The use of CAs in chronic respiratory diseases is still emerging, with few robust experimental studies published to date. Therefore, a scoping review was conducted to map current and emerging literature. The review followed the Arksey and O’Malley framework: identifying the research question; identifying and selecting relevant studies; charting the data; and collating, summarizing, and reporting the results [22]. Additionally, the review adheres to the PRISMA-ScR (Preferred Reporting Items for Systematic Reviews and Meta-Analyses extension for Scoping Reviews) guidelines [23] (Checklist 1) and was prospectively registered on the Open Science Framework [24].

This review aims to answer the following questions:

A systematic search was conducted across 9 databases (CINAHL, CENTRAL, Embase, IEEE Xplore, PubMed, ProQuest Theses & Dissertations, Scopus, Web of Science, and Google Scholar) for articles published from January 1, 2014, to February 22, 2025. Reference lists of relevant reviews and included studies were also manually screened. For Google Scholar, the first 200 records ranked by relevance were screened, a pragmatic approach commonly applied in gray-literature searching [25], and consistent with recommendations to focus on the first 200 to 300 Google Scholar results when the platform is used to supplement gray-literature searching [26]. ProQuest Theses & Dissertations was searched in full as the principal gray-literature source. References published after the search closing date are cited in the Introduction and Discussion sections for background and contextual purposes only. They were not included in screening, eligibility assessment, data extraction, or synthesis.

The full search strategy was developed using the Peer Review of Electronic Search Strategies (PRESS) checklist [27] and refined in consultation with a research librarian. Truncations (*) and Boolean operators (AND, OR) were used to maximize search results. Keywords and MeSH terms such as “chatbots,” “conversational agents” were truncated and exploded. Search strategies for each database are presented in Multimedia Appendix 2.

The first author (NBH) used the Joanna Briggs Institute (JBI) Data Extraction form to extract key information, including authors, year of publication, publication status, study location, intervention duration, study populations, methodology, key results, and features of CAs. A second reviewer (SQY) independently extracted data from a random sample of 20% reports (4/16) for verification as a pragmatic compromise given resource constraints, and the remaining extractions were checked against the source articles by the coauthors, with discrepancies resolved through discussion.

Where multiple reports originated from the same underlying study, the study was treated as the unit of analysis for participants, design, and setting. Reports were counted separately only for publication counts. Sloots et al [31] and ter Stal et al [32] report on the same cohort of 11 participants, and these participants are counted once. In the narrative synthesis, findings from these 2 reports were treated as one evidentiary contribution rather than independent corroboration.

Consistent with scoping review methodology, findings were mapped descriptively across study designs and stages of development rather than pooled, as the aim was to characterize the breadth of an emerging evidence base. An inductive content analysis approach [33] was used to synthesize the findings. For quantitative studies, relevant results sections were analyzed. For qualitative or mixed method studies, analysis included authors’ interpretations, participant quotations, and the interpretation of qualitative findings in the discussion. A 4-step process (decontextualization, recontextualization, categorization, and compilation) was used to synthesize and organize findings into categories and subcategories, which a second reviewer (SQY) subsequently reviewed. Tables and figures were created to support and illustrate findings. As a scoping review aims to outline existing literature, no quality appraisal of studies was conducted. However, to provide a descriptive overview of the evidence base, we recorded each study’s design, sample size, comparator presence and type, exposure duration, and use of clinically validated outcome instruments.

The Patterns, Advances, Gaps, Evidence for Practice, and Research Recommendations (PAGER) framework guided the discussion and critique of the included studies [34]. We adopted these reflective questions:

The search yielded 6275 citations, from which 2204 duplicates were removed. Following title and abstract screening, 27 articles were retrieved for full-text assessment and 16 reports from 15 studies [29-32,35-46] were included in this review (Figure 1).

Table 1 presents the characteristics of the included studies. The studies spanned 9 countries, including Switzerland (n=4) [35,39,41,42], the United Kingdom (n=3) [36,37,43], the United States (n=2) [40,44], the Netherlands (n=2; Sloots et al [31] and ter Stal et al [32], which report the same study), Denmark (n=1) [30], Saudi Arabia (n=1) [29], France (n=1) [38], Portugal (n=1) [45], and one additional study [46] conducted jointly in Denmark and Norway. A total of 3 [32,40,45] reports were conference papers and the remainder peer-reviewed journal articles.

