A mixed-methods study published in the International Journal of Mental Health Systems demonstrated that integrating a generative artificial intelligence assistant into asynchronous online psychotherapy led to a statistically significant increase in mental health professionals’ perceived competence and a marginal increase in their sense of autonomy. Asynchronous online psychotherapy, or AOP, has emerged as an essential modality for delivering mental healthcare across geographic distances and in crisis situations, including armed conflicts and pandemics, where face-to-face sessions become infeasible. However, therapists practicing within asynchronous digital frameworks frequently encounter substantial operational burdens, including heightened emotional strain, professional isolation, blurred practice boundaries, and elevated risk of occupational burnout. While autonomous generative AI chatbots have gained traction in self-guided digital mental health, technological risks such as diagnostic misinterpretations and conversational hallucinations have demonstrated the ongoing necessity of human clinical oversight. Seeking to support rather than displace human clinicians, the research team examined whether a specialized digital agent could directly empower practicing therapists by augmenting their execution of core psychotherapeutic tasks while preserving human clinical responsibility.
To investigate the effects of generative artificial intelligence on practitioner empowerment, the investigators developed a customized digital agent named APIA, which stands for Asynchronous Psychotherapeutic Intelligent Agent. Constructed on OpenAI’s GPT-3.5-Turbo-0613 and GPT-4-0613 language model architectures, APIA was specifically engineered to assist mental health specialists across three routine clinical tasks: formulating preliminary clinical diagnoses, recommending structured psychotherapeutic treatment plans, and drafting written responses to patient inquiries. The system utilized parallel and sequential API calls to optimize response latency, while strict system prompts restricted the conversation strictly to the domain of psychotherapy and defined the agent’s identity as a virtual assistant for psychotherapy. To mitigate the risk of conversational randomness and prevent model hallucinations, the developers configured the system with a low temperature setting of 0.35. Furthermore, to avoid the leakage of external cues or contextual contamination between experimental tasks, the prompt architecture ensured that the underlying model accessed only the immediate clinical context relevant to each individual scenario presented during evaluation.
The experimental evaluation was conducted with thirteen Ukrainian mental health professionals who possessed active clinical experience in asynchronous online psychotherapy during the ongoing armed conflict. The military conflict in Ukraine had generated substantial internal displacement, extensive client migration, and an acute escalation in cases of depression, anxiety, and posttraumatic stress disorder, thereby accelerating reliance on remote, asynchronous delivery models. Despite these challenging circumstances, the core clinical tasks assessed within the study—formulating diagnoses, constructing treatment strategies, and composing written therapeutic correspondence—remained fully consistent with established international and European psychotherapeutic standards. The participant cohort comprised eleven women and two men, with a mean age of 38.77 years, a median age of 38 years, and an average of 13 years of professional psychotherapeutic experience. Seven practitioners held doctoral degrees in psychology, and nine specialized clinically in anxiety, depression, or trauma-related conditions. At the time of data collection, seven participants resided in Ukraine, five were based in Switzerland, and one lived in the United Kingdom, all maintaining regular asynchronous therapeutic contact with clients across Europe.
The empirical assessment followed a within-subject, self-guided experimental protocol completed on a secure, university-hosted online survey platform where the chat interface of APIA was embedded. Each participating specialist completed two randomized clinical sessions to enable rigorous within-subject comparison: one session in which they completed therapeutic tasks entirely independently, and another session in which they collaborated interactively with the APIA digital agent. The evaluation utilized two authentic, anonymized real-world clinical patient scenarios and six distinct patient questions selected by a co-author who is a practicing psychiatrist and psychotherapist. All personal and identifying details were removed from the case materials, and no live patient interaction occurred during the evaluation, thereby excluding the protocol from requiring formal institutional ethical approval. Prior to testing, participants viewed an instructional video demonstrating navigation and interface interactions. During the experimental sessions, participants formulated preliminary diagnoses, devised therapeutic interventions, and composed answers to patient questions, having the opportunity in the AI-supported session to cross-examine hypotheses, review generated drafts, and refine their final submissions.
