Abstract
Artificial intelligence (AI) coaching offers scalable, on-demand support for professional development, yet research on its relational and process-oriented dimensions remains limited. This pilot study explored the design of a Cognitive Behavioural Coaching (CBC) AI chatbot aimed at enhancing resilience, drawing on expert feedback to assess theoretical and practical alignment. Using unstructured questionnaires and thematic analysis, findings indicate the chatbot effectively facilitated structured reflection, goal-setting and guided discovery, creating a safe and accessible reflective space. Limitations were noted in supporting deep cognitive-emotional work and relational nuance. The study highlights AI coaching's potential as a complementary tool and informs future design and research directions.
Keywords
Introduction
Artificial intelligence (AI) coaching is gaining increasing attention due to its potential to provide continuous, on-demand support for reflection and development (Arakawa & Yakura, 2024). AI coaching is defined as a structured, machine-facilitated process that supports clients in setting professional goals and generating solutions to achieve them (Graßmann & Schermuly, 2021). Although AI does not yet match human intelligence, it offers substantial potential to scale coaching and deliver basic support at significantly lower cost (Terblanche et al., 2022). The emergence of large language models (LLMs) and generative AI tools marked a major advance in natural language processing, enabling more sophisticated and human-like coaching chatbots (Ahmed et al., 2023; Terblanche, 2025).
Despite extensive debate in the broader AI literature, research on AI in coaching remains limited and primarily focused on efficacy and goal attainment (Bachkirova & Kemp, 2025). Few studies have examined relational and process-oriented dimensions such as insight generation, trust, confidentiality and resilience (Passmore & von Bartheld, 2024; Schiemann et al., 2019). Existing evidence often relies on student samples and health-related outcomes, with limited exploration of workplace coaching or complex goals requiring deep self-reflection (Arakawa & Yakura, 2024; Passmore et al., 2026b).
Contemporary organisational contexts are characterised by complexity and escalating leadership demands, making resilience a critical personal resource (Groysberg & Abbott, 2020; Rees et al., 2015; Visagie et al., 2016). Coaching is widely recognised as an effective method for building resilience due to its personalised, reflective and supportive nature (Bachkirova et al., 2010; Bozer & Jones, 2018; Lawton Smith, 2017; Vanhove et al., 2016), with Cognitive Behavioural Coaching (CBC) identified as the most effective approach for resilience coaching (Sipondo & Terblanche, 2024). CBC is a collaborative, results-focused approach that effectively links goal attainment with behaviour change through structured problem-solving, solution-focused strategies and goal setting (Carvalho et al., 2018). Its key strength lies in integrating external goal demands with internal cognitive barriers, promoting sustained self-regulation, long-term functioning and well-being rather than short-term fixes (Theeboom et al., 2014).
Although AI coaching chatbots use a variety of theoretical approaches, there is limited evidence on which methods are most effective for specific developmental outcomes, and research has largely focused on efficacy rather than coaching processes. Drawing on human coaching research that identifies CBC as particularly effective for resilience, this study addresses the gap by using CBC as the theoretical basis for a resilience-focused AI coaching chatbot. In this pilot the following aspects of this chatbot were explored:
To what extent does the AI coaching chatbot adhere to the principles of CBC? To what extent does the AI coaching chatbot align with established principles of effective resilience coaching? How does the AI coaching chatbot compare with human coaching in terms of efficacy and alignment with resilience coaching standards?
The study contributes to the AI coaching literature by clarifying current capabilities and limitations of AI coaching chatbots in the context of CBC and offering design-relevant insights to support the development of more effective, theoretically grounded AI coaching systems.
Literature Review
AI Coaching
Artificial intelligence (AI) coaching is attracting growing interest due to its capacity to provide continuous, on-demand support for reflection and development (Arakawa & Yakura, 2024). The emergence of LLMs and generative AI tools such as ChatGPT and Bard in 2022/2023 marked a significant advance in natural language processing, enabling more sophisticated, adaptive and human-like dialogue in coaching chatbots (Ahmed et al., 2023; Terblanche, 2025). LLM-based systems generate flexible natural language responses that move beyond rule-based constraints, supporting greater personalisation, reduced repetition and increased novelty in interactions (Carlbring et al., 2023; Huang et al., 2025; Lee et al., 2023). GPT-based chatbots are typically general-purpose, capable of producing innovative, human-like language and accessible through intuitive natural language interfaces (Plotkina & Sri Ramalu, 2024).