The study designs varied, including cross-sectional questionnaire (n=1), mixed methods (n=3), qualitative (n=2), quantitative (n=2), randomized controlled trial (n=1), single-arm quasi-experimental (n=1), and single-arm quasi-experimental collecting qualitative and quantitative data (n=6), summing to the 16 reports. Most studies were pilot feasibility or developmental studies, reflecting the nascency of CA application in chronic respiratory disorders.

A total of 860 participants took part across the included reports. This included people with asthma or COPD, caregivers, and parent-child dyads, health care professionals, and researchers. Kadariya et al [40] tested kBot with 8 asthma clinicians and 8 researchers, Easton et al [37] included 5 health care professionals among 23 participants, Høj et al [30] involved 5 medical professionals only, and Quinde et al [43] included 2 experts among 10 participants. Alabdulmohsen et al [29] did not involve human participants, instead evaluating 30 questions posed to ChatGPT (OpenAI). Sloots et al [31] and ter Stal et al [32] report the same 11 participants, who were counted once. Participant numbers were largely driven by 3 studies: Gross et al [39] (n=240), Gonsard et al [38] (n=194), and Cook et al [36] (n=150).

The findings were divided into 4 categories: (1) roles of CAs, (2) direct patient outcomes, (3) indirect patient outcomes, and (4) evaluation of the LLM output. The summary of synthesis findings is presented in Table 2 (detailed synthesis findings can be found in (Multimedia Appendix 3).

CAs supported patient needs across 3 key roles. First, they facilitate information dissemination and patient education [30-32,36-39,42,46]. Some provided on-demand responses to patient queries, while others delivered structured, tailored educational content to improve disease understanding and self-management [41,42]. Second, several agents monitor patient data, including physiological parameters, self-reported health status, and validated patient-reported outcome measures. Third, CAs assumed a coaching role by helping patients identify potential triggers, estimating exacerbation risk, and delivering timely reminders for medication adherence and physical activity [31,32,35-38,41-44,46]. Some CAs extend their coaching capabilities to emotional and psychosocial domains. By offering motivation, companionship, and empathy-driven dialog, they empower patients to maintain positive self-management behaviors and reduce the psychological burden of chronic disease [36,42,44].

Direct patient outcomes are quantifiable indicators of disease management (behavioral changes) and disease progression (clinical outcomes). A total of 6 reports described behavioral outcomes, such as lifestyle modifications and increased self-management [31,36,39,41,42,44]. Reported changes included improved inhaler technique [42], increased physical activity, and improved adherence to prescribed medications and treatment plans [44]. Participants in one study also reported greater empowerment and proactivity in managing their conditions [44]. All of these observations arise from single-arm or uncontrolled designs. Sloots et al [31] observed partial engagement, in which participants initiated components such as maintaining daily diaries but did not follow through with key recommended actions, including contacting case managers or demonstrating correct inhaler technique. The companion report by ter Stal et al [32], analyzing the same 11-participant cohort, described deficiencies in character design and responsiveness.

Only 4 studies measured clinical outcomes, including asthma control, symptom burden, functional exercise capacity, and lung function [31,36,41,42]. Of these, one was a randomized controlled trial (n=30), in which 21 participants were allocated to the intervention and 9 to the control group [41]. The remaining 3 studies did not include a control or comparison group. The single statistically significant clinical result was an 8.38% improvement in Asthma Control Questionnaire score reported by Cook et al [36] over a 28-day single-arm study without a comparator. Sloots et al (2021) [31] reported mixed findings. Taken together, these findings suggested that the current evidence does not provide a robust basis for drawing conclusions regarding the effects of CAs on clinical outcomes in asthma or COPD.

Indirect patient outcomes capture the practical aspects of CAs’ usage, including user satisfaction and engagement. Usability and feasibility were the most frequently reported outcomes across the reviewed studies [31,32,35-46]. Two studies [40,44] reported that participants described CAs as convenient, easy to use, and helpful in delivering accessible health information, and that participants valued immediate responses and interactive, personalized communication.

Other studies [32,36,45] reported substantial usability problems, including poor conversational flow, limited adaptability, and instances in which the CA failed to understand user inputs or intent, which participants linked to frustration and reduced engagement.

