Adoption of Artificial Intelligence in Organizational Coaching Processes

Abstract

Artificial intelligence (AI) is transforming how organizations develop human potential, offering scalable and data-driven support for coaching and capability building. This study proposes and validates a conceptual framework for integrating AI into organizational coaching processes to enhance competence development and strategic alignment. AI-supported coaching in this research is treated as an emerging organizational technology whose potential organizational value depends less on model capability and more on governance design, decision rights, and auditable evaluation outputs. Following a mixed-methods, multi-phase design, the research combined a Systematic Literature Review (SLR) with the construction of a layered design architecture in which OSCAR serves as the primary coaching-process scaffold, complemented by KSA for competency specification, Situational Leadership for adaptive guidance, and KPIs for monitoring and governance. The framework structures AI-supported coaching across 10 interrelated phases, from contextual anchoring to review and measurement, while preserving iterative re-entry to earlier phases whenever review evidence, contextual change, or insufficient progress makes adjustment necessary. Prototyping demonstrated feasibility and coherence across models, while the focus group provided qualitative expert feedback on the framework’s clarity, governance needs, and perceived usefulness for competence development. At this stage, however, the KPI structures generated by the framework and the descriptive comparison across AI tools should be interpreted as prototype-level outputs rather than as empirically validated performance measures or evidence of added value over baseline approaches. Because the evaluation relied on two fictional prototyping scenarios and a small expert-oriented focus group (n = 6), the findings should be interpreted as evidence of prototype demonstration and qualitative refinement rather than of real-world effectiveness or organizational impact. The study also does not include a control group or comparison with traditional human coaching, so the added value of the AI-supported framework over alternative coaching arrangements remains a question for future empirical testing. Findings suggest that AI can usefully support organizational coaching by personalizing dialogue, structuring reflection, and generating auditable development artefacts, provided ethical safeguards and human oversight remain integral. The research contributes a preliminarily validated, ethics-informed, and governance-aware framework for AI adoption in organizational coaching and offers practical insights for embedding AI-enabled development in learning organizations.

1. Introduction

Artificial intelligence (AI) is accelerating organizational innovation and reshaping knowledge-intensive work, including how organizations develop people and capabilities. In parallel, organizational coaching has matured into a strategic, ethically grounded intervention that balances individual growth with business needs and multi-stakeholder realities [1]. Where classic coaching centered on human-to-human dialogue, AI-enabled coaching—especially with recent advances in large language models (LLMs) and generative AI—is emerging as a complementary modality that can support clients across different stages of the coaching process [2,3]. The convergence of these trends raises a timely question for researchers and practitioners alike: How can organizations harness AI to amplify, rather than dilute, the developmental value of coaching?

Organizational coaching is distinguished from other coaching genres by its explicit alignment with strategic objectives and by the contextual complexity of organizational relationships. Coaches operate with “ethical clarity and contextual awareness,” enabling employees to exercise autonomy and make responsible decisions while serving organizational outcomes [1]. The profession has expanded rapidly, demonstrating effects on well-being, behavioral change, goal attainment, work engagement, communication, and self-regulation [4]. However, cost and access constraints limit the reach of traditional one-to-one interventions, particularly for large or distributed workforces. These constraints have catalyzed experimentation with AI coaching, which Graßmann and Schermuly (2021) [3] define as a machine-assisted, systematic process to help clients set professional goals and construct solutions efficiently. Building on this view, Terblanche (2024) argues that AI can emulate elements of coaching expertise and act as a facilitator before, during, and after human coaching sessions [5].

To date, evidence suggests several potential benefits. First, AI coaches and assistants can extend accessibility by providing 24/7 support, micro-interactions between human sessions, and scalable “always-on” follow-ups, which are especially valuable where budgets or geography limit human coaching capacity [6,7,8]. Second, AI can offer personalization and consistency through structured questioning, progress tracking, and prompts grounded in recognized coaching models (e.g., goal setting, reflective inquiry) [3]. Third, data captured across interactions can enable analytics-informed development (e.g., trends in goal progress, obstacles, or sentiment) while supporting the coach’s situational awareness [5]. Finally, when thoughtfully integrated, AI can function as coach support—proposing questions, prompting reflection, and sustaining momentum—rather than replacing the human relationship [5].

At the same time, early literature—and practitioner commentary—highlight non-trivial challenges. The most frequently cited risks concern a potential loss of empathy and the “human touch,” the difficulty of conveying nuanced emotions in complex cases, and the danger of over-mechanizing feedback [8]. Ethical and legal issues arise around privacy, data protection, and transparency—especially in interactions involving sensitive personal data and where standards for AI coaching are still evolving [9]. There is also a call for common design principles and theoretical grounding (e.g., in established coaching models) so that AI systems do not simply simulate coaching rhetoric but support learning and behavior change [3,10]. These tensions point not to a single broad deficiency in the literature, but to a specific design-and-implementation gap in AI-supported organizational coaching. Existing studies have advanced the field in three important yet largely separate ways. For example, they clarify the capabilities and limits of AI coaching [3,5], examine adoption factors and bounded AI-coaching applications [4,8,9], and propose design or ethical principles for trustworthy AI-enabled coaching [2,11,12]. However, these strands remain insufficiently integrated for organizational use.

More specifically, the literature still lacks a process-oriented framework that translates these insights into a coherent organizational coaching architecture. Prior work does not yet sufficiently explain how AI should be embedded across coaching phases, how competency development should be specified through explicit KSA targets, how progress should be monitored through KPI systems, and how governance requirements such as transparency, human oversight, and auditability should be operationalized within the coaching workflow [2,11,12,13,14,15,16]. Accordingly, the gap addressed in this paper is not primarily a lack of general empirical discussion about AI coaching, but the absence of an integrated, design-oriented, and implementation-ready framework for organizational coaching contexts.

This paper addresses that gap by proposing and evaluating a conceptual framework for integrating AI into organizational coaching to support competency development aligned with business strategy. The study, therefore, takes a design-oriented perspective: rather than testing whether AI coaching is universally effective across settings, it focuses on how an organizationally governed AI-supported coaching process can be structured, operationalized, and preliminarily validated. The guiding research question is: “How can an AI-supported organizational coaching process be designed and operationalized to support competency development aligned with business strategy while preserving governance and human oversight?” This question is intentionally design-oriented. Rather than testing the comparative effectiveness of AI-supported coaching against traditional coaching approaches, the study focuses on the development, operationalization, and preliminary validation of a conceptual framework. The framework is not conceived as a simultaneous fusion of multiple theories, but as a layered design architecture with differentiated functions. OSCAR provides the primary process scaffold for structuring the coaching dialogue from outcome definition to review [13]. KSA is introduced after action planning to translate coaching goals into explicit competency targets [14,15]. Situational Leadership is used as an adaptive layer to modulate the level of guidance according to the coachee’s readiness profile rather than to redefine the process structure itself [17]. KPIs are added as a monitoring and governance layer to make progress auditable and strategy-linked to organizational objectives [16]. At the current stage, these governance and value mechanisms should be interpreted as proposed features of the framework rather than as empirically confirmed organizational outcomes. The study evaluates whether such mechanisms can be coherently designed, instantiated, and judged plausible by expert participants; it does not directly demonstrate that they improve organizational performance in practice. Collectively, these design choices seek to balance process structure, competency specificity, adaptive support, and organizational accountability without treating the four models as interchangeable or equally central.

Because coaching is inherently developmental and unfolds over time, robust evaluation of coaching impact would ideally require longitudinal observation of behavioral change, skill transfer, retention, and progression. The present study does not provide that level of outcome evidence. Instead, it focuses on the design, operationalization, and preliminary validation of an AI-supported coaching framework, assessing whether the framework can be coherently instantiated, produce auditable development artefacts, and be judged useful and plausible by practitioners. Although the framework is presented through ordered phases for design clarity, it is intended to operate as an iterative coaching cycle rather than as a strictly linear pipeline: review points can trigger a return to earlier phases when goals require revision, situational understanding changes, or progress evidence indicates the need for adaptation.

This research makes three contributions. First, it clarifies the state of the art by distinguishing conceptual, implementation, and governance strands in the AI-coaching literature, thereby sharpening the specific research gap addressed by this study. Second, it proposes an integrated framework that combines OSCAR as a process scaffold with KSA-based competency targeting, situational tailoring, and KPI-based monitoring, addressing the lack of design-oriented and implementation-ready guidance in prior literature [13,14,15,16]. Third, it provides an initial operational validation of the framework through multi-model prototyping and practitioner focus group evaluation, yielding concrete implications for the ethical and practical adoption of AI-supported coaching in learning organizations.

The qualitative evaluation reported in this study should be interpreted accordingly. Because the focus group involved a small, expert-oriented sample and assessed perceptions of framework clarity, flow, governance value, and usefulness, it provides evidence of design quality and conceptual plausibility rather than direct evidence of usability or impact in real organizational settings. Recent review-based and critical literature also suggests that AI-supported coaching should not be theorized only through legacy coaching or managerial models in isolation. Bachkirova and Kemp (2024) [1] explicitly question whether AI coaching democratizes development or risks becoming an ersatz substitute, while Passmore, Olafsson, and Tee (2025) [18] show that the field remains emergent, methodologically uneven, and in need of stronger product and research standards. In parallel, recent reviews of responsible AI governance and human–AI collaboration argue that contemporary organizational AI systems must be interpreted through explainability, accountability, human oversight, and co-agency rather than through tool capability alone [19,20,21]. This broader review literature is therefore important for the present study because it frames the proposed framework not as a reaffirmation of older models in unchanged form, but as a contemporary, governance-aware reinterpretation of selected legacy scaffolds for AI-supported coaching.

The remainder of the paper proceeds as follows. Section 2 details the research methodology and design. Section 3 reviews related work on organizational coaching and AI-supported coaching, including conceptual models, empirical findings, and ethical frameworks. Section 4 presents the proposed framework and explicates the AI roles in each phase. Section 5 reports the prototyping and demonstration results. Section 6 reports the preliminary qualitative validation results from the focus group. Section 7 discusses the implications, limitations, and future research directions. Section 8 concludes the paper.

2. Research Methodology

This research follows a multi-phase, mixed-methods design [22] to (i) map the state of the art on AI-supported coaching, (ii) design and operationalize an AI-enabled coaching framework for organizational skill development, and (iii) evaluate its clarity, usability, and perceived value with practitioners. Importantly, this research design is not an outcome-effectiveness study. It does not compare AI-supported coaching with traditional coaching, nor does it measure behavioral change, performance improvement, or sustained competency development over time. Instead, it adopts a design-and-preliminary-validation logic focused on process coherence, operational feasibility, artefact completeness, and practitioner appraisal. The methodological logic of this study is pilot-level and design-oriented. The research does not test the framework through real organizational deployment, repeated coaching cycles, or outcome-based measurement of employee development. Instead, it examines whether the framework can be coherently instantiated, whether it produces complete and auditable coaching artefacts, and whether expert participants judge it to be clear, useful, and organizationally plausible.

For the same reason, the study does not empirically validate whether the KPI sets generated by the framework predict or improve real performance outcomes, nor does it compare framework-guided use against baseline conditions such as unguided AI prompting, human-only coaching, or business-as-usual development practice.

Also, the study does not include quantitative/statistical outcome analysis, controlled comparison with alternative coaching approaches, or inferential testing of impact. Its methodological purpose is narrower: to examine whether the framework can be coherently designed, instantiated, and qualitatively appraised under controlled prototyping conditions.

The phases comprise a Systematic Literature Review (SLR), theory-informed framework construction, operational prototyping with large language models (LLMs) and workflow automation, and qualitative validation via an online focus group, which informs an iterative refinement of a second framework version. An overview of this staged design appears in Figure 1.

Importantly, this research design does not aim to measure the comparative effectiveness of AI-supported coaching against human-only or business-as-usual coaching. Accordingly, the study design does not include a control group through which effectiveness, efficiency, or user satisfaction could be compared between the proposed AI-supported framework and traditional human coaching. Instead, it is intended to evaluate whether the proposed framework can be coherently instantiated, whether it produces complete and auditable coaching artefacts, and whether practitioners judge it to be clear, useful, and organizationally plausible. Thus, the study adopts a design-and-preliminary-validation logic rather than an outcome-effectiveness design or direct test of long-term coaching impact.