To capture changes in practitioner empowerment, the researchers assessed the cognitive dimensions of competence and autonomy following each experimental session using a five-point Likert scale adapted from the Technology-based Experience of Need Satisfaction–Task instrument, known as TENS-Task. Quantitative outcomes were analyzed through descriptive statistics, chi-squared tests for distributional normality, Cronbach’s alpha consistency metrics, and paired-sample t-tests. Scale reliability was confirmed with Cronbach’s alpha values of 0.75 for competence and 0.80 for autonomy. The quantitative analysis demonstrated that working with APIA produced a 12.7 percent increase in mean perceived competence, reflecting a statistically significant positive effect with a two-tailed p-value below 0.01 and a small effect size indicated by a Cohen’s d of 0.22. In terms of clinical autonomy, mean scores increased by 8.3 percent; a one-tailed paired t-test indicated a marginally significant positive effect with a p-value below 0.1 and a very small effect size of Cohen’s d equal to 0.06, whereas the corresponding two-tailed test did not achieve conventional statistical significance.
To complement the quantitative measurements, the researchers gathered extensive qualitative data from open-ended survey responses and conducted follow-up semi-structured interviews lasting between 30 and 45 minutes with ten of the thirteen clinicians over Microsoft Teams. The qualitative corpus was analyzed thematically through a bottom-up coding process utilizing MaxQDA software by two independent researchers, yielding a structured codebook of 52 distinct thematic codes. The qualitative findings illuminated the mechanisms underpinning the observed gain in perceived competence. Clinicians highlighted that APIA provided rapid, well-structured, and neutral drafts that significantly accelerated their response times and reduced cognitive strain during written communication. Many specialists utilized the digital agent as an immediate validation mechanism, cross-checking their clinical hypotheses against the model’s outputs. Confirming that the artificial agent’s structured reasoning aligned with their own clinical intuitions reinforced clinicians’ self-assurance, fostered reflective practice, and served as competence-promoting feedback during isolated asynchronous workflows.
Nevertheless, participant accounts also revealed clear conceptual boundaries regarding professional competence. Several experienced clinicians asserted that genuine psychotherapeutic capability arises from years of clinical practice, seasoned intuition, and deep emotional empathy—domains in which algorithmic systems remain inherently deficient. As one specialist observed, the digital agent treats the human individual primarily as an object of analysis, whereas the psychotherapist engages the patient as an intentional, subjective human being. Consequently, practitioners viewed the digital agent not as a substitute for human clinical intellect, but as an auxiliary instrument that provides rational intelligence and rapid information synthesis to complement the therapist’s emotional intelligence. Clinicians noted that blending algorithmic rationality with human relational empathy enabled complementary performance while counteracting the documented risks of standalone language models, which have previously been shown to misinterpret complex psychiatric states or downplay critical risks such as schizophrenia and suicide.
Regarding autonomy, qualitative feedback clarified why the quantitative effect was more modest. Because APIA functioned strictly as a passive, consultative assistant without direct therapeutic agency, clinicians retained complete control over whether to adopt, modify, or reject any generated suggestion. Several practitioners explained that their clinical autonomy is rooted in their internal understanding of human affective dynamics and immediate therapeutic opportunities, areas untouched by the software. However, participants noted that the assistant significantly enhanced their functional independence during moments of clinical uncertainty. In routine practice, asynchronous therapists must often consult senior supervisors or await peer review to resolve diagnostic ambiguities. By providing immediate rational counter-perspectives, APIA allowed specialists to independently resolve clinical impasses without waiting for formal external supervision, thereby expanding their operational flexibility and self-reliance.