Reports on the trustworthiness and credibility of CAs varied. Some participants described reluctance to follow CA-generated advice without consulting a health care professional [32], and some were hesitant to recommend the agent to others or rely on it for decisions about their care [38]. Other participants described the advice as helpful and were open to incorporating the agent into their self-management routines [36].

Participants expressed preferences regarding agent functionality and design, including gender, appearance, interaction style, and communication tone [32,36,37,39,43,45]. Visual design was reported to influence user acceptance, with participants in 1 study [32] favoring realistic representations such as a professionally dressed nurse over cartoon-like figures, which were described as more credible and authoritative. Reported gender preference favored female agents [32] or agents matching the participants’ own gender [42]. In terms of communication style, participants preferred a casual, empathetic tone over formal or clinical language [45].

From a functionality perspective, users valued tailored support features, including text-based interaction [32,37], multilingual support [35,45], and disease-tracking and management features [37,43]. These perspectives often varied by patient characteristics, including age, COPD severity, and recency of diagnosis [35,45], highlighting the importance of aligning CAs’ design with user demographics, communication preferences, and perceived authenticity to enhance engagement and trust.

A total of 2 [29,30] reports evaluated LLM output for respiratory disease care. Alabdulmohsen et al [29] posed 30 frequently asked asthma questions to ChatGPT across 2 devices in 2 locations and reported consistent and reproducible responses, while noting that readability exceeded recommended levels for patient-facing materials. Høj et al [30] had 5 specialists in allergic and respiratory disease rate ChatGPT responses on asthma and reported generally accurate content. These studies therefore provide evidence on the accuracy, consistency, and readability of LLM-generated asthma information under controlled test conditions, but they do not examine conversational interaction, patient use, trust, engagement, or support for self-management in community settings. For this reason, their findings are treated separately from studies of patient-facing CAs and were not used to support claims about patient experience, agent role, or implementation in community-based condition management.

The systems described in the reviewed studies extend beyond traditional symptom monitoring and tracking to include emotional and psychosocial support [35,37] and the promotion of physical activity [31,32,45]. This shift suggests that developers are increasingly treating psychosocial and behavioral domains as part of chronic respiratory care alongside disease monitoring, reflecting a more holistic and preventive care approach.

Managing chronic conditions is often stressful and associated with poorer mental health [50]. CAs have been proposed as virtual companions offering support, encouragement, and reassurance, with anonymity providing a nonjudgmental avenue for users to express concerns. Prior research has investigated CAs for mental health support [47,51], and similar approaches are being adapted for chronic respiratory disease management. However, in the current review, none of the included studies measured psychological outcomes using validated instruments, so this role is described but not evaluated. Questions also remain about the authenticity of such interactions and their capacity to establish a meaningful therapeutic relationship.

Despite growing interest in the use of CAs in health care, few studies have evaluated their impact on direct patient outcomes in asthma and COPD management, and those that have done so were not designed to support causal inference. It therefore remains premature to conclude whether these technologies lead to measurable improvements in clinical indicators. This observation aligns with previous reviews, which identified the scarcity of studies assessing the clinical impact of CAs and the absence of rigorous randomized controlled trials supporting effectiveness in disease management [48,52]. One plausible explanation is that much of the existing literature is pilot or developmental in nature, prioritizing feasibility, usability, and patient perceptions over clinical endpoints. Another possible explanation could be that the short exposure durations typical of this literature may not be sufficient to detect changes in outcomes such as exacerbation frequency or lung function. Both reflect the early and exploratory phase of integrating CAs into chronic disease care [52].

While several studies reported high usability and feasibility scores for CAs, these findings derive largely from small, self-selected samples in single sessions and may not reflect the needs of all users, particularly older adults with COPD. A systematic review of digital health adoption among older adults with chronic disease identifies limited digital literacy, physical and cognitive challenges, infrastructural deficits, usability problems, and mistrust as recurring barriers, with older women showing lower adoption [53]. A longitudinal pilot study of a COPD app for older adults found that technology-related barriers decreased with familiarity, whereas perceived lack of relevance and competing health concerns persisted [54]. Language complexity may pose further difficulty for users with lower health literacy, a concern reinforced by the readability problems documented in the LLM evaluations [29]. Health care providers should account for specific users’ literacy, digital competencies, and engagement when considering CAs [55].

