their feedback on this manuscript and their help with data curation. This study received no funding.\u003c/p\u003e\n\u003ch2\u003e\u003cstrong\u003eData Availability\u003c/strong\u003e\u003c/h2\u003e\n\u003cp\u003eThe data in this study comprises personal health information that is considered highly sensitive and confidential. Thus, these data are available upon request for researchers based upon compliance with legal, regulatory, confidentiality, and patient privacy requirements. Please contact the Yale Human Research Protection Program with requests ([email protected])\u003c/p\u003e\n\u003ch2\u003e\u003cstrong\u003eCompeting Interests\u003c/strong\u003e\u003c/h2\u003e\n\u003cp\u003eDr Ward, Dr Hawrilenko, Dr Brown, and Dr Chekroud reported being employed by and holding equity in Spring Care Inc outside the submitted work. In addition, Dr Chekroud reported being the lead inventor on 3 patent submissions relating to treatment for major depressive disorder (US Patent and Trademark Office number Y0087.70116US00 and provisional application numbers 62/491 660 and 62/629 041) outside the submitted work. Finally, Dr Chekroud reported holding equity in Carbon Health Technologies Inc, Wheel Health Inc, Parallel Technologies Inc, Healthie Inc, and UnitedHealthcare; receiving consulting fees from Fortress; and providing unpaid advisory services to health care technology startups outside the submitted work. No other disclosures were reported.\u003c/p\u003e\n\u003ch2\u003e\u003cstrong\u003eAuthor Contributions\u003c/strong\u003e\u003c/h2\u003e\n\u003cp\u003eCRediT: EJW: Conceptualization, Data curation, Formal analysis, Investigation, Methodology, Project administration, Resources, Software, Validation, Visualization, Writing \u0026ndash; original draft, Writing \u0026ndash; review \u0026amp; editing; MH: Conceptualization, Data curation, Methodology, Project administration, Resources, Supervision, Writing \u0026ndash; review \u0026amp; editing; MB: Conceptualization, Resources, Supervision, Writing \u0026ndash; review \u0026amp; editing; AMC: Conceptualization, Supervision, Writing \u0026ndash; review \u0026amp; editing.\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\n\u003cli\u003eStubbe DE. The Therapeutic Alliance: The Fundamental Element of Psychotherapy. \u003cem\u003eFocus J Life Long Learn Psychiatry\u003c/em\u003e. 2018;16(4):402. doi:10.1176/appi.focus.20180022 \u003c/li\u003e\n\u003cli\u003eBaier AL, Kline AC, Feeny NC. Therapeutic alliance as a mediator of change: A systematic review and evaluation of research. \u003cem\u003eClin Psychol Rev\u003c/em\u003e. 2020;82:101921. doi:10.1016/j.cpr.2020.101921 \u003c/li\u003e\n\u003cli\u003eGesner E, Dykes PC, Zhang L, Gazarian P. Documentation Burden in Nursing and Its Role in Clinician Burnout Syndrome. \u003cem\u003eAppl Clin Inform\u003c/em\u003e. 2022;13(5):983-990. doi:10.1055/s-0042-1757157 \u003c/li\u003e\n\u003cli\u003eAMIA Survey Underscores Impact of Excessive Documentation Burden | AMIA - American Medical Informatics Association. Accessed September 30, 2025. https://amia.org/news-publications/amia-survey-underscores-impact-excessive-documentation-burden \u003c/li\u003e\n\u003cli\u003eTierney AA, Gayre G, Hoberman B, et al. Ambient Artificial Intelligence Scribes to Alleviate the Burden of Clinical Documentation. \u003cem\u003eNEJM Catal Innov Care Deliv\u003c/em\u003e. Published online February 21, 2024. doi:10.1056/CAT.23.0404 \u003c/li\u003e\n\u003cli\u003eTierney AA, Gayre G, Hoberman B, et al. Ambient Artificial Intelligence Scribes: Learnings after 1 Year and over 2.5 Million Uses. \u003cem\u003eNEJM Catal\u003c/em\u003e. 2025;6(5):CAT.25.0040. doi:10.1056/CAT.25.0040 \u003c/li\u003e\n\u003cli\u003eStults CD, Deng S, Martinez MC, et al. Evaluation of an Ambient Artificial Intelligence Documentation Platform for Clinicians. \u003cem\u003eJAMA Netw Open\u003c/em\u003e. 2025;8(5):e258614. doi:10.1001/jamanetworkopen.2025.8614 \u003c/li\u003e\n\u003cli\u003eDoes AI-Powered Clinical Documentation Enhance Clinician Efficiency? A Longitudinal Study | NEJM AI. Accessed September 24, 2025. https://ai.nejm.org/doi/full/10.1056/AIoa2400659 \u003c/li\u003e\n\u003cli\u003eKroenke K, Spitzer RL, Williams JBW, L\u0026ouml;we B. The Patient Health Questionnaire Somatic, Anxiety, and Depressive Symptom Scales: a systematic review. \u003cem\u003eGen Hosp Psychiatry\u003c/em\u003e. 2010;32(4):345-359. doi:10.1016/j.genhosppsych.2010.03.006 \u003c/li\u003e\n\u003cli\u003eL\u0026ouml;we B, Un\u0026uuml;tzer J, Callahan CM, Perkins AJ, Kroenke K. Monitoring depression treatment outcomes with the patient health questionnaire-9. \u003cem\u003eMed Care\u003c/em\u003e. 2004;42(12):1194-1201. doi:10.1097/00005650-200412000-00006 \u003c/li\u003e\n\u003cli\u003eSpitzer RL, Kroenke K, Williams JBW, L\u0026ouml;we B. A Brief Measure for Assessing Generalized Anxiety Disorder: The GAD-7. \u003cem\u003eArch Intern Med\u003c/em\u003e. 2006;166(10):1092-1097. doi:10.1001/archinte.166.10.1092 \u003c/li\u003e\n\u003cli\u003eOehlert GW. A Note on the Delta Method. \u003cem\u003eAm