In each phase, several activities are executed, as follows:

• Phase 1—Systematic Literature Review (SLR): Phase 1 consisted of a structured review of the literature intended to map the state of the art on AI-supported coaching and to inform the design requirements of the proposed framework. Databases/search engines included Scopus (primary), the EBSCOhost interface, Elicit (an AI-assisted search tool), and Google Scholar for complementary and grey literature. The review synthesized adoption factors, perceived benefits and risks, fielded systems, design principles, governance requirements, and the available empirical evidence on AI-supported coaching outcomes. Notably, the corpus was uneven: studies on usefulness, intention to use, and illustrative applications were more common than rigorous evaluations of goal attainment, behavioral change, working alliance, KPI attainment, or organizational performance. This imbalance informed the structure of the background review and reinforced the need to distinguish clearly between adoption evidence and outcome-based evidence in the literature. Although the review was conducted systematically in terms of source selection and thematic synthesis, it was not reported as a fully reproducible PRISMA-compliant systematic review. Therefore, the findings from Phase 1 should be interpreted as a structured knowledge base for framework design rather than as a formal systematic review with complete protocol disclosure.
• Phase 2—Theory-informed framework construction: Drawing on established coaching and management concepts, the initial framework version (V1) integrates: OSCAR (Outcome–Situation–Choices–Actions–Review), KSA (Knowledge, Skills, Abilities), Situational Leadership for competence/commitment profiling (D1–D4), SMART goal setting, and KPI design for progress tracking. The resulting ten-stage sequence personalizes coaching to align with organizational goals and employee context. It includes a Review command for automated progress recaps and a DONE command to generate a consolidated report of the session.
• Phase 3—Operational prototyping with LLMs and prompt engineering: To test feasibility, workflow logic, and artefact generation, the framework was instantiated as prompts and exercised in two case scenarios within a fictional firm (“CreativeTech): (1) a UX manager (Noah) balancing research quality with constraints; and (2) a frontend developer (Emily) transitioning to Angular. The tests used ChatGPT 4.5, Gemini 3, and DeepSeek-V3.2 (free tiers), following recognized prompt-engineering heuristics—clarity/precision, explicit output formats, logical structure, and iterative refinement [23,24,25]. The purpose of this phase was not to conduct a rigorous performance benchmark or effectiveness evaluation, but to assess whether different LLMs could consistently operationalize the same coaching flow and generate the expected framework artefacts with sufficient completeness and coherence. Still in this phase, the artefact included workflow automation and a transparent reasoning trace. A proof-of-concept automation was implemented in n8n to simulate end-to-end AI-assisted coaching sessions. The flow used consisted of a Trigger node (chat message), a Tools Agent node seeded with the framework prompt, LLaMA3-70B-8192 as the LLM (via Groq credentials), and a Memory node to preserve conversational context. The purpose of this automation layer was to demonstrate workflow orchestration, state carry-over, and traceability at the prototype level. It was not designed or evaluated as a production-grade enterprise deployment. It therefore did not test scalability under concurrent use, robustness under failure conditions, advanced error handling, security hardening, or operational constraints such as authentication, observability, and infrastructure resilience.
• Phase 4—Qualitative evaluation via synchronous online focus group: To assess perceived clarity, flow, utility for competency development, and acceptance of AI as a coach, we conducted a single synchronous online focus group (cost-effective, inclusive, and rapid to organize) with six participants recruited from coaching, psychology, learning-organization practice, and human resources. The focus group was held via Microsoft Teams, and the session was audio-recorded and transcribed for analysis. A structured Focus Group Guide (nine prompts) probed participants’ understanding of the study aims, phase clarity, and transition fluency, contribution to competence development, perceptions of AI empathy, and recommendations. Data were coded in NVivo 15, with categories/sub-categories developed inductively; NVivo’s AI-assisted summarization supported concise within-code synthesis. Phase 4 was designed as a qualitative expert appraisal rather than a user-based field usability study. Its purpose was to elicit informed judgments about framework clarity, governance requirements, perceived usefulness, and refinement needs. Consequently, the focus group evidence should not be interpreted as broadly generalizable to employees or organizations, nor as a direct measure of real-world coaching impact. Although the study adopts a mixed-methods design in the sense of combining structured literature review, prototyping, workflow demonstration, and qualitative expert appraisal, its empirical validation component remains exploratory and small-scale. Accordingly, the mixed-methods label should not be interpreted as implying broad generalizability or large-sample validation.
• Refinement: Iteration—From Version 1 to Version 2: Feedback from the focus group was used to refine Version 1 into Version 2, primarily by (i) tightening governance/decision gates, (ii) specifying minimum required fields for core deliverables, and (iii) clarifying escalation and manager involvement.
• Phase 5—Conclusions: The results of this research are summarized, and new research directions are pointed out.

In Figure 2, the tools supporting the execution of each phase are depicted.

In terms of ethical, governance, and validity considerations, the work is informed by ICF AI coaching standards (transparency, disclosure, explainability, bias mitigation, data transparency, informed consent) [11] and EU Trustworthy AI principles emphasizing safeguards, auditability, and human override [12]. In the present study, these principles are operationalized at framework level through: (i) mandatory disclosure that the user is interacting with an AI coaching assistant; (ii) explicit human decision authority over goals, choices, and final acceptance of outputs; (iii) scope boundaries defining what the AI may and may not coach on; (iv) data minimization by limiting inputs to role, goal, and work-context data necessary for the coaching task; (v) auditable logging of artefacts and review decisions; and (vi) escalation rules that transfer sensitive, ambiguous, or out-of-scope cases to human oversight. These controls guided both the prototyping logic and the focus group protocol, even though full enterprise compliance testing remains outside the scope of the present study.

In terms of credibility and dependability, methodological rigor derives from triangulation across background research, prototyping/automation traces, and practitioner feedback. Because Phase 1 was conducted as a structured literature review rather than as a fully reported PRISMA systematic review, its role in this study is to provide a transparent conceptual and evidential basis for framework design. In contrast, the later phases provide operational and qualitative validation.

3. Background Research

In this Section, we present and discuss the state-of-the-art in AI-supported organizational coaching, as defined in Phase 1 of the methodology presented in Section 2.

3.1. What Organizational Coaching Delivers—And Where AI Adds Value

Organizational coaching is a structured, goal-oriented intervention that promotes behavior change, self-regulation, goal attainment, engagement, and communication, and operates explicitly in alignment with strategic, multi-stakeholder organizational objectives [1,3,4]. In the reviewed corpus, AI consistently contributes three practice-relevant benefits: (i) extended access and continuity between human sessions (24/7 micro-interactions), (ii) personalization and consistency through structured questioning and progress tracking, and (iii) analytics-supported reflection and monitoring for coaches and coachees [3,5,7].

At the same time, salient risks recur possible erosion of empathy/”human touch,” mechanistic feedback, and unresolved issues of privacy, bias, and transparency—hence the need for explicit governance [8,9,10].

The literature converges on the concept of complementarity: AI should be positioned to support (before/during/after) rather than replace human coaching, leveraging its strengths in scaffolding, reminders, and data capture while preserving the relational core of the working alliance [5]. At the same time, claims about “value” in this literature should be interpreted carefully, because evidence on adoption and perceived usefulness is more mature than evidence on validated coaching outcomes.

3.2. Adoption Evidence: Informative but Not Sufficient for Effectiveness Claims

A first empirical strand in the AI-coaching literature concerns adoption rather than outcomes. Across studies applying UTAUT and related acceptance lenses, performance expectancy (perceived usefulness) is often reported as the strongest predictor of intention to use AI coaching, with social influence, facilitating conditions, ease of access, and platform fit also shaping uptake [4,8]. These studies are valuable because they explain why users may engage with AI coaches and under which conditions acceptance is more likely.

However, adoption evidence should not be conflated with effectiveness evidence. Intention to use, perceived usefulness, and positive attitudes toward AI coaching do not by themselves demonstrate that AI-supported coaching improves goal attainment, behavioral change, competency development, or organizational performance. For that reason, adoption findings are relevant to implementation readiness, but they are only one part of the empirical foundation needed to understand the actual impact of AI-supported coaching [8].

Broader Organizational AI Literatures Relevant to Coaching Design and Adoption

Recent high-quality studies beyond the coaching field are increasingly relevant to AI-supported coaching because they address the organizational conditions under which AI systems are trusted, adopted, and made useful at work. In HRM and leadership development, recent work argues that generative AI requires a rethinking of people management frameworks, leadership capabilities, and leader learning processes, especially where AI is used to augment—not replace—human judgment [26]. Related research on human–AI collaboration has moved beyond a simple tool-user model toward co-agency and human-complementary design, highlighting trust, role clarity, and calibrated human oversight as central design requirements [21,27].

A second relevant stream concerns trustworthy and responsible AI in organizations. Recent governance research shows that explainability, accountability, stakeholder involvement, and auditable practices are critical for translating AI principles into operational reality [19,20].

This is particularly visible in HR and people analytics, where case-based evidence shows that organizations are beginning to use review boards, codes of conduct, and governance arrangements to address bias, privacy, and accountability concerns [28]. In parallel, adoption research suggests that digital capabilities, organizational change capacity, and staff skills strongly shape whether AI moves from experimentation to sustained organizational use [29,30]. These adjacent studies strengthen the present study by situating AI-supported coaching within the broader challenges of organizational AI adoption, governance, and capability building.

3.3. Outcome-Oriented Evidence: What Has Been Measured So Far

Compared with the adoption literature, empirical outcome-based research on AI-supported coaching remains limited but is beginning to emerge. One of the strongest comparative studies is Terblanche et al. (2022) [31], who examined goal-attainment efficacy across equivalent longitudinal randomized control trial designs over 10 months. Their study reported that both human coaching and an AI chatbot coach were significantly more effective than control groups in supporting client goal attainment, and that the AI coach reached goal-attainment levels comparable to those of human coaches in a narrow, goal-focused coaching context.

Additional evidence comes from process and relational outcome studies. Barger (2025) [32], in a mixed-methods randomized controlled trial, found that participants developed similarly moderate-to-high levels of working alliance with both a simulated AI coach and a human coach during a single coaching session, with no significant difference between treatments. Because working alliance is widely recognized as a central process variable in coaching, this finding is important. However, it should not be interpreted as proof of sustained developmental or organizational impact.

More context-specific studies also suggest bounded but relevant developmental effects. Terblanche and Tau (2024) [9] reported that first-time graduate employees who used a goal-attainment AI chatbot coach for four weeks perceived it as effective in helping them progress toward career goals through actionable steps, measurable outcomes, reflection, and self-awareness, while still missing the flexibility and human touch of a human coach. In a hybrid human-AI setting, Terblanche et al. (2024) [33] found that chatbot-assisted coaching was perceived as useful for goal tracking, accountability, convenience, and, when endorsed by the human coach, as psychologically safe within the broader coaching relationship.

Taken together, these studies suggest that AI-supported coaching can contribute to structured goal pursuit, reflection, accountability, and, in some settings, working alliance formation. However, the evidence base remains narrow in at least five respects: samples are often small; many studies are short-term; several rely on self-reported perceptions; most examine bounded coaching tasks rather than full organizational coaching programs; and few assess longitudinal behavioral change, KPI attainment, or business-performance outcomes. Thus, the empirical literature provides promising initial evidence, but not yet a sufficiently broad basis for strong general claims about the effectiveness of AI-supported coaching across organizational settings.

3.4. Process Scaffolds That Transfer Well to AI-Mediated Coaching

Among classical coaching models, OSCAR (Outcome–Situation–Choices–Actions–Review) stands out as a solution-focused scaffold that emphasizes clear outcomes, actionable steps, and regular review, mapping well to contemporary AI affordances (prompting, state tracking, automated follow-ups) [13]. Complementary lenses—KSA (competence targeting) and KPI frameworks (measurement/feedback)—anchor development in observable capabilities and close the loop with performance [15,16]. In this synthesis, OSCAR emerges as the primary coaching-process lens because it provides the conversation backbone through which goal clarification, situational diagnosis, option generation, action planning, and review can be structured over time. KSA and KPI logic are not treated as alternative coaching theories, but as supporting constructs that add competency specificity and measurable follow-up to an OSCAR-based process. Accordingly, the framework developed in this paper is grounded first in OSCAR as its dominant coaching lens, with KSA and KPI logic serving complementary operational functions.

Combining OSCAR + KSA + KPI yields a competence-centric, measurable pathway that AI can operationalize (phase-specific prompts; micro-check-ins; periodized reviews).

3.5. Design Frameworks: How to Build AI Coaches That People Trust and Use

The DAIC (Designing AI Coach) framework synthesizes human-coaching efficacy elements (trust, empathy, transparency), validated theoretical models, ethical conduct, and narrow task focus with chatbot design best practices (human-likeness cues, expectation management, transparency about learning limits, graceful failure, explicit disclosure) [2]. The central design lesson is focus: AI is most effective when scoped to a specific coaching outcome and implemented with clear disclosure and relationship-supportive behaviors; complete substitution of human coaches is neither realistic nor desirable at present [5].