Systematic reviews indicate that AI coaching can be effective and well accepted for specific functions, particularly goal attainment, psychological support and structured reflection, with performance in some cases comparable to human coaches (Passmore et al., 2026a; Plotkina & Sri Ramalu, 2024). Empirical evidence suggests that certain AI coaches may function effectively as complements or alternatives to human coaching, with some models outperforming both human coaches and GPT-4 on selected criteria (Brown et al., 2025). Further, Passmore et al. (2026a) reported that an AI coaching chatbot met the International Coaching Federation (ICF) Associate Certified Coach (ACC) competency standards, while other studies show that AI coaches can provide feedback, support goal tracking and act as accountability partners, and are often perceived as accessible, convenient and psychologically safe (Passmore et al., 2026b; Terblanche et al., 2023, 2024).
Despite the availability of general design guidelines for AI coaching chatbots (e.g., Terblanche, 2020), research examining how interaction modalities influence coaching effectiveness remains limited (Terblanche & Prywes, 2025). Successful deployment of AI coaching requires careful design and implementation oversight (Passmore et al., 2026b), particularly as most current systems rely on narrow AI and expert-system approaches that simulate human coaching skills (Terblanche et al., 2022, 2024). Further research is needed to determine whether coaching approaches effective with human coaches translate effectively to AI contexts, and whether AI-specific coaching frameworks are required (Barger, 2025). Given current limitations, focusing AI coaches on well-defined topics and outcomes, rather than broad, open-ended coaching agendas, appears more feasible (Terblanche, 2020; Passmore et al., 2026a), underscoring the need for context-specific research into AI coaching design and application (Terblanche, 2025).
Resilience Coaching
Research consistently demonstrates that coaching is an effective intervention for enhancing resilience, wellbeing, and health outcomes (Moore & Jackson, 2014). It is particularly effective in supporting leaders and managers to build resilience in demanding organisational contexts (Jouali et al., 2024) and has been shown to outperform resilience training interventions by providing a confidential, reflective space that supports proactive skill development (Vanhove et al., 2016). Syntheses of the evidence indicate that coaching positively influences employees’ work-related experiences and psychological wellbeing, including resilience and coping capacity (Theeboom et al., 2014). Empirical studies further show that coaching interventions, such as solution-focused coaching, can strengthen resilience, including among middle managers facing complex workplace demands (Bennett & Lemoine, 2014; Grant et al., 2009; Sherlock-Storey et al., 2013).
A scoping review by Sipondo and Terblanche (2024) identified core processes underpinning effective resilience coaching, including the development of psychological resources and a strong coach–coachee relationship that fosters trust, cognitive and emotional shifts and goal-directed behaviour. Grounded in psychological theory, coaching uses targeted questioning and reflective techniques to enhance self-awareness, accountability and behavioural change (Cardwell & Wright, 2023). Through structured reflective dialogue, coachees challenge limiting beliefs, reframe stressors, and independently identify adaptive actions, thereby strengthening resilience and coping capacity (Joe, 2025; Jouali et al., 2024).
Despite robust evidence for human-delivered resilience coaching, empirical research on AI-based resilience coaching remains limited. To date, to our knowledge only Ellis-Brush (2021) has reported improvements in self-resilience using an AI coaching chatbot grounded in cognitive behavioural techniques; however, this study's methodological rigour has been questioned (Passmore et al., 2026a). No empirically tested, theory-driven AI coaching interventions have been identified that explicitly inform the design of AI coaches for resilience, highlighting a clear need for further research to define and empirically examine the design elements required for effective AI-supported resilience coaching.
Cognitive Behavioural Coaching
CBC applies cognitive-behavioural interventions within a coaching context, grounded in well-established CBT principles and methods (Hultgren et al., 2013; Lungu et al., 2021). CBT posits that thoughts influence behaviours, can be identified and modified, and that changing thought patterns leads to desired behavioural outcomes (Minzlaff, 2018). CBC supports clients in overcoming negative thoughts, unhelpful behaviours and distressing emotions, enhancing emotional regulation, cognitive flexibility and adaptive action towards personal goals (David & Cobeanu, 2016; Sipondo, 2021).