Stat\u003c/em\u003e. 1992;46(1):27-29. doi:10.1080/00031305.1992.10475842 \u003c/li\u003e\n\u003cli\u003eR Core Team. \u003cem\u003eR: A Language and Environment for Statistical Computing\u003c/em\u003e. R Foundation for Statistical Computing; 2022. https://www.R-project.org/ \u003c/li\u003e\n\u003cli\u003eCross S, Bell I, Nicholas J, et al. Use of AI in Mental Health Care: Community and Mental Health Professionals Survey. \u003cem\u003eJMIR Ment Health\u003c/em\u003e. 2024;11(1):e60589. doi:10.2196/60589 \u003c/li\u003e\n\u003cli\u003eHipgrave L, Goldie J, Dennis S, Coleman A. Balancing risks and benefits: Clinicians\u0026rsquo; perspectives on the use of generative AI chatbots in mental healthcare. \u003cem\u003eFront Digit Health\u003c/em\u003e. 2025;7:1606291. doi:10.3389/fdgth.2025.1606291 \u003c/li\u003e\n\u003cli\u003eSadeh-Sharvit S, Camp TD, Horton SE, et al. Effects of an Artificial Intelligence Platform for Behavioral Interventions on Depression and Anxiety Symptoms: Randomized Clinical Trial. \u003cem\u003eJ Med Internet Res\u003c/em\u003e. 2023;25(1):e46781. doi:10.2196/46781 \u003c/li\u003e\n\u003c/ol\u003e"},{"header":"Footnotes","content":"\u003col\u003e\u003cli\u003e\u003cspan\u003e Pairs needed to have 45 days or more within the study period after the first session to be included.\u003c/span\u003e\u003c/li\u003e\u003c/ol\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":true,"hideJournal":true,"highlight":"","institution":"","isAcceptedByJournal":false,"isAuthorSuppliedPdf":false,"isDeskRejected":"","isHiddenFromSearch":false,"isInQc":false,"isInWorkflow":false,"isPdf":false,"isPdfUpToDate":true,"isWithdrawnOrRetracted":false,"journal":{"display":true,"email":"[email protected]","identity":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true},"keywords":"","lastPublishedDoi":"10.21203/rs.3.rs-9589129/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-9589129/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003cp\u003eClinical documentation is a major burden, particularly in psychotherapy where notes must meet clinical, compliance, and billing requirements. Across 850,000 appointments within a nationwide digital mental health benefit, we evaluated the feasibility, adoption, efficiency, and clinical impact of AI-powered notetaking by making these tools available to over 8,700 providers and 189,000 patients. Mutual patient-provider consent was required for session recording and AI summarization, and nearly half of providers and 66.2% of patients consenting regularly. Although providers were more reticent, adoption rates increased steadily over the study period. Those providers who used AI-generated notes completed documentation 25% faster, submitted notes 8.2 hours sooner, and had reduced after-hours work compared to sessions without AI. Integration into workflows was common, with AI-generated text incorporated in over 73% of submitted notes. Importantly, continuity of care was slightly higher for adopters and clinical outcomes—including depression and anxiety symptom trajectories—were equivalent between AI adopters and non-adopters, indicating no adverse impact on patient care. These findings demonstrate that AI-powered notetaking can be scaled in psychotherapy, and its use is associated with improved provider efficiency and quality of life without impacting clinical effectiveness.\u003c/p\u003e","manuscriptTitle":"AI Notetaking in Psychotherapy: Consent, Efficiency, and Clinical Outcomes from a Large-Scale Evaluation","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2026-05-09 00:56:04","doi":"10.21203/rs.3.rs-9589129/v1","editorialEvents":[{"type":"communityComments","content":0}],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true}}],"origin":"","ownerIdentity":"cd1d183d-a09b-4fb3-a595-455dd40f0416","owner":[],"postedDate":"May 9th, 2026","published":true,"recentEditorialEvents":[{"type":"decision","content":"Revision requested","date":"2026-05-06T07:56:52+00:00","index":"","fulltext":""},{"type":"editorAssigned","content":"","date":"2026-05-02T07:38:58+00:00","index":"","fulltext":""},{"type":"checksComplete","content":"","date":"2026-05-02T07:38:38+00:00","index":"","fulltext":""},{"type":"submitted","content":"Scientific Reports","date":"2026-05-01T20:22:05+00:00","index":"","fulltext":""}],"rejectedJournal":[],"revision":"","amendment":"","status":"posted","subjectAreas":[{"id":67405836,"name":"Scientific community and society/Business and industry"},{"id":67405837,"name":"Health sciences/Health care"},{"id":67405838,"name":"Health sciences/Medical research"},{"id":67405839,"name":"Biological sciences/Psychology"},{"id":67405840,"name":"Social science/Psychology"}],"tags":[],"updatedAt":"2026-05-09T00:56:04+00:00","versionOfRecord":[],"versionCreatedAt":"2026-05-09 00:56:04","video":"","vorDoi":"","vorDoiUrl":"","workflowStages":[]},"version":"v1","identity":"rs-9589129","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-9589129","identity":"rs-9589129","version":["v1"]},"buildId":"XKTyCvWXoU3ODBz1xrDgd","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}