Viewed in light of more recent critical and review-based literature, this design lesson should also be interpreted cautiously. Bachkirova and Kemp (2024) [1] warn that AI coaching may risk becoming an ersatz substitute if relational and ethical safeguards are not preserved, while [18] show that AI coaching research is still maturing and requires stronger evidence standards as well as more careful translation into product and practice design. These perspectives reinforce the need to treat older coaching and managerial models as design resources to be adapted under current AI conditions, rather than as sufficient stand-alone justification for organizational deployment.

Grounding assistants in established but critically interpreted process scaffolds (e.g., OSCAR), designing for trust and clarity, and setting explicit capability boundaries are necessary conditions for responsible AI-supported coaching.

3.6. Ethics, Governance, and Validation—Principles That Must Be Operationalized

Professional guidance from the International Coaching Federation (ICF) and the EU Ethics Guidelines for Trustworthy AI converge on a common baseline for AI-supported coaching: users must know that they are interacting with AI; the system must operate within declared limits; sensitive data must be minimized and protected; potentially harmful or biased outputs must be reviewable; and humans must retain override authority [11,12]. In organizational coaching, these principles only become meaningful when translated into concrete controls.

Five operational risk areas are especially relevant in AI-supported coaching. First, hallucination risk: AI systems may generate confident but unfounded developmental recommendations, inappropriate competency suggestions, or misleading KPI formulations. In a coaching setting, such errors are particularly problematic because they can shape self-perception, learning priorities, and action commitments. Second, privacy and confidentiality risk: coaching conversations may involve sensitive personal, interpersonal, or organizational information, requiring strong data minimization, access boundaries, and clear disclosure of how information is used. Third, bias risk: AI-generated competency assessments, development-level classifications, and recommended actions may reflect hidden assumptions or produce unfair or exclusionary guidance if not reviewed critically. Fourth, overdependence risk: if AI support becomes overly directive or normalized as a substitute for human judgment, coachees or organizations may defer excessively to automated outputs and weaken critical reflection. Fifth, escalation risk: some coaching situations exceed the appropriate scope of AI support and require transfer to a human coach, manager, or organizational approver.

Accordingly, ethics in AI-supported coaching should not be treated as a generic declaration of compliance, but as a set of enforceable workflow controls. These include disclosure, scope restriction, data minimization, auditable logging, human validation of outputs, periodic review for bias/harms, and explicit escalation pathways. In this sense, the present study treats ethics and governance as an operational design problem rather than as a purely normative one.

3.7. Findings

The review highlights three cross-cutting gaps.

• Implementation gap. Few studies detail how organizations should deploy AI-supported coaching in practice, including governance arrangements, role design, process integration, review mechanisms, and measurable outputs.
• Design-integration gap. Existing studies discuss AI-coaching concepts, bounded assistants, adoption factors, or design principles. Yet, they do not sufficiently integrate coaching process structure, competency modeling, performance monitoring, and organizational governance into a single framework suitable for organizational deployment.
• Ethics-to-practice gap. Ethical principles are widely stated, but actionable patterns such as disclosure rules, audit trails, bias testing, escalation logic, and human-override mechanisms are rarely operationalized within concrete coaching workflows and with enough detail to support accountable implementation.

Taken together, these gaps indicate that the central contribution of this study is design-oriented and implementation-focused: the paper addresses the absence of an integrated organizational framework for AI-supported coaching, rather than claiming to resolve the broader empirical question of AI-coaching effectiveness.

The present research responds by (i) grounding AI assistance in the OSCAR process while coupling it with KSA targets and KPI tracking; (ii) specifying AI roles before, during, and after sessions; and (iii) embedding operational governance controls such as disclosure, human validation, scope restriction, auditable logging, review cadence, and escalation pathways to translate ethical principles into workflow-level safeguards. This translation is especially important for concrete operational risks such as hallucinated developmental advice, confidentiality breaches, biased competency judgments, excessive reliance on automated recommendations, and unclear thresholds for escalation to human intervention.

4. Solution Proposal

In the previous section, the main topics and concepts were discussed. In this Section, the primary theoretical lens of the framework—OSCAR—is presented together with the supporting design constructs that extend its operational relevance in organizational settings, namely KSA, Situational Leadership, and KPI logic.

4.1. Proposal Background

4.1.1. Rationale for the Chosen Foundations

The framework proposed in this study was designed as a layered architecture rather than as a simple aggregation of concepts. The objective was not to combine multiple established models because they are individually well known, but to assign each selected component a distinct design function within an AI-supported organizational coaching process.

OSCAR was selected as the primary process scaffold because the framework required a structured conversational sequence that could be operationalized in prompts, automated workflows, and review cycles. Among classical coaching models, OSCAR was preferred because of its explicit progression from outcome definition to review, its solution-focused orientation, and its suitability for instrumented transitions between phases. In other words, OSCAR defines the order of the coaching process.

However, OSCAR alone does not specify what kind of capability is being developed. For that reason, KSA was selected as a secondary design layer. Its role is not to structure the conversation, but to translate coaching outputs into explicit competency targets—knowledge to acquire, skills to practice, and abilities to mobilize. KSA, therefore, gives semantic precision to the development objective.

The framework also required a mechanism for adapting the style of AI support to differences in learner readiness. Situational Leadership was selected for this purpose because it offers a parsimonious way to vary the degree of direction, persuasion, participation, or delegation according to competence and commitment. Its function in the framework is therefore adaptive rather than structural: it changes how guidance is delivered, not the process sequence itself.

Finally, KPI logic was incorporated because organizational coaching in this study is explicitly strategy-linked and governance-sensitive. KPIs were not included as an additional coaching theory, but as a monitoring and accountability layer that translates the coaching plan into auditable indicators, review cadence, and escalation triggers, through which value can be reviewed rather than assumed.

Accordingly, the four components are neither equivalent nor simultaneous in function. OSCAR structures the coaching journey; KSA specifies the development target; Situational Leadership modulates the intensity and style of support; and KPIs make progress reviewable and organizationally actionable. The added value of the framework lies precisely in this functional differentiation and sequencing.

The older constructs used in this study should therefore be interpreted as bounded design scaffolds rather than as uncontested current theories. OSCAR is retained for process structuring, SMART for pragmatic goal specification, KSA for competency articulation, and Situational Leadership for adaptive guidance, but the framework does not assume that these legacy models remain universally valid in unchanged form. For example, SMART is used in a pragmatic sense and should be interpreted cautiously, given the later critique that overly formulaic objectives can become reductionist when detached from context and learning processes [34]. More broadly, these legacy models are complemented in this article by recent review-based literature on AI coaching, responsible AI governance, human–AI collaboration, and organizational adoption. The contribution of the paper, therefore, lies not in revalidating decades-old theories but in reconfiguring them for a contemporary AI-supported organizational context.

It is important to explain why these components are important and what each adds.

A key design decision in this research was to avoid treating all candidate models as equally central. The framework, therefore, uses a selection logic based on four design requirements: process structure, competency specification, adaptive guidance, and measurable review.

First, the framework needed a coaching process scaffold suitable for sequential prompting, state persistence, and automated review. OSCAR was retained for this function because it offers clear phase transitions and an explicit review component. Although GROW and CLEAR were considered as alternatives, OSCAR was preferred because it more directly supports solution construction, action commitment, and review cadence in ways that are easier to operationalize in AI-mediated interactions.

Second, the framework needed a way to express development goals in competency terms rather than only as broad aspirations. KSA was selected because it provides a practical and transparent bridge between coaching dialogue and trainable capability targets. Without this layer, the framework would risk producing action plans without a sufficiently explicit competency model.

Third, the framework needed to adapt the style and depth of AI guidance to the coachee’s development level. Situational Leadership was selected because it offers a compact and operational vocabulary for varying guidance intensity according to competence and commitment. Without this layer, the framework would remain process-consistent but pedagogically less adaptive.

Fourth, the framework needed auditable monitoring and organizational follow-up. KPI logic was therefore selected to convert developmental intent into measurable progress signals, review cadence, and governance triggers. Without this layer, the framework would support coaching conversations but would be weaker as an organizational instrument.

For clarity, the necessity of each component can be expressed as a conceptual ablation logic: removing OSCAR would remove the process spine; removing KSA would weaken competency specificity; removing Situational Leadership would reduce adaptive tailoring; and removing KPIs would weaken measurement, review, and governance. The framework’s contribution is therefore not that it “adds more models,” but that it assigns each component a bounded and non-redundant function within a coherent design architecture.

4.1.2. Theoretical Positioning of the Framework

The theoretical positioning of this study can therefore be summarized as follows. The dominant lens is coaching theory, operationalized through OSCAR. The supporting constructs—KSA, Situational Leadership, and KPI logic—do not replace OSCAR, but extend it in three bounded ways: they specify what capability is being developed, how guidance is tailored, and how progress is reviewed in organizational settings. AI workflow design, in turn, is not treated as a separate theoretical lens, but as the implementation logic through which this OSCAR-centric architecture is operationalized in prompts, state transitions, review loops, and auditable artefacts.

This positioning clarifies that the article’s novelty lies not in theorizing across many unrelated domains at the same level, but in extending a coaching-process lens with a small number of functionally bounded supporting constructs.

4.1.3. OSCAR Framework as the Process Scaffold

OSCAR (Outcome–Situation–Choices–Actions–Review) was selected after being compared with GROW and CLEAR because it transitions from problem analysis to solution construction, while maintaining structured reflection and accountability [13]. Using OSCAR, the researcher understands its meaning as follows:

○

Outcome clarifies desired end-states and success criteria.

○

Situation surfaces present realities, constraints, and strengths.

○

Choices broaden the set of options and anticipate consequences.

○

Actions convert intent into dated, resourced commitments.

○

The review establishes a cadence, provides feedback, and captures learning.

The research details guiding questions for each phase, as listed in

Table 1. Questions by OSCAR phase.

OSCAR Phase	Questions	Objective
Outcome	For short-term outcomes, the opening issue to be coached on: What would you like to talk about today? What would you like to cover today? What issues would you like to focus on? For the long-term outcome, some questions could be asked: What is your long-term desired outcome around this issue? What would success look like/feel like to you? How will you know you’ve achieved it? What will happen? Which outcome do you want to focus on? How important is it to you to achieve this outcome? What will it cost you if you don’t? What impact will not achieving this outcome have? What impact will achieving this outcome have? Over what period do you want to achieve this outcome?	Helping coaches and coachees effectively define and articulate their goals enables them to specify exactly what they want to achieve and what result they will focus on.
Situation	These are some questions that the coach may ask the individual to encourage discussion: What is the current situation? What do you consider the primary issues at present? Where are you now in terms of your goals? What impact is that having on you? (your family, your performance, etc.)? What impact is that having on others? Who is contributing to the problem? What are you doing that could be contributing to this problem/issue? What do you feel now, and how does that impact the issue? What do you do now, in this situation or a similar one, that works well for you? What made you aware that you need to do something different?	Enable coaches to identify and understand their current situation before planning for the future. The coachee begins by describing their context and evaluating their progress toward their goals, which helps build a clear image of where they are now. This enables the identification of critical issues and challenges that may be hindering progress, as well as their impact on employers and others.
Choices and Consequences	These are some questions that the coach may ask the individual to encourage discussion [13]: What can you do to start resolving the situation? What do you need to do or say to start resolving the situation? What choices do you have? What’s stopping you from doing that now? How far towards your desired outcome would that option take you? What advice would you give to somebody else in your position? What else could you do? Who else could help you? What’s stopping you from asking them? What would you like to do differently? What’s stopping you? What’s really stopping you? What would you do if you weren’t afraid? What would be the consequences? Upsides and downsides? What would you do if there were no downsides? What’s the worst thing that could happen? What’s the best thing that could happen? Which choice/choices will best move you toward your outcome? How far would that take you towards achieving your outcome?	Help coaches explore different options and strategies for achieving their goals.
Actions	These are some questions that the coach may ask the individual to encourage discussion: What actions are you going to take? What will you do to move yourself forward? What specific actions are you willing to take? When will you take them? What support will you need? Where will you find that support? How will you maintain your motivation? What actions will you take in the next 24 h to move forward? On a scale of 1 to 10, how committed are you to take this action? What’s stopping it from being 9 or 10?	Measure progress toward goals and clarify a concrete action plan. They help coaches reflect on previously established options and decide on the next steps simply and systematically.
Review	These are some questions that the coach may ask the individual to encourage discussion: How will you review your progress? What will you do to check whether your actions are moving you toward your outcome? How will you measure your success? How will you celebrate your success? When will you and I get together to review your progress? What would you like to tell me next time you see me? How will you maintain your momentum? What support do you need?	Assessing progress toward the goal, enabling a structured reflection on the actions taken and the lessons learned throughout the process.

, and argues that OSCAR’s focus on goal clarity, ownership, and regular review makes it highly suitable for long-horizon skill development and for AI mediation. For each phase, a specific number of questions exist, as listed in Table 1, defined based on [13], along with corresponding objectives.