Drawing upon the empirical observations and existing service literature, the researchers delineated three distinct architectural archetypes for integrating generative digital agents into asynchronous online psychotherapy: psychotherapist-centric, patient-centric, and therapy-centric. In the psychotherapist-centric archetype, the digital agent operates strictly as a backstage assistant supporting the clinician. There is no direct communication between the software and the client. The agent functions before, during, or between therapeutic interactions as an information retrieval tool, summarizing past sessions, organizing longitudinal records, or providing theoretical perspectives from established psychological frameworks. This model minimizes clinical and ethical risk by keeping all AI outputs behind the clinician’s evaluative barrier, allowing the therapist to formulate high-quality responses while maintaining an unmediated personal connection with the patient.
In the patient-centric archetype, the digital agent interacts directly with the patient on the frontstage under the therapist’s active clinical supervision. While the human therapist remains the primary therapeutic authority and manages the treatment alliance asynchronously, the patient can engage the digital agent synchronously between sessions for psychoeducation, symptom tracking, or structured cognitive-behavioral homework exercises. The system synthesizes patient interactions and provides analytical summaries back to the supervising clinician, closing the diagnostic feedback loop. This configuration expands the availability of continuous support, but investigators emphasized that system developers must carefully address secondary mental models to ensure that patients do not experience depersonalization, develop misplaced emotional reliance, or encounter unvetted therapeutic recommendations.
The third archetype, therapy-centric integration, positions the digital agent as an omniscient background co-therapist and supervisor. In this framework, the agent is continuously aware of the ongoing asynchronous message exchange between patient and clinician. Operating from a meta-position without directly intervening in client dialogue, the agent assists the clinician by monitoring longitudinal progress, generating automatic session summaries, highlighting potential cognitive errors, and identifying subtle transference dynamics where clients redirect unresolved emotions toward the therapist. By acting as a reflective sounding board or Socratic opponent, the therapy-centric agent provides systemic oversight that mitigates therapist isolation, supports complex decision-making, and assists clinicians in maintaining objectivity amidst demanding therapeutic cases.
Despite the recognized operational advantages, the study highlighted significant clinical, technical, and ethical limitations associated with generative AI integration. Participants raised serious concerns regarding the confidentiality of sensitive patient data, noting that the deployment of large language models requires stringent data protection mechanisms that comply with regulatory standards such as the General Data Protection Regulation. Methodological compatibility emerged as another critical limitation; psychoanalytically oriented clinicians noted that providing automated psychoeducational explanations disrupts fundamental therapeutic mechanisms such as free association. Furthermore, specialists cautioned that inexperienced or junior practitioners might over-rely on algorithmic recommendations, unthinkingly adopting flawed diagnostic outputs without adequate critical appraisal. Finally, researchers emphasized the exploratory nature of the study, noting that its sample was restricted to thirteen Ukrainian professionals evaluating simulated clinical tasks, and called for broader investigations with larger, cross-cultural cohorts examining live implementations in diverse clinical settings.
Subject of Research: Empowering mental health professionals in asynchronous online psychotherapy through generative artificial intelligence
Article Title: Empowering mental health professionals in asynchronous online psychotherapy with GenAI
Article References: Bucher, A., Vashkite, I., Staehelin, D., Dolata, M., & Schwabe, G. (2026). Empowering mental health professionals in asynchronous online psychotherapy with GenAI. International Journal of Mental Health Systems, 20(1), Article 10. https://doi.org/10.1186/s13033-026-00700-5
Image Credits: AI Generated
DOI: 10.1186/s13033-026-00700-5
Keywords: asynchronous online psychotherapy, generative artificial intelligence, digital agents, APIA prototype, therapist empowerment, perceived competence, clinical autonomy, mental healthcare technology, mixed-methods evaluation, psychotherapy workflows, large language models, blended care, psychotherapist-centric archetype, patient-centric archetype, therapy-centric archetype, clinical decision support, technology-based need satisfaction, mental health practitioner burnout
News Source: Glenn Wilkins. (November 20, 2026). Generative AI Assistant Enhances Therapists’ Perceived Competence in Asynchronous Online Psychotherapy. Scienmag.