Several CBC models exist. The ABCDE model structures the identification and challenge of unhelpful beliefs and automatic thoughts. The SPACE model fosters awareness of dysfunctional patterns to improve functioning. Palmer's PRACTICE model integrates cognitive and solution-focused techniques across seven steps: problem identification, goal setting, generating alternatives, evaluating and selecting solutions, implementation and outcome review (Palmer, 2007). PRACTICE has been applied successfully in coaching, psychotherapy, and stress management, enhancing resilience and well-being (Minzlaff et al., 2025).
Current AI coaching chatbots draw on a range of theoretical approaches, yet there remains a significant knowledge gap regarding which coaching approaches are most appropriate for addressing specific developmental outcomes. While the broader AI literature extensively examines capabilities and limitations, research on AI in coaching is still emerging and has largely focused on efficacy and outcomes rather than coaching processes (Bachkirova & Kemp, 2025). Although AI coaches have been trained using diverse methodologies, including goal theory, motivational interviewing, solution-focused coaching and cognitive behavioural approaches, there is little empirical guidance on which approaches are best suited to specific coaching problems (Passmore et al., 2026b). While CBC has informed comparative work on AI coaching design, particularly in conceptualising AI's application in coaching and clarifying its capabilities and limitations across the coaching process (Graßmann & Schermuly, 2021), no studies have yet evaluated CBC-based AI chatbots with expert oversight in resilience coaching contexts. Drawing on evidence from human coaching research identifying CBC as the most effective approach for resilience development (Sipondo & Terblanche, 2024), this study addresses this gap by adopting CBC as the theoretical foundation for a resilience-focused AI coaching chatbot.
Methodology
This pilot study forms a phase in a multi-stage research project investigating the design of an AI coaching intervention to enhance resilience. The purpose of the current phase was to gather insights from CBC experts regarding the relevance and potential efficacy of a CBC-based AI chatbot and to assess whether it incorporated the core elements of CBC necessary for effective resilience-focused coaching. Findings from this phase informed iterative refinements to the AI coach, to be subsequently deployed in later stages of the project.
Pilot studies commonly adopt flexible designs to explore feasibility in depth (Orsmond & Cohn, 2015). This pilot study employed a qualitative approach to capture the perspectives of CBC experts regarding the design, functionality and theoretical alignment of an AI coaching intervention (Aschbrenner et al., 2022). By examining the AI resilience coaching chatbot on a smaller scale, the study evaluated the viability of AI coaching methods and procedures, identified areas for refinement, and assessed whether the intervention is suitable for broader application and more extensive testing (Teresi et al., 2022).
Research Settings
Participants were granted unrestricted access to the chatbot for a period of 1 week, during which they were able to interact with it and provide feedback. The chatbot was deployed through WhatsApp and was accessible at all times, permitting participants to engage with it as frequently as desired, with a minimum requirement of one interaction.
The Chatbot
The participants interacted with an AI chatbot coach called Ziqamo. Ziqamo is a CBC coaching chatbot designed to help coachees to develop resilience by helping them identify internal resources, reframe irrational beliefs and adopt flexible thinking and behaviour (Sipondo & Terblanche, 2024). The underlying CBC coaching model used by Ziqamo is the PRACTICE model (Palmer, 2007). Ziqamo is a hybrid scripted and generative AI design, that uses ChatGPT 4o, and is based on the DAIC framework (Terblanche, 2020).
Figure 1 demonstrated an example of the chatbot interface of coach Ziqamo, which was accessed by participants accessed via WhatsApp at no cost.
Example of coach ziqamo interface.
Sampling
To obtain in-depth insights, purposive sampling was used to select coaching and counselling professionals with expertise in CBC or therapy (Etikan et al., 2016). In line with pilot study recommendations suggesting an optimal sample of six to eight participants for this type of study (Crabtree & Swainston, 2023), the study recruited seven experts with demonstrated proficiency in CBC (Etikan et al., 2016). Table 1 presents the demographic characteristics of the participants.
Demographic Characteristics of the Participants.