OSCAR supplies the conversation backbone for an AI assistant, featuring phase-aware prompts, state persistence, and transitions that can be instrumented (e.g., transitioning from Outcome prompts to Situation probes, then to Choice enumeration and Action scheduling), along with automated Review reminders.

Although the OSCAR model was initially designed for a relationship between a manager and a coachee, in the proposed framework, this dynamic will be reinterpreted. In the context of this research, the role traditionally played by the manager will be taken on by an artificial intelligence engine. In other words, the AI will be responsible for guiding, supporting, and monitoring the development of the employees’ skills.

4.1.4. Supporting Constructs Extending the OSCAR-Based Framework

The following constructs are introduced as supporting extensions to the OSCAR-based framework rather than as coequal theoretical lenses. They provide additional specificity for competency targeting, adaptive support, and measurable review in organizational settings. These include the Knowledge-Skills-Abilities (KSA) model, the Situational Leadership theory, and Key Performance Indicators (KPIs). In this section, the relevance and contribution of these elements to the development of the proposed framework are described.

The Knowledge–Skills–Abilities (KSA) model provides a practical template for translating high-level outcomes into trainable targets, encompassing factual/procedural knowledge, demonstrable skills, and enabling abilities [14,15]. This research positions KSA as the bridge between coaching dialogue and development plans.

Situational Leadership [17] tailors support to the learner’s development level, ranging from D1 to D4 (enthusiastic beginner, disillusioned learner, capable but cautious contributor, and self-employed professional), by pairing it with styles S1–S4 (directive, persuasive, participative, and delegative). The research uses this to modulate how guidance is delivered, not just what is delivered: for example, more stepwise scaffolding at D1, and more autonomy and delegation at D4.

To ensure coaching translates into organizational value, the research adopts KPI practice [16] and the Balanced Scorecard Institute’s Measure–Plan–Review–Adapt (MPRA) cycle for performance management [35]. Good KPIs combine leading and lagging indicators, remain strategy-aligned, and are reviewed on a fixed cadence.

The role of each supporting construct, and its specific contribution to the OSCAR-centric framework, is summarized in

Table 2. Supporting constructs within the OSCAR-centric framework.

Component	Design Function in the Framework	Why Was It Selected	What Would Be Lost if Omitted?
OSCAR	Primary process scaffold	Provides a sequential coaching structure (Outcome–Situation–Choices–Actions–Review) that can be operationalized in prompts, state transitions, and review cycles.	The framework would lose its process spine and become a set of disconnected coaching prompts.
KSA	Competency specification layer	Translates broad development goals into explicit knowledge, skills, and abilities, linking coaching dialogue to trainable capability targets.	The framework would produce goals and actions, but with less precision about what capability is being developed.
Situational Leadership	Adaptive guidance layer	Adjusts the style and granularity of support according to competence and commitment (D1–D4/S1–S4).	The framework would remain structured, but the AI support would be less tailored to readiness differences.
KPI	Monitoring and governance layer	Converts coaching plans into measurable indicators, reviews cadence, and auditable follow-up aligned with organizational accountability.	The framework would support reflection and planning, but not systematic review, traceability, or organizational oversight.

.

Taken together, these constructs form a layered design architecture rather than a simultaneous theoretical fusion. They extend an OSCAR-based coaching process in bounded ways: KSA adds competency specificity, Situational Leadership adds adaptive guidance, and KPI logic adds measurable review and organizational accountability. This background underpins the framework evaluated later: OSCAR provides the primary theoretical lens and process spine, while KSA, Situational Leadership, and KPI logic function as secondary design extensions. This distinction is important because it clarifies that the framework does not treat the four components as redundant or interchangeable, but as bounded design layers addressing different requirements of AI-supported organizational coaching:

• Outcome-first orchestration. Begin with explicit success criteria and maintain a phase-aware state machine that enforces the OSCAR flow [13].
• Competence-centric planning. Translate goals into KSA targets to make development observable and coachable [16,22].
• Adaptive scaffolding. Detect D1–D4 profiles and adjust prompt depth, exemplars, and autonomy [17].
• Measurement by design. Co-define leading/lagging KPIs and embed MPRA reviews and audit trails [16,35].
• Human agency preserved. AI facilitates option generation and planning but leaves decisions and accountability with the coachee/organization.
• Reusability and governance. Express the framework as modular prompts/templates, incorporating data minimization, transparency notices, and review logs for audit purposes.

This background underpins the framework evaluated later: OSCAR provides the primary process spine, KSA specifies competency targets, Situational Leadership modulates the style of support where needed, and KPIs provide the monitoring and governance layer through which value can be reviewed.

4.2. Proposal

The proposal consists of a conceptual and operational framework that structures AI-supported organizational coaching as a staged process beginning with organizational context and ending with systematic review and measurement. The framework is organized around a primary process scaffold and three supporting layers. OSCAR governs the progression of the coaching sequence; KSA is activated when development targets need to be specified; Situational Leadership is activated when the AI must adjust the style of support to learner readiness; and KPI logic is activated when progress must be formalized into measurable and reviewable indicators.

This layered organization is intended to reduce conceptual overlap while preserving process coherence, adaptive support, and organizational accountability. Although the framework is presented through 10 ordered phases for design clarity and implementation purposes, it is not intended to operate as a rigid one-pass sequence. Instead, review evidence, contextual change, insufficient progress, or revised goals may trigger a return to earlier phases, making the framework cyclical and adaptive in line with coaching practice.

The framework embodies the role reversal adopted in this research: an AI agent performs many of the manager-coach functions (prompting, scaffolding, structuring, and monitoring), while humans retain final decision rights, accountability, and governance control.

More specifically, across the 10 phases, the AI may elicit information, organize dialogue, propose classifications, generate options, and draft artefacts. Still, it does not autonomously approve scope, validate development priorities, commit the coachee to an action path, authorize KPI governance, or decide escalation. These decisions remain with the employee/coachee and, where relevant, the organizational approver or manager.

For this framework, we define several key actors and artifacts. The actors are:

○

Company—provides strategy, mission, constraints);

○

Employee/coachee—supplies work context, chooses options, commits to actions);

○

AI coach/assistant—structures dialogue and outputs; adapts style by development level.

The key artifacts generated by the framework are also defined, as follows:

○

Organization profile, role profile;

○

SMART goal statement [34,36];

○

Situation analysis;

○

Development level (D1–D4) [17];

○

Option set and decision rationale;

○

Action plan;

○

KSA triplets [14,15];

○

KPI set and review cadence [16]

○

Running review log.

The proposal is organized as a 10-phase coaching cycle, as follows:

○

Step 1—Company Setup (Context Anchoring): The session begins by loading organizational information—sector, strategy, principles, and targets—so that subsequent dialogue and recommendations are context-aware rather than generic. This step ensures strategic alignment and reduces the need for later rework. Actor: company. Output: minimal org profile used by the AI’s prompts and safety rails.

○

Step 2—Role and Work Context (Clarification of Role Profile): The coachee provides their job title, unit, core responsibilities, and current projects. That informs the tailoring of examples, resources, and constraints in later phases. Actors: employee + AI. Output: role profile.

○

Step 3—Outcome (goal definition with SMART): Guided by the OSCAR step Outcome [13], the AI assists the coachee in crafting a SMART goal (specific, measurable, achievable, relevant, and time-bound), clarifying the value and costs of non-achievement. Actors: employee + AI. Output: SMART goal statement; initial success criteria.

○

Step 4—Situation (present-state analysis): Following OSCAR’s Situation, the AI prompts for the current status, constraints, enablers, and emotions affecting performance, thereby increasing self-awareness and surfacing hidden blockers. Actors: employee + AI. Output: concise situation map and candidate obstacles/resources.

○

Step 5—Level (development profiling via Situational Leadership): The AI administers brief questions on competence (experience, knowledge, practical skill) and commitment (motivation, confidence, initiative) to classify the coachee as D1–D4, then adjusts its style accordingly (S1 directive → S4 delegative). Actors: employee + AI. Output: development-level tag with Rationale; style settings for subsequent phases.

○

Step 6—Choices (option generation and consequences): Consistent with OSCAR’s Choices, the AI facilitates brainstorming, helps articulate pros/cons, and feasibility (including “what would you do if you weren’t afraid?” prompts), while preserving human agency—the coachee makes the decision. Actors: employee + AI. Output: option set, decision rationale, risk/assumption notes.

○

Step 7—Actions (commitment and planning): The chosen path becomes a dated, resourced action plan (near-term micro-steps + longer-term milestones). The AI elicits a commitment rating (1–10) and explores what would increase it, thereby enhancing the likelihood of execution. Actors: employee + AI. Output: action checklist with owners, dates, and support needs.

○

Step 8—KSA Generation (Competency Targeting): To anchor development, the AI proposes KSA triplets aligned with the SMART goal (knowledge to acquire, skills to practice/demonstrate, abilities to leverage), which the coachee can accept/edit. Actor: AI (with human confirmation). Output: approved KSA plan linked to actions.

○

Step 9—KPI generation (measurement by design): The AI co-creates KPIs—balanced leading and lagging indicators—and a review cadence (e.g., weekly micro-check-ins; monthly milestone review). KPIs map to the SMART goal and KSA plan. Actor: AI (with human confirmation). Output: KPI sheet (target, source, frequency, owner).

○

Step 10—Review (monitoring, learning, celebration, and re-entry): The Review phase operates as the framework’s main iterative control point. It has two layers: (a) scheduling of check-ins from the first session; and (b) ongoing evaluation that tracks actions and KPIs, surfaces drift, captures learning, and celebrates milestones to reinforce motivation. Crucially, Review does not only close the cycle; it may also reopen it. When progress is insufficient, contextual conditions change, the chosen option proves inadequate, or the original SMART goal needs revision, the framework returns to the relevant earlier phase (e.g., Outcome, Situation, Level, Choices, Actions, KSA, or KPI) rather than forcing a purely forward sequence. Actors: employee + AI. Output: review log; adjustments to actions/KPIs; risk note where relevant; re-entry decision where needed; next-session agenda.

The framework is summarily presented in

Table 3. Proposed Framework.

Step	Step Designation	Objective	Actor	Concept Used	Output
1	Company setup	Context anchoring	Company	—	The AI prompts and safety rails use the minimal org profile
2	Role and work context	Clarification of role profile	Employee + AI	—	Role profile.
3	Outcome	Goal definition with SMART	Employee + AI	OSCAR + SMART	SMART goal statement; initial success criteria
4	Situation	Present-state analysis	Employee + AI	OSCAR	Concise situation map and candidate obstacles/resources
5	Level	Development Profiling via Situational Leadership	Employee + AI	Situational Leadership	Development-level tag with rationale; style settings for subsequent phases
6	Choices	Option generation and consequences	Employee+ AI	OSCAR	Option set, decision rationale, risk/assumption notes
7	Actions	Commitment and planning	Employee + AI	OSCAR	Action checklist with owners, dates, and support needs
8	KSA generation	Competency targeting	AI (with human confirmation)	KSA	Approved KSA plan linked to actions
9	KPI generation	Measurement by design	AI (with human confirmation)	KPI	KPI sheet (target, source, frequency, owner)
10	Review	Monitoring, learning, celebration	Employee + AI	OSCAR + KPI

.

Although the framework already specifies governance controls and decision gates, the allocation of decision authority across the 10 phases also needs to be made explicit at the level of the coaching cycle itself.

Table 4. Decision authority across the 10-phase coaching cycle.

Step	Framework Phase	What the AI May Do	What Must Be Decided or Validated by Humans	Final Authority
1	Company Setup	Ingest organizational profile, apply predefined constraints, and prompt boundaries	Define strategy-relevant scope, permissible coaching topics, and governance constraints	Company/organizational approver
2	Role and Work Context	Elicit and structure role/context information; summarize inputs	Confirm role accuracy and relevance of the contextual information	Employee/coachee
3	Outcome (SMART goal)	Scaffold SMART formulation; propose refinements and clarification prompts	Select the actual development goal; validate relevance and, where applicable, approve the topic for organizational coaching	Employee/coachee; organizational approver where required
4	Situation	Prompt reflection: summarize current blockers, enablers, and context factors	Confirm the accuracy of the situation diagnosis and whether it is sufficient to proceed	Employee/coachee
5	Level (development profiling)	Propose a provisional D1–D4 classification and adapt style settings	Treat the classification as provisional and validate it if it affects planning, timeline realism, or approval	Employee/coachee; manager/approver where relevant
6	Choices	Generate options, articulate pros/cons, and surface feasibility considerations	Choose among options; reject unsuitable options; validate feasibility where resources, budget, or policy are implicated	Employee/coachee; manager/approver where relevant
7	Actions	Draft sequenced actions, milestones, and support needs	Commit to the action path; validate timeline, resources, ownership, and feasibility	Employee/coachee; manager/approver where relevant
8	KSA Generation	Propose KSA targets linked to the chosen development path	Confirm whether the proposed KSAs are relevant, realistic, and aligned with the intended development objective	Employee/coachee; organizational approver where formal capability mapping is required
9	KPI Generation	Draft KPI definitions, targets, sources, frequency, owner, and review cadence	Approve what will actually be monitored, by whom, how often, and under which governance rules	Employee/coachee; manager/approver
10	Review	Evaluate progress against KPIs, summarize drift, suggest adjustments, and flag escalation conditions	Decide whether to continue, revise, re-enter earlier phases, or escalate to human oversight	Employee/coachee for reflection and revision; manager/approver/human coach for escalation and high-stakes decisions

, therefore, distinguishes, for each phase, what the AI may do as a bounded facilitative agent, what must be decided or validated by humans, and who holds final authority at that stage.