Trust was a recurring concern, and reports of hesitancy were more common for text-based and embodied agents than for a single voice-based agent in this review. Trust in conversational systems is consistently associated with perceived social presence and with perceived warmth and competence, and is undermined by communication delays [56-58]. Voice interfaces can convey tone, timing, and natural turn-taking cues that text cannot, making them richer in these aspects. They may also reduce access barriers that are especially relevant for people with COPD, many of whom are older adults, by avoiding difficulties related to small text, typing accuracy, and visual acuity [53]. Nevertheless, this review included only one voice-based agent, evaluated in 4 participants [35]; therefore, future studies should further examine whether the text-versus-voice distinction remains relevant in chronic respiratory care.

Across the wider literature, many users may still see CAs as less trustworthy than clinicians for complex needs. Clearly communicating the limits of these systems, including that they are not substitutes for professional advice, is essential for setting realistic expectations. Across health care apps, a hybrid model has been proposed in which the agent acts as a first point of contact, with structured escalation to a health care professional when needed [59], and it has been evaluated in mental health, where an AI-delivered program with clinician oversight achieved outcomes comparable to face-to-face cognitive behavioral therapy, while reducing clinician time by up to eightfold [60]. Clinician endorsement and hybrid care models have also been identified as facilitators of digital health adoption among older adults with chronic disease [53]. No included study evaluated such a model in asthma or COPD, so future research should test whether this approach is feasible and effective in chronic respiratory care.

A notable descriptive finding is that users expressed specific preferences regarding CA’s characteristics, including gender, age, appearance, interaction styles, and communication tone. Consistent with prior work, younger and female agents were generally preferred over older and male agents [61]. Possible explanations include that female agents may be perceived as more empathetic or nurturing because of wider social stereotypes about gender [62]. Nevertheless, the preference findings in this review are based on only 2 reports [32,42], and neither study tested the reasons behind these preferences. These findings should therefore be treated as design signals for future testing, rather than as evidence of broader societal attitudes. Personalization may improve comfort and sustained use, but it should be implemented carefully to avoid reinforcing stereotypes or introducing design bias. An inclusive and authentic design approach may help improve both user experience and ethical practice.

As patients may increasingly use LLMs for treatment information or self-management advice, a parallel body of work has evaluated the accuracy of LLMs in answering standardized respiratory questions after the search closing date in this review [15-17,63,64]. Like the 2 LLM evaluations included in our review, they assess model output under test conditions rather than patient interaction, intervention delivery, or self-management support. Collectively, they indicated that LLM responses are often accurate, but readability, understandability, and performance on more complex scenario-based tasks remain concerns. Future work should examine how patients actually use LLMs for respiratory self-management and whether this use is safe, understandable, and clinically appropriate.

A key strength of this review is its position as the first to map roles and outcomes of CAs in asthma and COPD management. Its methodological strengths include a broad search across 9 databases, the inclusion of gray literature, and the use of the PAGER framework.

However, several limitations should be noted. Restricting the review to English-language publications reduced cultural and linguistic diversity. Most included studies were conducted in Western countries; this pattern may partly reflect the English-language restriction, which could have led to the underrepresentation of research from Eastern and low- and middle-income countries. This limits the generalizability of the findings to Eastern and resource-limited contexts. Data were extracted by a single author (NBH), with verification by coauthors (SQY and YJ), rather than through full independent duplicate extraction. This approach may have introduced a risk of extraction error. Only one randomized controlled trial [41] with a small sample was identified, highlighting the lack of robust comparative evidence. Additionally, many studies were still in the formative phase, focused on development or preliminary testing. Consequently, further large-scale, rigorous research is needed before these findings can inform routine clinical practice or policy.

This scoping review maps the roles that CAs have been designed to perform in asthma and COPD, namely patient education, symptom monitoring, and psychosocial support. Patients are the end users of these systems, and this review shows that problems at the point of interaction, including poor conversational flow, limited responsiveness, character design, and lack of personalization, were repeatedly associated with disengagement. In one project, these problems were also linked to noncompletion of the recommended self-management actions the agent was designed to prompt [31,32]. Patients should therefore be central to co-design from the outset, through sustained involvement across iterative development cycles rather than one-off consultation at the requirements stage. Several included studies did involve patients or caregivers in co-design [37,45,46], and these provided the most detailed and actionable insights into user requirements in this review.