Data Collection
Data were collected using an unstructured, open-ended email questionnaire, suitable for exploratory research in emerging fields such as AI coaching (McGuirk & O’Neill, 2016; Teresi et al., 2022). Multiple attempts were made to schedule synchronous interviews; however, many potential participants did not respond, due to professional demands and limited availability. Consequently, an asynchronous email approach was adopted to enable participation without the constraints of scheduled interviews. Although this method limits opportunities for real-time probing (Harris et al., 2024), this method is justified due to the study's exploratory focus, the novelty of AI-based coaching and the need for timely access to expert perspectives in a rapidly evolving field. This method is supported by literature indicating that response quality and participant experience can be comparable to in-person, telephone or video interviews, particularly for busy professionals who benefit from time to reflect and provide thoughtful, data-rich responses (Harris et al., 2024). The format also ensured consistency by presenting identical questions in the same order to all participants (Tombs & Strange, 2024) and supported candid, in-depth insights from experts selected for their experience in CBC (Amirova et al., 2024; Tombs & Strange, 2024). Table 2 provides examples of the questions included in the questionnaire.
Example of Questions in the Study Questionnaire.
Data Analysis
The unstructured questionnaire transcripts were analysed using thematic analysis (TA) following Braun and Clarke's framework (2012). Consistent with this recursive approach, analysis involved ongoing movement between the full dataset, coded extracts, and emerging interpretations. Step 1 entailed deep familiarisation with the data through repeated reading of verbatim transcripts and the recording of initial analytic notes in relation to the research questions concerning the feasibility of a CBC AI coach for improving resilience. In Step 2, initial inductive codes were generated and applied to meaningful text segments ranging from sentences to paragraphs, for example coding statements such as ‘It asked questions that prompted self-reflection’ as ‘encouraging reflection’ and ‘It didn’t feel like it was judging me’ as ‘non-judgemental’. Step 3 involved collating related codes into higher-order categories (e.g., combining ‘AI coach values’ and ‘AI coach navigation’ into ‘AI coach characteristics’), followed by Step 4, which entailed systematic review and refinement of these categories across the dataset to ensure internal coherence and clear distinction between categories. In Step 5, categories were integrated into overarching themes that captured broader patterns of meaning, such as combining ‘chatbot advantages’ and ‘chatbot coaching support’ under ‘chatbot benefits’. This process resulted in 87 codes grouped into nine categories, from which four key themes emerged. The analysis was documented throughout the process and culminated in a coherent, evidence-based thematic account addressing the study's research questions.
Ethical Approval
This study received ethical approval from the Research Ethics Committee of the researchers’ university on 7 November 2024 (Project ID: SBS-2024-32095). Written informed consent was obtained from all participants prior to data collection. The AI coach was developed exclusively for research purposes, and the authors declare that they derive no financial or commercial benefit from its use.
Credibility and Trustworthiness
Credibility and trustworthiness were ensured through multiple strategies. The unstructured, open-ended questionnaire was informed by a comprehensive literature review and consultation with expert stakeholders, ensuring clarity, relevance and alignment with the study aims (Tombs & Strange, 2024). Expert participants with CBC experience provided rich, reflective responses, enhancing authenticity. Data were analysed iteratively using Braun and Clarke's (2012) thematic analysis framework, with systematic coding, categorisation and theme development documented for transparency. Researchers maintained regular discussions throughout data collection and analysis to ensure consistency and reflexivity. Combined with expert validation, iterative analysis and reflective review, these measures strengthened methodological rigour and ensured the findings accurately captured participants’ nuanced perspectives on the AI coaching intervention.
Findings
The four themes that emerged from the study were: Alignment with CBC, Reflective space, Chatbot benefits and Comparison with human coaching.
Alignment With CBC
Participants reported that the chatbot aligned well with CBC, particularly in addressing simple tasks, identifying unhelpful thinking, breaking problems into manageable actions and focusing on specific behavioural outcomes. They appreciated its probing questions regarding emotions and somatic experiences associated with negative thought patterns. The chatbot was also seen to support structured reflection, facilitate recognition of unhelpful patterns and prompt cognitive reframing, guiding users towards small, actionable steps.