4.2.1. Iterative Logic and Re-Entry Rules

Although the framework is described through 10 ordered phases, its operational logic is cyclical rather than strictly linear. This distinction is important because coaching rarely unfolds as a single forward pass; instead, new reflection, new evidence, and changing circumstances often require a return to earlier issues.

In the proposed framework, Review functions as the main recursive control point. At each review, the AI and coachee assess whether the current goal remains appropriate, whether the situational diagnosis is still valid, whether the selected option remains feasible, whether the action plan is realistic, whether the KSA targets still fit the developmental need, and whether the KPI set remains meaningful. If misalignment is detected, the process returns to the relevant earlier phase instead of advancing mechanically.

The framework, therefore, uses ordered phases for clarity, comparability, and implementability, but allows adaptive re-entry according to coaching needs. In practical terms, examples of re-entry include: returning from Review to Outcome when goals require reformulation; from Review to Situation when new blockers emerge; from Review to Level when competence/commitment is reassessed; from Review to Choices when the original option proves unsuitable; and from Review to Actions, KSA, or KPI when execution and measurement require adjustment.

This iterative design aligns the framework more closely with established coaching logic, where structured progression and recursive adaptation coexist.

Mapping the framework to its theoretical positioning clarifies that the four components are activated at different points and for different purposes. Steps 3, 4, 6, 7, and 10 follow the OSCAR process logic. Step 5 introduces Situational Leadership only as an adaptive mechanism for calibrating the style of guidance. Step 8 introduces KSA to convert the selected development path into explicit competency targets. Step 9 introduces KPI logic to make progress measurable, reviewable, and auditable. Thus, the framework should be read as a staged composition of bounded functions rather than as the simultaneous application of four full models at every step.

Operationally, the framework can be instantiated as a phase-aware state machine: the AI maintains the current phase, required inputs, and produced artefacts; it guides transitions between phases, preserves context for follow-ups, and supports controlled re-entry to earlier phases when review evidence or contextual change indicates that adaptation is needed.

Although full deployment-grade compliance testing remains outside the scope of the present study, the framework already specifies operational governance touchpoints: (i) company setup defines permissible scope and data boundaries; (ii) human decision authority is preserved in level, choice, and approval stages; (iii) KPI and review stages generate an auditable record of progress and interventions; and (iv) prompts, logs, and escalation rules provide a basis for disclosure, review, and later DPIA-style documentation.

A coaching cycle is considered complete when the following artifacts exist and are mutually consistent: SMART goal, situation summary, D1–D4 profile & style, decisioned option, action plan, KSA triplets, KPI set with cadence, and a review log with at least one completed check-in. These outputs are subsequently used for (i) prototyping with LLMs and (ii) practitioner evaluation.

4.2.2. Ethical Governance and Operational Controls

To make the framework operationally enforceable, ethics is implemented through explicit workflow controls rather than treated as a general statement of alignment with ICF or EU principles. In this framework, the AI coach is not an autonomous decision-maker; it is a bounded drafting and prompting agent operating under organizational rules.

At a minimum, the framework specifies the following controls:

• Disclosure and informed use. At the beginning of the interaction, the coachee is informed that the session is AI-supported, that outputs are advisory drafts rather than binding decisions, and that human oversight remains available.
• Data minimization. Only the minimum information required for the coaching task is processed: organizational context, role profile, goal, current situation, development level, action plan, and review-related progress notes. Sensitive personal data not required for the coaching objective should not be requested or retained.
• Scope restriction. The AI is constrained to the approved development topic and organizationally accepted coaching scope. If the conversation drifts into out-of-scope, sensitive, or high-risk matters, the system must flag this and redirect or escalate.
• Human validation and override. All key outputs—SMART goals, selected options, action plans, KSA maps, KPI sheets, and review decisions—are treated as draft artefacts subject to human confirmation. Final authority remains with the coachee and, where applicable, the organizational approver or manager. For clarity, this confirmation does not occur only at the end of the process; it is distributed across the coaching cycle at the specific phase-level decision points defined in Table 4.
• Auditability. The framework requires a persistent review log recording outputs, changes, KPI readings, review decisions, and escalation events. This log supports accountability, traceability, and later governance review.
• Bias and harms review. Generated outputs should be periodically checked for unfair assumptions, exclusionary recommendations, inappropriate tone, or strategically misaligned suggestions. In the present research, this control is specified at the framework level and highlighted as a required review activity, although full deployment-grade testing remains future work.
• Escalation pathways. The framework requires escalation to human oversight when outputs concern high-stakes issues, when repeated KPI stagnation suggests risk of inappropriate persistence, when scope boundaries are exceeded, or when ethical ambiguity arises.

In this way, ethical principles are translated into concrete process rules governing what the AI may do, what data it may use, how outputs are reviewed, and when humans must intervene.

Table 5. Operational ethics and governance controls are embedded in the framework.

Control Area	Operational Mechanism in the Framework	Purpose
Disclosure and informed use	Initial notice that the interaction is AI-supported and that outputs are advisory drafts subject to human validation	Ensures transparency and prevents misleading automation authority
Data minimization	Limit inputs to organizational context, role data, goal, situation, actions, KSA, KPI, and review-relevant progress data	Reduces unnecessary data exposure and supports privacy by design
Scope restriction	AI is limited to approved coaching topics and must flag or escalate out-of-scope or sensitive matters	Prevents role drift and constrains inappropriate use
Human validation and override	All major outputs require the coachee and/or organizational approval before adoption	Ensures accountability remains with humans
Auditability and logging	Review logs, KPI records, decision notes, and escalation events are retained as governance artefacts	Supports traceability, review, and organizational oversight
Bias and harms review	Periodic review of outputs for unfair assumptions, exclusionary suggestions, inappropriate tone, or strategic misalignment	Supports responsible monitoring of AI-generated recommendations
Escalation pathways	Defined triggers for referral to a manager, approver, or human coach	Provides a safety mechanism for ambiguous or high-risk cases
Role-based governance	Decision rights are assigned to the employee, AI, manager, and organizational approver at different stages	Clarifies accountability and prevents autonomous AI decision-making

describes in a systematic way the operational mechanisms in the framework.

To make this translation fully explicit,

Table 6. Traceability of ethical principles into executable framework controls.

Ethical Principle	Operational Meaning in This Study	Executable Control in the Framework	Where Enacted	Auditable Evidence
Transparency and disclosure	The user must know that the interaction is AI-supported and that outputs are advisory drafts	Disclosure and informed-use notice before or at the beginning of the interaction	Before Step 2/session start	Disclosure statement recorded in the session setup or system configuration
Human oversight and override	AI may structure and propose, but humans retain final authority over decisions	Human validation and override for goals, options, actions, KSAs, KPIs, and review decisions	Steps 3, 6, 7, 8, 9, and 10	Approved artefacts; review log; decision records
Scope restriction	AI must remain within approved developmental boundaries and flag drift	Scope boundaries and escalation rules are linked to the approved topic and organizational constraints	Steps 1, 3, 4, and 10	Scope statement; guardrail rule; flagged deviations; escalation note
Data minimization	Only the minimum necessary information should be processed for the coaching task	Restricted input structure limited to organizational context, role profile, goal, situation, development level, actions, KSA plan, KPI set, and review data	Throughout the workflow	Input schema; retained artefact list; logging policy
Auditability and accountability	Outputs and key decisions must be reviewable after the session	Persistent review log, KPI sheet, decision notes, and escalation events	Steps 9 and 10, plus governance logging across the cycle	KPI sheet; review log; escalation record; versioned artefacts
Bias and harms review	Outputs must be reviewable for unfair assumptions, inappropriate tone, or exclusionary recommendations	Periodic human review of competency assessments, recommendations, and review outputs	Steps 5, 8, 9, and 10	Reviewer comments, revised outputs; bias/harms review note
Escalation to human intervention	Sensitive, ambiguous, high-risk, or out-of-scope cases must be transferred to humans	Explicit escalation pathways to manager, approver, or human coach	Steps 3, 6, 9, and 10	Escalation trigger; escalation outcome; approver decision
Role-based governance	Decision rights must be allocated explicitly rather than left implicit	Decision authority is assigned across employee/coachee, AI, manager, and organizational approver	Across the 10-phase cycle	Decision-rights table; approved outputs; governance records

maps the main ethical principles referenced in the manuscript to the corresponding executable controls embedded in the framework, indicating where each principle is enacted and what auditable evidence it generates. This traceability is important because it shows that the framework does not merely declare alignment with ICF and EU guidance, but specifies concrete workflow mechanisms through which those principles can be implemented and reviewed.

4.2.3. Risk Controls and Governance Enforcement

To make the framework operationally enforceable, governance is implemented through explicit risk controls rather than through general ethical statements alone. In this framework, the AI coach is not treated as an autonomous decision-maker, but as a bounded drafting and prompting agent operating under organizational rules.

Five risk domains are addressed directly. First, hallucination risk is controlled by treating all AI-generated outputs—goals, KSA suggestions, KPI proposals, and action plans—as draft artefacts requiring human confirmation. The AI may structure and propose, but it does not decide. Second, privacy and confidentiality risk are controlled through data minimization: only the minimum information necessary for the coaching task should be processed (organizational context, role profile, goal, situation, development level, actions, KSA plan, KPI set, and review-related progress data). Sensitive personal information not required for the coaching objective should not be requested or retained.

Third, bias risk is controlled through review and override mechanisms. Competency assessments, development-level classifications, and recommended actions should be treated as provisional and subject to human review, especially where they influence timelines, approval, or perceived capability gaps. Fourth, overdependence risk is controlled by preserving human agency and limiting the AI to a facilitative role. The framework, therefore, requires that key outputs be accepted, revised, or rejected by the coachee and, where relevant, by the organizational approver. Fifth, escalation risk is controlled through explicit criteria for handoff to human oversight, including out-of-scope topics, repeated KPI stagnation, sensitive issues, ethical ambiguity, or concerns regarding feasibility, budget, or role boundaries.

In operational terms, these controls are supported by scope boundaries, review logs, KPI cadence, escalation rules, and auditable decision points. In this way, ethical and governance principles are translated into concrete process rules governing what the AI may do, what data it may use, how outputs are reviewed, and when human intervention becomes mandatory.

To make the governance layer of the framework more explicit and operational,

Table 7. Key risks in AI-supported coaching and corresponding governance controls.

Risk Area	Why It Matters in Coaching	Control Is Built into the Framework
Hallucination in developmental recommendations	Incorrect or unfounded advice may distort goals, actions, or perceived development needs.	AI outputs are treated as draft artefacts requiring human confirmation; review checkpoints and decision authority remain with humans.
Privacy and confidentiality of coaching conversations	Coaching may involve sensitive personal, interpersonal, or organizational information.	Data minimization, role-bound access, disclosure, and auditable logging are required; unnecessary sensitive data should not be requested or retained.
Bias in competency assessments or development-level classification	AI may generate unfair or exclusionary judgments about competence, commitment, or readiness	Assessments are treated as provisional; human review, override, and escalation are required where outputs affect approval, scope, or progression
Overdependence on automated coaching systems	Users or organizations may defer excessively to AI suggestions and weaken critical reflection.	The AI is restricted to a facilitative role; final acceptance of goals, plans, KSAs, KPIs, and reviews remains human.
Escalation threshold failures	Some cases exceed the appropriate scope of AI support and require human intervention.	Escalation rules are attached to scope boundaries, KPI review gates, feasibility concerns, and ethical ambiguity; managers/approvers intervene when thresholds are triggered

summarizes the main risk domains associated with AI-supported coaching and the corresponding controls embedded in the proposed framework. This mapping is intended to show that ethical and governance concerns are not treated only at the level of principles, but are translated into specific process safeguards linked to human validation, data minimization, auditability, review checkpoints, and escalation rules.

5. Prototyping and Demonstration

In the previous section, the proposal was designed and presented. In this section, a prototype will be developed and applied to use cases for evaluation.