Health care professionals also have an important role. The functions mapped in this review, including education, monitoring, and psychosocial support, align closely with nursing responsibilities in chronic respiratory care [65]. Nurses are therefore well placed to contribute to system design, guide patients in use, support digital literacy, and help interpret CA-generated information [66]. Doing so would require structured training and institutional support [67].

Implementation also depends on factors beyond the individual patient–agent interaction, many of which were largely absent from the included studies. These include reimbursement and procurement pathways, medical device and AI regulation, interoperability with electronic health records, digital infrastructure and connectivity, workforce training, and governance of patient data. Evidence from policy-driven digital health interventions suggests that these structural and policy conditions often determine whether clinical and system-level benefits are realized, regardless of the tool itself [68].

Privacy and data security require particular attention. Information collected by CAs must comply with relevant privacy regulations and ethical standards [69], and weaknesses in technical safeguards or the use of unregulated “shadow IT” may reduce acceptance, especially among younger users who are more skeptical of data harvesting [70]. Mixed user experiences and concerns about trust also point to the need for clinical oversight to ensure that CA-generated information is accurate and appropriate [71].

Nursing and clinical leadership will therefore be important in shaping the policies, governance arrangements, and training needed for safe and equitable use. Overall, the gap between the developmental studies reviewed here and real-world deployment remains substantial, and depends as much on policy and infrastructure as on the technology itself.

Most included studies were early-phase work, pilot, or developmental in design, limiting evidence to short-term feasibility outcomes and offering little insight into whether behavioral changes persist or translate into meaningful clinical improvements over time. Future research should prioritize longitudinal, rigorously designed trials to evaluate the sustained effectiveness of CAs in chronic respiratory care. Specifically, randomized controlled trials with adequate sample sizes, active comparators, and clinically meaningful endpoints are urgently needed. Adopting mixed methods frameworks in future work would further enable the integration of quantitative effectiveness data with qualitative insights into the patient experience, producing a more complete and actionable evidence base.

There is a pressing need to incorporate direct clinical outcome measures into CA research on chronic respiratory disease management. In the absence of evidence that CAs improve disease control, reduce exacerbations, or enhance medication adherence, their routine clinical use cannot be recommended. Future research should therefore prioritize validated disease-specific endpoints such as the Asthma Control Test (ACT), the COPD Assessment Test (CAT), or spirometric measures, and should report sample size calculations against those endpoints. To enable cross-study synthesis and meta-analysis, future work should align with established frameworks such as the Core Outcome Measures in Effectiveness Trials (COMET) initiative [72] to define and adopt a minimum core outcomes set for CA research in chronic respiratory disease.

A critical gap identified in this review is the near-total absence of research from low- and middle-income countries, despite their disproportionate burden of obstructive lung disease and rapidly growing interest in digital health solutions [1]. As noted in the limitations, our English-language restriction may have contributed to this gap. Future work should assess the feasibility, safety, and acceptability of CAs across diverse socioeconomic and health system contexts, with explicit attention to infrastructure, language, health literacy, and cultural norms. Older adults, who represent a large proportion of the COPD population and who demonstrated more variable and cautious attitudes toward CA use in this review, also remained understudied [53].

Co-design should place patients first. Given the technical and clinical complexity of CA implementation, co-design with end users, namely patients and caregivers, alongside health care professionals, including nurses, should be embedded in the development cycle from the outset rather than pursued retrospectively. This review found that deficiencies in character design, conversational flow, and responsiveness were frequently linked to reduced engagement and less favorable patient outcomes, underscoring the value of iterative, user-centered design processes. Future studies should document and report their co-design methodology transparently, including who was involved, at which stages, and how their input changed the resulting system, allowing replication and knowledge transfer across settings.

In summary, this review mapped the use of CAs in the management of asthma and COPD and found that the current evidence base remains too limited to support conclusions about clinical benefit. Across the 15 included studies [29-32,35-46], the evidence base was mainly small feasibility, developmental, and pilot studies. Descriptively, CAs have been developed to support education, health data collection, and psychosocial support, while user-experience factors such as conversational flow, responsiveness, trust, and personalization consistently emerged as important for engagement. Future studies should be adequately powered, longitudinal, and use clinically meaningful, standardized outcomes across diverse settings, with patient input and clinical oversight. Until then, CAs in asthma and COPD remain a promising research area rather than tools ready for routine care. AI should complement, not replace, human relationships in chronic respiratory care.