P1: The chatbot reframed things positively, which is consistent with CBC/CBT's emphasis on building adaptive thinking patterns. P5: At one point the chatbot asked me to articulate what I was thinking in a certain situation and then asked me what evidence I had that contradicted that belief. Furthermore, behaviours were suggested that would have catalysed healthier cognitions.
However, the participants felt that most of the alignment with CBC was at a superficial level since the chatbot was not able to explore their schema, way of thinking, what informs their views and perspective. For example, P1 would have preferred more nuanced responses based on the content they shared.
P2: CBT is based on the interconnectedness of an individuals’ thoughts, emotions and behaviours, the understanding of its interplays and how it influences or direct the current status quo. The chatbot does this very superficially.
Some participants felt that the chatbot fell short from really getting down to unearthing belief systems and or mental models that lurks in the unconscious. While there is an option to explore the problem further, P2 did not feel as though the chatbot was exploring their thinking but more guided the conversation towards solutions. Furthermore, P4 did not see how the chatbot could explore one's thought structure, history and why they struggle with what they do. The participants felt that the chatbot struggled to ‘move them forward’/challenge them when they presented it with fixed unhelpful thinking.
Reflective Space
Another theme that emerged was a reflective space that the participants emphasised that the chatbot offered. According to P5, the suggestions offered triggered valuable thinking.
P5: The suggestions offered at various points (e.g., lists of potential action points/ homework) triggered valuable thinking for me …. I appreciated the probing questions about emotions and somatic experiences accompanying negative thinking patterns.
P7 highlighted that the chatbot had a good use of questions, follow up questions, opportunity to reflect, directing conversation back to the coachee. Participants felt the chatbot offered strength-based reflections, which felt personalised and constructive. The chatbot asked questions that prompted self-reflection without pushing participants towards a particular answer and instead brought structure and reflection in a digestible, supportive format.
P5 highlighted that the chatbot asked plenty of questions and was distinctly non-directive unless suggestions were specifically ‘requested’. P6 felt that the prompting by the was valuable, offering structured nudges and reflective questions. The chatbot facilitated reflective and non-directive conversations well by facilitating self-awareness without being overly directive.
Chatbot Benefits
The study findings highlighted various chatbot benefits. First, the chatbot was seen as intuitive.
P1: I was positively surprised by how insightful the chatbot was. It developed thoughtful questions and gave me useful prompts to explore what I was trying to achieve. It didn’t come across as robotic or superficial – in fact, it showed a level of psychological alignment that's often missing from digital tools.
Second, the participants were of the view that the chatbot was easy to use. Participants mentioned that the navigation was straightforward, and the simplicity of the interface was a plus. Participants mentioned that the interaction with the chatbot felt easy and straightforward. Immediacy and accessibility of the chatbot was seen as valuable. P1 appreciated the chatbot's structure in helping them break down goals into tangible steps.
Third, participants felt that the chatbot provided psychological safety in that it didn’t feel like the chatbot was judging or trying to fix them – instead, it created space to pause and consider things. Participants appreciated the warmth, empathy and enthusiasm of the chatbot communication. The chatbot was found to be ‘personable’ and ‘easy’ to communicate with. The use of participant's name and reflection of what the participants had written was seen as valuable and personable.
Comparison With Human Coaching
Participants were of the view that the chatbot could not compare to human coaching. Participants felt that it was evident that the two approaches have distinctly different outcomes with both having a purpose, but the purposes differ. Compared to traditional coaching, some participants felt that the chatbot cannot offer full relational nuance. P7 mentioned that after a few exchanges with the chatbot, they were reminded of the chatbots’ inability to continue building on that rapport and in some respects, the ‘relationship’ stagnates.
It emerged that the chatbot may fall short for those needing higher levels of emotional containment or more complex processing.
P2: If a user is interested to really grow and transform as e.g., a leader, then no. The chatbot lacks the ability to have in depth quality conversations that uncovers what lies beneath what is conscious. It fails in having transformational conversations that facilitates real shift in an individual's thoughts, behaviours and emotions.
Some participants reported deeper dives experienced as frustrating as they lacked depth or ‘real meaningful conversation’. P2 felt that the AI technology is not yet at the level of real human-to-human transformational conversations. The chatbot lacks the sophistication of really locking onto human emotion, and ‘reading-between-the-lines’. P3 felt that the chatbot didn’t reflect during the conversation or do a check in during the conversation to get a sense of how the person is feeling.