5.1. Framework Instantiation

This section explains how the framework was made operational for the demonstration use cases. It summarizes the instantiation choices (tools, prompts, data, guardrails), the execution protocol, the comparison dimensions, and the artefacts captured so that other teams can reproduce the prototyping procedure under comparable conditions.

This section reports the prototyping stage with the level of detail required for replication. In particular, it specifies the models used, the prompting conditions applied across tools, the execution protocol followed in each run, the criteria used to compare outputs, and the limits of the comparison design. This stage aimed to assess framework instantiation and artefact-generation coherence rather than to conduct a formal benchmark of model superiority.

5.1.1. Implementation Environment and Tooling

The following tools and environment have been set up:

• LLM services—Three general-purpose LLM services were used to exercise the framework logic and check portability across providers: ChatGPT [ChatGPT 5.2], Gemini [Gemini 2.5], and DeepSeek [DeepSeek V4]. The tests were conducted using the publicly available free-tier versions accessible at the time of data collection. These models were not benchmarked against each other; the aim was to verify whether each service could follow the same coaching flow and generate the expected artefacts (SMART goal, D-level, Choices, Action plan, KSA, KPIs, Review plan). These were not benchmarked against each other; the aim was to verify that each model could follow the same coaching flow and generate the expected artifacts (SMART goal, D-level, Choices, Action plan, KSA, KPIs, Review plan). Accordingly, this stage should be interpreted as operational prototyping rather than as a formal experimental comparison of model performance. The focus was on framework instantiation, artefact generation, and procedural coherence, not on estimating comparative model superiority or intervention effectiveness. The purpose of this prototyping stage was therefore not to determine which model was more effective, nor to infer the effectiveness of AI-supported coaching as an intervention. Rather, it was to assess whether the framework could be consistently operationalized across tools and whether the expected artefacts could be generated with sufficient completeness and coherence.
• Workflow orchestration—A proof-of-concept n8n flow was built to automate the coaching conversation, preserve context, and render a simple interface. The pipeline used a Trigger node → Tools Agent seeded with the coaching prompt → LLaMA-3-70B-8192 → Memory node for state carry-over. At the prototype level, this setup demonstrated that the framework could be orchestrated as a traceable workflow and that coaching artefacts could be generated and preserved across steps. However, this layer should be interpreted as a workflow demonstration rather than as an enterprise-ready implementation. The prototype did not evaluate concurrent-user scalability, queueing, rate-limit handling, node failure recovery, provider fallback, access control, secrets management, monitoring dashboards, or deployment architecture constraints.
• Project context—A fictional firm (“CreativeTech”) provided an organizational backdrop (mission, values) to anchor the coach’s questions and suggestions in a business context, ensuring comparable runs across models.

The n8n automation component [37] should be interpreted as a proof-of-concept orchestration layer rather than as a production deployment. Its contribution in the present study is to show that the proposed coaching framework can be translated into a stateful, traceable workflow with explicit artefacts and review logic. This is an important feasibility result, but it is narrower than enterprise readiness.

For enterprise deployment, at least five technical dimensions would require further engineering and evaluation. First, scalability: the present prototype was not stress-tested under multiple simultaneous sessions, large user populations, or sustained organizational load. Second, robustness: the workflow was not evaluated against node failures, provider outages, partial-response handling, memory corruption, or interrupted sessions. Third, error handling: the prototype did not implement a full exception-management strategy for malformed inputs, incomplete outputs, timeout events, or escalation to fallback flows. Fourth, operational governance: deployment-grade implementations would require authentication, role-based access control, secrets management, observability, version control of prompts and flows, and audit-safe retention policies. Fifth, infrastructure and cost constraints: enterprise deployment would need to consider hosting architecture, latency, vendor dependency, model availability, and cost/performance trade-offs.

Accordingly, the present automation component demonstrates workflow feasibility, traceability, and orchestration logic, but not enterprise readiness in the strict technical sense.

To clarify the technical contribution of the automation layer without overstating its maturity,

Table 8. Technical interpretation of the n8n automation layer.

Dimension	What the Current Prototype Demonstrates	What Was Not Evaluated in the Present Study
Workflow orchestration	Phase-aware orchestration of the coaching flow through Trigger, Tools Agent, LLM, and Memory nodes	Production-grade orchestration under sustained organizational load
Traceability	Loggable artefact generation and visible workflow steps	Enterprise observability stack, monitoring dashboards, and incident management
State handling	Context carry-over through the Memory node	Long-session resilience, rollback logic, recovery from corrupted or incomplete state
Scalability	Feasibility of automated session flow at proof-of-concept level	Concurrent-user load, queueing, throughput limits, horizontal scaling
Robustness	Basic end-to-end execution in controlled prototype conditions	Node-failure recovery, provider outages, retry policies, fallback routing
Error handling	Implicit manual supervision during prototype use	Formal exception handling, timeout management, malformed-input handling, and automatic recovery
Security and access	General governance intent in framework design	Authentication, authorization, secrets management, environment hardening, and deployment isolation
Deployment readiness	Conceptual plausibility for organizational workflows	Full enterprise deployment architecture, cost/performance analysis, service-level validation

distinguishes between what the current n8n/LLaMA-3 prototype demonstrably supports and what remains outside the scope of the present study. This distinction is important because the automation component contributes evidence of workflow orchestration, traceability, and stateful session handling at a proof-of-concept level. Still, it does not yet provide a sufficient basis for claims regarding enterprise-grade scalability, robustness, security, or deployment readiness.

5.1.2. Prompt Scaffold Mapped to Framework Phases

A single phase-aware prompt instantiated the nine steps of the framework (work/role context → Outcome/SMART → Situation → Level/D1–D4 → Choices → Actions → KSA → KPI → Review). To ensure cross-tool consistency, the same prompt skeleton, organizational context (“CreativeTech”), role vignette, phase order, and control commands (REVIEW and DONE) were used across all three LLM services. No tool-specific prompt rewriting was introduced during the comparison runs. The intention was to hold the prompting conditions constant so that observed differences in outputs could be interpreted primarily as differences in model/service behavior rather than differences in prompt design. Two control commands standardized closure and evidence capture across models:

○

REVIEW—recap decisions, check progress blockers, refresh following actions.

○

DONE—generate a consolidated session report (goal, plan, KSA, KPIs, dates).

Prompt design followed the good practices of prompt-engineering principles (clarity, explicit outputs, logical flow, one step at a time, minimal bias) [38].

The prompts presented in

Table 9. Prompts to use.

Step	Framework Phase	Prompt
1	Objective and Company Data	Objective: The goal is to develop a framework that utilizes artificial intelligence to support the coaching process in a learning organization, thereby helping employees develop their skills. This framework will be applied to coach employees at CreativeTech. The AI coach should be friendly and empathetic, utilizing the “You-to-You” approach. CreativeTech’s Core Mission: “At CreativeTech, we are committed to helping our clients, people, and communities thrive. Working with the best, we offer a comprehensive range of services to diverse industries. Whether we’re helping transform organizations, protecting financial markets, or working with governments to support societies, the impact we make every day is profound and far-reaching.
2	Employee’s position in the company	The coaching framework follows a structured, step-by-step approach, guiding employees through the learning and skill development process. The framework includes six phases: Phase 1: Work Context Before starting the mentoring process, it’s essential to understand the person’s role and the context in which they work in order to tailor the questions and strategies more effectively Phase 1 Questions: What is your current role? Which department or area do you work in? What are the main responsibilities and tasks you perform on a daily basis?
3	Outcome	This phase focuses on identifying the desired outcomes and defining SMART objectives for the employee. SMART Objective Questions: 1. What specific outcome would you like to focus on? 2. What is the long-term goal you are trying to achieve? 3. How will you measure success in this area? 4. How important is this result to you, and why? 5. What impact would not achieving this result have on you or your role? 6. How will you know you’ve achieved your desired outcome? 7. What timeline are you aiming for to achieve this result?
4	Situation	This phase assesses the employee’s current knowledge, skills, and abilities (KSAs). Situation Questions: 1. What is the current situation you’re facing? 2. What do you think are the main challenges or obstacles right now? 3. How is this impacting your personal or professional life? 4. Who or what is contributing to this challenge? 5. What actions are you currently taking that might be contributing to the issue? 6. What has worked well in this situation or similar ones? 7. What made you realize that something needs to change? Realize what knowledge, skills, and abilities the person has.
5	Level	This phase evaluates the employee’s competence and commitment. Competence Assessment: 1. Have you performed this task before? 2. How often do you engage in this task? 3. How comfortable would you feel teaching this task to someone else? 4. How would you feel about solving an unexpected problem in this area? 5. Are you familiar with the best practices for completing this task efficiently?Commitment Assessment: 1. How motivated are you to improve in this area? 2. Would you be willing to take on a project related to this skill? 3. How confident are you in making decisions related to this task? 4. When an unexpected issue arises, do you try to solve it independently or seek guidance? 5. If no one supervised your work, would you continue to give it your best effort? Based on the answers, categorize the employee into one of the following profiles and adjust communication style accordingly: - D1—Enthusiastic Beginner (Directive) - D2—Disappointed Apprentice (Persuasive) - D3—Capable but Cautious Contributor (Participative) - D4—Self-Employed (Delegative)
6	Choices	This phase helps the employee explore various options and evaluate their pros and cons. Choice Exploration Questions: 1. What are the potential actions you could take to resolve the situation? 2. What’s stopping you from acting on those options? 3. What would be the consequences (both positive and negative) of each option? 4. Who could support you in taking action? 5. What would you do if there were no downsides to your decision? 6. Which option is most likely to help you achieve your desired outcome?
7	Actions	This phase focuses on creating an actionable plan and setting short-term goals. Action Planning Questions: 1. What concrete actions are you ready to take? 2. How committed are you to moving forward with these actions (on a scale from 1 to 10)? 3. When will you start acting, and what support will you need? 4. How will you maintain your motivation throughout this process? 5. What are your immediate next steps within the next 24 h? 6. What could prevent you from being fully committed to these actions?
8	KSA	In this phase, the AI will generate the Knowledge, Skills, and Abilities (KSAs) related to the objective you’ve set. These KSAs are crucial for your development and will help guide your learning path. The AI will analyze your goal and create a customized list of key skills and abilities (KSAs) required to achieve proficiency in the target area. Must show the KSA generated.
9	KPI Generation	In this phase, the artificial intelligence (coach) must generate specific KPIs to measure progress toward the objectives defined in the previous phase. These KPIs should be customized to the coachee’s goals, helping track development and ensuring the mentee understands how their progress will be evaluated.1. Analyze the Coachee’s Goal: The AI should review the goal set in Phase 1 (Results) 2. Generate Customized KPIs: Based on the goal, the AI creates specific KPIs that will be used to measure the coachee’s progress throughout the learning process. These KPIs may include indicators related to task completion, quality of work, time to completion, and the level of independence required for work. 3. Present the KPIs: The AI must present the generated KPIs to the coachee in a clear and detailed manner so that the coachee knows how their progress will be tracked and what metrics will be used to evaluate their growth.
10	Review	When the ‘REVIEW’ keyword is entered in capital letters, review questions should be asked, assuming that X amount of time has passed since the last conversation. To understand the person’s progress. Ask one question at a time This phase emphasizes reviewing progress, celebrating milestones, and measuring success. Review Questions: 1. How will you track your progress toward your goal? 2. When will you check in with yourself to assess whether you’re making progress toward the desired outcome? 3. How will you celebrate when you hit milestones or achieve success? 4. What support do you need from others to stay on track? 5. When will we meet again to review your progress? 6. What would you like to be able to report at that time? 7. Ask about the KPI that was generated in Phase 5.1 When the ‘DONE’ command is entered in capital letters, a report must be made of all phases. Note: During each phase, ask one question at a time, allowing the employee to reflect and respond before proceeding. Use their responses to inform the next question and tailor your advice based on their specific needs and goals accordingly. Only provide solutions if the employee struggles to answer. Start by asking for the name, and from now on, act like the coach and follow the framework. Ask one question at a time, pay attention to go through all the steps.

do not represent the full conceptual flexibility of the proposed framework. Instead, they were structured as a simplified sequential script of questions for prototyping purposes, so that all phases could be followed consistently under the constraints of the free LLM versions used in testing.

This methodological choice resulted from the limitations of the free versions of the tools used during testing: ChatGPT, Gemini, and DeepSeek, which presented relevant constraints such as:

○

Lack of long-term memory between extended phases;

○

Limited context capacity, which hindered coherence over more prolonged interactions;

○

Lower reliability in state management can lead to potential phase skipping or contextual loss.

Consequently, the adoption of a simpler, more linear prompt ensured that the framework could be executed coherently during prototyping. This should be understood as a methodological simplification for testing rather than as the full theoretical logic of the framework, which is intended to support adaptive re-entry and iterative adjustment across coaching cycles. Although the conceptual proposal envisaged using more advanced and adaptive prompts suitable for LLMs with broader context windows and memory functions, the practical implementation was intentionally simplified to align with the tools available at the time.

The test prompt is presented in Table 9, split by each framework phase, for better understanding.

To assess deployability beyond ad hoc chats, the same prompt and sequencing were wired into N8N, enabling: (i) stateful sessions (Memory node), (ii) traceability (node-level logs), (iii) scheduled reviews (timers), and (iv) a lightweight coaching UI. The automation reproduced the full flow and produced the same artefacts, demonstrating feasibility for enterprise environments that require audit trails. Figure 3, Figure 4 and Figure 5 depict the N8N [37] architecture, terminal trace, and conversation UI.

The AI proposes options and structures plans but does not decide, allowing the coachee to retain agency; the tone is supportive rather than prescriptive.

Sessions were logged for research evidence, and no sensitive personal data was required. The study follows the ethical considerations outlined (ICF/Trustworthy-AI guidance).

From a governance perspective, the prototype was intentionally bounded. The demonstration scenarios did not require sensitive personal data, and the retained artefacts were limited to framework outputs such as the goal, situation summary, D-level, options, action plan, KSA map, KPI sheet, and review log. Even so, the prototype should not be interpreted as eliminating core AI risks. Hallucinated recommendations, inappropriate competency inferences, overconfident KPI proposals, and boundary drift remain possible failure modes. For that reason, the framework assumes that humans review outputs and that review logs are retained for auditability. That escalation to human oversight is required when outputs appear implausible, ethically ambiguous, or operationally misaligned.

The N8N flow is: Trigger → Tools Agent → LLM → Memory. And repeat. From a governance perspective, the prototype’s data flow was intentionally limited. The Trigger node received the session input; the Tools Agent structured the phase-specific prompt; the LLM generated draft coaching outputs; and the Memory node preserved only the minimum conversational context required to continue the session coherently. The artefacts retained for review were limited to the framework outputs (goal, situation summary, D-level, options, action plan, KSA map, KPI sheet, and review log). No sensitive personal data was required for the demonstration scenarios. This architecture supports traceability and reviewability by making the data path, generated artefacts, and logging points explicit.

To assess deployability beyond ad hoc chats, the same prompt and sequencing were wired into n8n, enabling: (i) stateful sessions (Memory node), (ii) traceability (node-level logs), (iii) scheduled reviews (timers), and (iv) a lightweight coaching UI. The automation reproduced the full flow and produced the same artefacts, demonstrating workflow feasibility for environments that require traceability and reviewability. However, this should not be interpreted as evidence of full enterprise readiness, because the prototype was not evaluated for scale, resilience, advanced error handling, or deployment hardening.

5.2. Test Data and Use Case Setup

Two role-specific scenarios were prepared so the same framework could be tested in distinct domains. These scenarios were intentionally fictional and were used as structured prototyping cases rather than as empirical field interventions. The prototyping stage cannot determine whether the generated KPI sets are valid indicators of real developmental progress, nor whether the framework adds measurable value relative to baseline approaches such as unguided AI usage, framework-free prompting, or human-only coaching support. Accordingly, they do not support direct measurement of coaching impact, quantitative/statistical outcome analysis, or comparison with alternative approaches such as human-only coaching, business-as-usual development practice, or non-AI coaching interventions. Their purpose was to test whether the framework could be applied consistently across different role profiles and development goals, not to evaluate coaching outcomes in real organizational settings:

• Use case 1 (Noah)—Senior UX manager balancing usability rigor with sprint constraints; the desired outcome was an MVP-grade testing pipeline aligned with product cadence. Outputs compared in Table/Figure summaries for use case 1 in Table 11.
• Use case 2 (Emily)—React developer transitioning to Angular under time pressure; the desired outcome was a time-bounded upskilling plan with demonstrable productivity milestones. Outputs compared for use case 2 in Table 12.

Each run started with identical context snippets (role, current project constraints) and the same prompt skeleton, so differences in outputs reflected model behavior rather than setup variance. For each model-service and each use case, the comparison was based on a single complete run under fixed prompting conditions. Repeated trials were not conducted in this prototyping stage. Accordingly, the outputs should be interpreted as representative instantiations of the framework under controlled conditions, not as estimates of within-model variance or stability across repeated runs. Three generative-AI tools were used: ChatGPT, Gemini, and DeepSeek, in their respective free versions.

At this stage, an execution protocol has been established and followed:

• Initialize session (LLM chat or N8N UI) and paste the framework prompt.
• Inject CreativeTech context + role vignette (UX manager or frontend developer).
• Walk through the phases in order. The agent asks one phase at a time, discourages compound questions, and requires the user to confirm completion before advancing.
• At the level, the model gathers indicators of competence (experience, knowledge, practical skill) and commitment (motivation, confidence, initiative) to assign D1–D4 and adapt tone/detail (directive → delegative).
• Choices → Actions turn options into a dated, resourced plan; collect a commitment rating and “what would raise it” nudges.
• KSA generation (knowledge to acquire, skills to practice; abilities to leverage) is tied to the specific goal and plan.
• KPI generation with leading/lagging indicators and review cadence (e.g., weekly micro-check-ins; monthly milestone review).
• Trigger REVIEW and DONE to record the final recap/report for analysis.
• Export artefacts (SMART goal, D-level, plan, KSA, KPIs, Review plan) for cross-model comparison.

All steps mirror the operational flow documented and the framework logic in the previous Section.

A run was considered a successful instantiation if it produced the full minimum artefact set required by the framework:

• A SMART goal with clear business relevance;
• A D-level classification with rationale;
• At least three options with pros/cons and a selected option;
• A dated action plan with owners/support;
• A KSA triplet (or set) aligned to the goal and actions;
• A KPI set with targets, sources, and frequency; and
• A review plan and a session report via DONE.

Subjectivity in judging outputs was controlled in three ways. First, all tools were evaluated against the same predefined artefact criteria and acceptance rules. Second, the comparison focused primarily on explicit output features (presence of required fields, alignment across phases, reviewable KPI structure, and action-plan specificity) rather than on general impression. Third, comparative summaries in Table 6 and Table 7 were written at the level of observable output characteristics (e.g., tone, structure, level classification, KPI type) rather than claims of superiority. However, because no blinded inter-rater scoring exercise was conducted at this stage, some interpretive subjectivity remains and should be treated as a limitation of the prototyping design.

These criteria were used as framework-instantiation checks rather than as evidence of coaching effectiveness. In other words, they indicate whether the model can produce outputs that are structurally complete and usable within the proposed coaching architecture, not whether those outputs are necessarily superior, behaviorally effective, or practically optimal.

To make the prototyping stage more transparent and reproducible,

Table 10. Replication details for the prototyping stage.

Element	Reported Condition
LLM services used	ChatGPT; Gemini; DeepSeek
Model version reporting	[Insert exact model labels if visible in the interfaces; otherwise, report provider free-tier default model as accessed on date]
Use cases	Use case 1: Noah (UX manager); Use case 2: Emily (frontend developer)
Prompting condition	Same prompt skeleton, same organizational context, same role vignette structure, same phase order, same REVIEW and DONE commands
Tool-specific prompt adaptation	None during comparison runs.
Run structure	[Single-run per model/use case OR insert repeated-trial design]
Comparison dimensions	Artefact completeness; internal coherence; procedural adherence; field-level actionability
Primary artefacts assessed	SMART goal; D-level; options set; action plan; KSA map; KPI set; review plan
Subjectivity control	Predefined artefact criteria; same protocol across tools; comparison based on observable output features; no blinded inter-rater scoring

summarizes the key procedural conditions under which the comparison across LLM services was conducted. The table consolidates the essential replication details—models used, prompting conditions, run structure, comparison dimensions, assessed artefacts, and subjectivity controls—so that the prototyping logic can be interpreted consistently and, where possible, reproduced under comparable conditions.

To make the prototyping stage more explicit methodologically, the generated outputs were examined against three evaluation dimensions. First, artefact completeness: whether the required fields of each framework deliverable were present. Second, internal coherence: whether outputs generated in later phases remained consistent with earlier outputs (e.g., whether the KSA plan aligned with the SMART goal and action plan, and whether KPIs matched the stated development objective). Third, procedural adherence: whether the model preserved the intended phase order and transitioned through the framework without skipping critical stages. These dimensions support assessment of framework operationalization, but they do not replace expert judgment, longitudinal outcome evidence, or formal benchmarking.

These deliverables match the expected outputs in the Operational application of the proposed framework and were used to compare ChatGPT/Gemini/DeepSeek outputs in both use cases, as presented in

Table 11. Analysis of Use Case 1.

Step	Framework Phase	ChatGPT	Gemini	DeepSeek
1	Company setup	(Minimal AI role; assumes predefined organizational context)	(Minimal AI role; assumes predefined organizational context)	(Minimal AI role; assumes predefined organizational context)
2	Role and work context	Formal and structured language; profile D3—Capable but Cautious	Technical and detailed language; profile D4—Self-Employed	Technical and action-oriented language; profile D4—Self-Employed
3	Outcome	Defines goal: learning UX research, workshops, and continuous testing	Defines hybrid goal (UX + continuous testing)	Defines a goal based on comparison with React to facilitate adaptation
4	Situation	Context and skill gap analysis identifies UX and agile integration challenges.	More technical description focused on usability gaps and iteration	Emphasizes framework transition and adaptation barriers
5	Level	D3—shows moderate confidence and caution	D4—demonstrates autonomy and self-direction	D4—demonstrates self-sufficiency and results orientation
6	Choices	Generates multiple options (workshops, continuous testing, stakeholder focus)	Suggests rapid integration and MVP testing	Recommends a practical comparison between React and Angular for decision-making
7	Actions	Defines a minimum viable plan, pilot for the next sprint, and stakeholder alignment	Defines MVP testing, rapid integration, and alignment session	Defines workshop, evaluation criteria, recap document, and plan B supported by ROI data
8	KSA Generation	Knowledge
		- UX Methods - Stakeholder - Communication - Agile UX Integration - Usability Testing - Methodologies - Data Analysis & Synthesis	- Lean UX Principles - Agile Development - Usability Testing - Methodologies - Stakeholder - Management - Data Analysis and Synthesis	- Agile UX Research - Stakeholder - Management - Rapid Testing Tools
		Skills
		-	- Facilitation - Communication - Time Management - Problem-Solving - Adaptability	- Workshop Facilitation - Insight Synthesis - Metrics Advocacy
		Abilities
		-	- Strategic Thinking - Collaboration - Critical Thinking - Influence - Organization	- Balancing Rigor vs. Speed - Influence Without - Authority
9	KPI Generation	KPIs: testing time reduction, stakeholder satisfaction, testing frequency, insight actionability	KPIs: UX testing cycle time, adoption of UX recommendations, continuous testing frequency	KPIs: testing cycle time, stakeholder adoption, insight quality score
10	Review	Weekly check-ins, team celebration	Weekly check-ins, team brunch	Weekly check-ins, team celebration

and

Table 12. Analysis of Use Case 2.

Step	Framework Phase	ChatGPT	Gemini	DeepSeek
1	Company setup	(Minimal AI role; assumes predefined organizational context)	(Minimal AI role; assumes predefined organizational context)	(Minimal AI role; assumes predefined organizational context)
2	Role and work context	Technical and structured tone; profile D2—Disappointed Apprentice	Structured and motivational tone; profile D1—Enthusiastic Beginner	Inspirational and direct tone; profile D1—Enthusiastic Beginner
3	Outcome	Defines a logical and measurable goal focused on mastering Angular structure	Defines clear objectives with visualization of progress and coaching tone	Defines goal through React–Angular comparison to ease adaptation
4	Situation	Identifies knowledge gaps in Angular core concepts (dependency injection, lifecycle hooks)	Maps technical weaknesses and learning potential in TypeScript and RxJS	Focuses on adaptation barriers from React to Angular and learning speed
5	Level	D2—moderate competence but low confidence	D1—beginner profile, high enthusiasm	D1—beginner with fast learning potential
6	Choices	Suggests structured learning: exercises → Todo App → complex projects	Suggests guided path: tutorials → small projects → avoid React comparisons	Suggests a comparative project, React vs. Angular (Login Flow)
7	Actions	Defines a sequential roadmap focusing on progression and feedback	Defines learning milestones tracked weekly on Trello	Defines a self-commitment plan using GitHub activity as motivation
8	KSA Generation	Knowledge
		- Angular architecture, - Dependency injection, Routing, Forms and validation, RxJS, Lifecycle hooks	- Angular Architecture - TypeScript - RxJS - Angular CLI - Angular Lifecycle Hooks - Dependency Injection - Angular Modules - Angular Services - Routing	- Angular Architecture - RxJS Fundamentals - State Management
		Skills
		- Angular architecture, - Dependency injection, Routing, Forms and validation, RxJS, Lifecycle hooks	- Component - Development - Template Syntax - Form Handling - HTTP Communication - State Management - Debugging - Testing - Problem Solving	- Translating React - Patterns - Debugging Angular - Reactive APIs
		Abilities
		- Problem solving - Debugging - Collaboration - Project structuring	- Adaptability - Collaboration - Continuous Learning - Time Management - Communication - Critical Thinking - Self-directed Learning	- Architectural Thinking - Performance - Optimization - Coaching-readiness Potential
9	KPI Generation	KPIs: Component development mastery, Routing proficiency, Dependency Injection understanding, RxJS use, Test coverage, Code review feedback	KPIs: Independent task completion, Code review quality, Contribution to projects, Learning roadmap progress	KPIs: Task independence, RxJS proficiency, Speed-to-market
10	Review	Fixed review scheduled for the following week	Weekly Trello check-ins	Weekly review based on deliverables

. We treat these outputs as measurable deliverables rather than conversational by-products: each run either produces the full deliverable set or fails the instantiation criteria. This allows replication and tool-to-tool comparison using deliverable completeness and field-level quality checks (e.g., KPI sheet includes target/source/frequency/owner; review cadence is stated).