Participants suggested that to maximise the chatbots’ effectiveness, it should be positioned as a complement to human coaching. It's not a replacement for human coaching, but it could be a great supplement or entry point. The chatbot was seen as a useful co-pilot, but likely not a substitute for deeper relational work.
The chatbot was seen as beneficial as an extra support tool to individuals who already had a coaching session with a coach. Participants emphasised that the chatbot would be most beneficial for between-session coaching support and targeted issues or micro-challenges. The chatbot was seen as having ‘something in your pocket’ that offers strong convenience and potential cost saving benefits compared to traditional coaching models. P2 felt that the chatbot was useful if the user wants a quick ‘mentor-in-your-pocket’ that provides immediate suggestions to immediate concerns or problems.
Discussion
The discussion is structured around three key areas corresponding to the study's research questions. First, it examines the extent to which the AI coaching chatbot adheres to the principles of CBC. Next, it explores how the chatbot aligns with established practices of effective resilience coaching. Finally, the discussion considers how the AI coaching chatbot compares with human coaching in terms of efficacy and alignment with resilience coaching standards.
The Extent to Which the Chatbot Aligns With Established Principles of CBC
The study found that the AI coach was able to perform several core CBC functions, including helping coachees identify unhelpful thinking patterns. This aligns with the central premise of CBC that challenges can be addressed by examining and reassessing maladaptive beliefs and behaviours (Karas & Spada, 2009). As resilience depends on cognitive and behavioural flexibility, CBC seeks to increase awareness of irrational beliefs that impede goal attainment, with cognitive change driving subsequent emotional and behavioural adaptation (Neenan, 2018).
In addition, the AI coach in this study supported the reframing of unhelpful thinking and facilitated perspective shifts, consistent with CBC's focus on challenging cognitive distortions and maladaptive core beliefs (Dinos & Palmer, 2015). Resilience is shaped less by adversity itself than by individuals’ appraisals and responses, and coaching supports this process by enabling reframing and adaptive coping (Fontes & Dello Russo, 2021).
The AI chatbot in this study also demonstrated the use of probing and reflective questions, aligning with coaching practices that employ targeted questioning to enhance self-awareness, accountability and behavioural change (Bruning & Boak, 2026; Cardwell & Wright, 2023). In CBC, Socratic questioning is particularly important in short-term interventions, as it supports problem-solving and the expansion of psychological resources (Lautsten et al., 2018; Sipondo & Terblanche, 2024).
The study further showed that the AI coach could break complex problems into manageable actions and nudge users towards behavioural follow-through, consistent with CBC's solution-focused, goal-oriented approach to enabling goal attainment (Carvalho et al., 2018). In this role, the AI coach functioned as both a navigational aid and a facilitator of goal clarification and action experimentation (Barger, 2025; Joe, 2025).
However, the application of CBC by the AI coach was largely superficial, with limited capacity to explore underlying schemas or challenge deeper assumptions. This reflects broader limitations of current AI systems in capturing contextual nuance and individual experience (Blyler & Seligman, 2024). Consistent with Arakawa and Yakura (2024), the findings suggest that AI coaches can support single-loop learning through repeated goal-focused clarification but have limited capacity for double-loop learning that requires deeper challenge. For example, Passmore et al. (2026a) found that AI coaches struggled to generate original metaphors to facilitate deeper insight. Overall, these findings indicate that AI coaching chatbots are currently better suited to performance-focused or transactional coaching, rather than complex developmental or psychologically advanced coaching tasks (Bruning & Boak, 2026; Rasool, 2025).
The Extent to Which the Chatbot Adheres to Recognised Principles of Effective Resilience Coaching
Identification of Issues and Goal Setting
This study demonstrated that the AI coach supported coachees in identifying issues and setting goals through effective questioning that kept the conversation coachee-centred and focused on concrete behavioural outcomes. Coaching literature characterises coaching as a structured process that promotes self-exploration and goal attainment by enabling coachees to challenge limiting beliefs, reframe difficulties and independently identify actions (Joe, 2025). Consistent with the PRACTICE model and CBC, the AI coach engaged coachees in guided discovery rather than providing solutions, supporting them to generate their own insights and strategies (Neenan & Palmer, 2001). Through this process, the AI coach in this study contributed to resilience development by helping coachees identify and mobilise internal and external resources and build cognitive skills to manage challenges more effectively (Jouali et al., 2024). Goal-directed effort within coaching has been shown to strengthen resilience and self-efficacy by enabling individuals to overcome setbacks and progressively manage greater levels of challenge (Fontes & Dello Russo, 2021).
Non-Directive Questioning
Participants in this study valued the AI coach's use of non-directive questioning, which provided structure while breaking complex problems into manageable actions and gently nudging users toward practical next steps. Non-directive coaching supports coachees in generating insight and identifying actions aligned with their personal goals (Barger, 2025). By adopting a solution-focused and forward-looking stance, the AI coach created space for coachees to explore alternative ways of addressing current challenges, thereby supporting resilience development (Cardwell & Wright, 2023).
Reflective Space
The findings indicated that the AI coach effectively supported reflective conversations. Self-reflection is central to resilience, as increased self-awareness enables improved emotional regulation and the restoration of psychological balance (Joe, 2025). Coaching provides a reflective space in which coachees can explore personal and professional issues, reducing stress and anxiety while supporting intentional cognitive and behavioural change (Lawton Smith, 2017). By facilitating reflection on thoughts, emotions and behaviours, the AI coach may contribute to resilience by supporting self-regulation, which functions as a key protective mechanism (Joe, 2025; Sipondo & Terblanche, 2024).
Psychological Safety
The study found that the AI coach fostered psychological safety by adopting a non-judgemental stance and creating a safe space for coachees to pause and reflect. Prior research suggests that individuals may disclose more openly to AI coaches, which are often perceived as psychologically safe due to reduced fear of judgement and impression management (Diller et al., 2024; Ellis-Brush, 2021). These findings align with AI coaching literature indicating that unbiased, objective environments can enhance coaching effectiveness and goal attainment (Graßmann & Schermuly, 2021; Khandelwal & Upadhyay, 2021).
The Extent That the Coaching Experience Provided by the AI Chatbot Compares to Human-Delivered Coaching
The coaching relationship is central to resilience development, as it enables coachees to broaden perspectives and adopt more flexible interpretations of adversity (Sipondo & Terblanche, 2024). However, this study showed that the AI coach lacked relational depth and continuity, resulting in stagnation of rapport over time. This is significant given that a sustained coaching relationship, grounded in trust, mutual respect and collaboration, is widely recognised as critical to effective coaching and resilience outcomes (de Haan et al., 2016; Graßmann et al., 2019; Joe, 2025). CBC, in particular, relies on a cooperative partnership to guide structured problem-solving, underscoring the continuing importance of human coaches in maintaining engagement (Arakawa & Yakura, 2024; Neenan & Palmer, 2001).
The findings further indicated that the AI coach may be insufficient for coachees requiring deeper emotional containment or complex psychological processing. The chatbot struggled to engage in depth-oriented dialogue, ‘read between the lines’, or attune to subtle emotional cues. This aligns with prior research showing that AI systems have limited capacity to interpret nuanced human factors, including emotions, personal context and learning preferences, which may result in missed needs or insensitive responses (Panda & Mishra, 2024). Although AI can generate evidence-informed insights, it lacks subjective experience and contextual understanding, preventing it from participating in joint inquiry, which is an essential element of effective coaching (Bachkirova & Kemp, 2025). Joint inquiry is particularly important for resilience, as coaches facilitate this process through challenge, feedback and reflective dialogue that support self-discovery and ongoing evaluation (Sipondo & Terblanche, 2024). Through facilitated reflection, human coaches help coachees explore multiple perspectives, build self-insight and develop action plans, thereby enhancing self-regulation, coping and problem-solving capacities (Grant & Kinman, 2014). Coaching for resilience often involves asking challenging questions that provoke challenge and deep reflection, an area where AI coaches currently appear limited (Arakawa & Yakura, 2024).
Rather than competing with human coaches, the findings suggest that AI coaching chatbots are best positioned as complementary tools, particularly for between-session support and targeted micro-challenges. This aligns with existing research emphasising that AI coaching is most effective when supporting, rather than replacing, human coaches (Bruning & Boak, 2026; Khandelwal & Upadhyay, 2021). Large-scale survey evidence similarly indicates that AI chatbots should be designed to complement human coaching roles (Diller et al., 2024). When used as a ‘pocket coach’, AI can provide accessible, on-demand support via smart devices, enhancing inclusivity and continuity between sessions (Arakawa & Yakura, 2024; Diller, 2024). Between-session engagement is particularly important for resilience development, as it enables coachees to practise skills and implement action steps, with AI offering structured prompts and reflection support (Sipondo & Terblanche, 2024). However, effective integration requires clarity regarding when AI or human coaches are guiding the process to preserve the distinct value of human connection (Diller, 2024).
Finally, the study found that the AI coach did not routinely check in on the coachee's emotional state during conversations. While AI can simulate empathic responses, it does not experience emotions and may therefore struggle to provide genuine emotional support (Passmore et al., 2026a). This limitation is well documented, as interactions requiring human self-awareness, empathy, and intuition remain contested domains for AI coaching (Bruning & Boak, 2026). Given that emotional support and attentive listening are critical to resilience, supporting emotional regulation, confidence, and adaptive change, developers of AI resilience coaches should prioritise empathic design features, while recognising that human coaches remain essential for forming robust coaching relationships (de Haan et al., 2016; Jouali et al., 2024).
Implications for Theory and Practice
This study advances theory by demonstrating the capabilities and limitations of AI coaching in the context of CBC and resilience development. Findings indicate that AI chatbots could support goal clarification, reflective questioning and problem-solving, consistent with CBC principles, but have limited capacity to engage in deeper schema exploration or transformational cognitive change. Mapping the AI coach to the PRACTICE model highlights which stages are effectively supported (e.g., problem identification, goal setting, reflection, action planning) and which require human input, providing a foundation for developing hybrid AI–human coaching frameworks that integrate relational and cognitive dimensions.
For practitioners, the findings suggest that AI coaching chatbots are best positioned as complementary tools rather than substitutes for human coaches. AI coaching chatbots can support resilience development by facilitating between-session reflection, delivering micro-challenges and guiding users through structured, goal-directed practice. Effective implementation requires clearly defined coaching tasks, attention to usability and integration within broader coaching programmes, recognising the continued importance of human relational, emotional and transformational capacities. These insights offer practical guidance for designing AI-supported resilience interventions and inform strategies for hybrid coaching models that optimise both AI and human strengths.
Limitations and Future Research
The short duration of this pilot study may limit the generalisability of the findings, as longer-term engagement with the AI coach could yield different results. Future research should include longitudinal designs to determine whether the observed outcomes are sustained over time. Additionally, the small sample size restricts the broader applicability of the findings; thus, studies involving larger and more diverse participant groups are necessary to enhance generalisability. The simplicity of the chatbot used in this study may have constrained its capabilities. Further empirical research is needed, employing more advanced AI coaching systems with enhanced features, to better assess the potential and limitations of AI in coaching contexts.
Conclusion
This study provided insights into expert perceptions of using a CBC AI chatbot, highlighting both its potential and current limitations. The chatbot demonstrated a strong alignment with CBC techniques in facilitating structured reflection, identifying unhelpful thoughts, and supporting goal-focused behaviour. It created a psychologically safe, accessible and reflective space for coachees, offering non-directive prompts and fostering guided discovery through Socratic questioning. However, the findings also revealed significant limitations in the chatbot's ability to support deeper cognitive-emotional exploration, such as schema-level work and relational nuance, which are essential for sustained behavioural change and complex coaching contexts. While the chatbot fell short in replicating the depth and emotional intelligence of human coaching, it showed promise as a complementary tool, particularly for between-session support or addressing targeted micro-challenges. By engaging CBC experts, this study contributes to a more nuanced understanding of AI coaching and its theoretical alignment, offering implications for future chatbot design and calling for more longitudinal and large-scale studies using advanced AI tools.
Footnotes
Funding
The authors received no financial support for the research, authorship and/or publication of this article.
Declaration of Conflicting Interests
The authors declared no potential conflicts of interest with respect to the research, authorship and/or publication of this article.