These demonstrations should be interpreted as single-run instantiations of the framework rather than as longitudinal coaching interventions. Their purpose was to test whether the framework could generate complete and coherent artefacts across phases, not to assess whether the resulting plans produced sustained performance improvement, behavioral change, or skill retention over time.

The approaches proposed by ChatGPT, Gemini, and DeepSeek offer systematic and personalized blueprints for skill acquisition in Angular, each with a unique focus. ChatGPT is more technical and future-focused in tone, Gemini focuses on learner motivation and autonomy, and DeepSeek uses a comparison to React as a scaffolding approach.

All the alternatives, despite their differences, agree that the definition of goals is clearly established by utilizing performance indicators (KPIs) and continuous monitoring. This approach leads to the potential use of artificial intelligence in supporting coaching processes within organizational learning environments.

5.3. Prototyping Results Summary

These case studies illustrate how the framework can be instantiated in practice and how its phases articulate throughout the coaching process. However, they should be interpreted as prototyping evidence rather than as a rigorous experimental evaluation of output quality or coaching effectiveness.

Although the free versions of the AI engines have limitations that prevent implementing all the details provided in the framework, the examples presented demonstrate the framework’s application potential and the practical plausibility of AI-supported coaching workflows.

Based on the tests conducted, the proposed framework appears practically viable and feasible for organizational contexts at a prototyping level. While certain limitations were identified—such as the need to regulate the communication tone, the inability to achieve depth in some stages, and the challenge of simulating aspects like scheduled reviews—the overall structure and sequence of the framework were maintained mainly as intended.

AI showed a tendency to ask several questions at once, which can compromise the flow of the conversation. The incorporation of the organizational context, although simplified, enabled the AI to generate contextual responses. The logic of progression between phases, although sometimes requiring intervention, was workable. At the same time, the tests also highlighted that real coaching interactions benefit from more adaptive looping than the simplified prototyping script could fully express, reinforcing the need to distinguish the linear test prompt from the framework’s broader iterative logic. These results suggest that enhancing the tools used, particularly by adopting more advanced AI versions, may enable a more comprehensive implementation of the framework and support more robust technology-enabled competency development processes.

In the tests conducted with the framework for adopting artificial intelligence in the coaching process, we aimed to evaluate not only how the AI tools perform but, more importantly, how the framework operates in practice. This includes its applicability across different organizational contexts and its ability to generate coherent results. The tests were conducted using three primary tools: ChatGPT, Gemini, and DeepSeek. These tools were applied in two distinct scenarios within the fictional company CreativeTech—one involving Noah, a UX manager, and the other with Emily, a frontend developer transitioning to Angular.

In all tests, the tools followed the framework’s stages: starting with the assessment of the work context, defining objectives and outcomes, mapping the challenges faced, analyzing the current level of competence, exploring available options, building the action plan, automatically generating the required knowledge, skills, and attitudes (KSAs), producing AI-driven key performance indicators (KPIs), and finally, setting up a follow-up and review plan. In addition to these tests, an automated experiment was conducted using N8N, where an automated coaching flow was configured with LLaMA3.

This experiment showed that it is possible to create a transparent and traceable structure capable of simulating automated follow-up and advisory sessions at a proof-of-concept level. However, the automation result should be interpreted narrowly. It demonstrates orchestration feasibility and artefact traceability, not production-grade deployment maturity. Important technical dimensions remain underdeveloped in the present study, including concurrent-session scalability, workflow robustness under failure conditions, explicit exception handling, infrastructure hardening, and deployment constraints linked to provider dependency, latency, and operational governance.

While a set of guiding questions was provided, the AI systems can (and should) adjust their approach according to the specific needs of each scenario, offering even more personalized and context-sensitive coaching.

Regarding the role of AI throughout the framework, it is evident that the AI intervenes continuously from Step 2 to Step 10, although with varying degrees of intensity. In some phases, the AI assumes a more passive or supportive role, organizing and structuring the user’s responses (e.g., in the Role & Work Context and Situation phases).

In others, its role becomes more analytical or collaborative, assisting in defining SMART goals, assessing development levels, or structuring action plans.

Finally, in later stages—such as KSA (Knowledge, Skills, and Abilities) generation and KPI (Key Performance Indicator) generation—the AI assumes a more active and generative role, producing new outputs for user validation.

Overall, the tests demonstrated that the framework is practical, applicable, and versatile as a design artefact and prototyped workflow. It adapts well to different contexts, enabling AI tools to provide not just generic advice but structured and personalized coaching. The use of automation further reinforces the model’s potential relevance for organizational workflows by making processes more traceable and metrically structured at the prototype level. Even so, a stronger technical evaluation is still required before making claims about enterprise-grade efficiency, resilience, or deployment readiness.

The demonstration supports a narrower set of claims, namely:

• Operational completeness—For both roles, the framework generated the expected artefact stack from goal setting to monitoring (SMART goal, D-level, option set, action plan, KSA map, KPI set, and review plan), indicating that the phases can be instantiated coherently across different LLMs.
• Procedural coherence—Across models and use cases, the framework logic remained recognizable and the intended phase sequence was largely preserved, suggesting that the process design is stable enough to be operationalized in software.
• Organizational plausibility—The generated outputs were sufficiently structured to support later governance-oriented assessment (e.g., review cadence, measurable fields, and auditable artefacts), which is relevant for organizational implementation.

These findings should not be interpreted as formal evidence of output superiority, coaching effectiveness, or benchmarked system performance. Rather, they indicate that the framework can be instantiated consistently enough to justify more rigorous future evaluation. This limitation also applies to the KPI outputs. Although each model generated KPI sheets with targets, sources, frequency, and ownership fields, the study did not test whether these indicators were more valid, more useful, or more predictive than KPIs produced through unguided AI use, human coaching, or ordinary managerial planning. The current evidence therefore supports the framework’s capacity to generate structured monitoring artefacts, but not the empirical validity of those artefacts as performance measures in practice.

Despite the limitations encountered during testing, further refinement and enhancement, particularly in unlocking AI’s adaptive capacities, could allow the framework to achieve an even greater impact and deliver more robust and meaningful results.

At the same time, these demonstrations do not establish long-term coaching value in the full developmental sense. Because coaching typically unfolds through repeated reflection, action, review, and adaptation cycles, stronger evidence would require longitudinal observation of whether the proposed plans lead to sustained behavioral change, competency growth, retention, and progression. This remains outside the scope of the present study and should be treated as a central direction for future empirical validation. Also, the prototype should not be interpreted as a complete compliance implementation. The present study demonstrates how ethical principles can be translated into scope boundaries, review logs, human validation, and auditable workflow steps. Still, it does not yet provide full enterprise-grade bias testing, formal DPIA documentation, or production-level access-control validation. These remain important next steps for deployment research.

6. Preliminary Qualitative Validation

In this Section, the preliminary qualitative validation of the framework is described following its prototype demonstration.

The proposed framework was subjected to a preliminary qualitative validation through a focus group (FG) study designed to: (i) assess the clarity and flow of the framework phases and guiding questions; (ii) judge its perceived usefulness for competence development in learning organizations; and (iii) elicit perceptions of AI acting as a coach (including empathy and acceptance). Accordingly, the focus group was used to assess perceived clarity, usability, acceptance, and refinement needs, rather than to measure effectiveness or impact in practice. This validation should be interpreted as conceptual and design-oriented rather than outcome-based. It concerns clarity, flow, usefulness, governance value, and organizational plausibility, but it does not establish behavioral change, performance improvement, or real-world coaching impact.

This validation should be interpreted as preliminary and design-oriented. It concerns the framework’s clarity, logical flow, perceived usefulness, governance value, and deliverable structure, rather than the direct effectiveness of AI-supported coaching in improving employee performance relative to alternative coaching approaches.

6.1. Design and Participants

We adopted a synchronous online focus group format [39] because it provides rich interaction at a low cost, facilitates the inclusion of geographically dispersed experts, and accelerates scheduling and data capture.

The methodological Rationale—contrasting asynchronous versus synchronous modalities—is documented in the dissertation and grounded in the literature synthesized from the focus group studies. Six practitioners were purposively recruited from coaching, psychology, learning-organization practice, and human resources, as listed in

Table 13. Selected Participants’ Profile.

Participant ID	Professional Domain	Experience Level	Highest Degree	Relevance to Framework Evaluation
P1	Coaching practice	10–15 years	Bachelor	Practical view of coaching flow, dialogue quality, and coachee engagement
P2	Psychology	15–20 years	Master	Relational, empathy, and behavioral change perspective
P3	Learning organizations/e-learning	30+ years	Master	Learning design, capability-building, and organizational learning perspective
P4	Human resources management	15–20 years	Bachelor	Governance, approval, role-boundaries, and organizational implementation perspective
P5	Information systems/higher education	30+ years	Ph.D.	Process structure, digital systems, and evaluation perspective
P6	Artificial intelligence research	20–25 years	Ph.D.	AI capability, technical risk, and responsible-use perspective

.

Across the six participants, the sample represented 129 cumulative years of professional experience (mean = 21.5 years; range = 12–32 years). Recruitment was purposive and criterion-based: participants were selected to ensure informed critique from complementary perspectives relevant to the framework, namely coaching practice, psychology, learning organizations, HRM, information systems, and AI. Because the participant pool was small and professionally specialized, the reporting strategy prioritized non-identification: participants are referred to only by anonymized IDs (P1–P6), and no one-to-one mapping between participant IDs and personally identifying professional details is disclosed beyond the grouped profile shown in Table 13.

The composition of the focus group was intentionally expert-oriented because the objective of this phase was to obtain an informed critique on framework design, governance, and organizational plausibility. However, this sampling strategy also limits generalizability: the findings reflect expert perceptions rather than the experiences of end-users operating in real coaching contexts, and they should not be interpreted as representative of broader organizational populations.

This expert-oriented composition was appropriate for design critique and governance-oriented appraisal, but it also limits generalizability. The focus group findings should therefore be interpreted as an informed qualitative evaluation rather than as representative evidence of broader end-user populations. The purpose of this phase was therefore depth of expert feedback rather than breadth of representation, and the resulting findings should be interpreted as analytically informative but not statistically or organizationally generalizable.

6.2. Focus Group Execution

Materials have been provided in advance. Participants received a short contextualization video, the framework’s phases and question set, and a list of evaluation prompts. The nine evaluation prompts used during the session are listed in

Table 14. Focus Group Evaluation prompts.

Evaluation Prompt	Description
1	Perceived importance and clarity of objectives
2	Clarity and logical organization of phases
3	Smoothness of transitions
4	Contribution to competence development
5	Support for reflection/planning of specific competencies
6	Support for setting realistic learning goals and actions
7	Views on AI “as a coach.”
8	Perceived empathy/understanding
9	Recommendations for future work.

.

The focus group was conducted via Microsoft Teams; the session was audio-recorded and transcribed for research purposes. No personally identifiable information is reported. The session took place on 22 July 2025 and lasted 120 min. All participants provided informed consent for recording and analysis, and transcripts were pseudonymized prior to coding.

The audio recording was transcribed and pseudonymized before analysis. Qualitative analysis was conducted in NVivo 15 using an inductive coding strategy aligned with the focus group’s purpose. In a first coding cycle, segments of text were coded openly to capture recurring judgments, concerns, refinement suggestions, and governance-related observations. In a second cycle, related codes were consolidated into higher-order categories and sub-categories corresponding to the main analytical dimensions of the study, including phase clarity, flow, competence development, governance, empathy, and future design recommendations.

To strengthen dependability, a codebook was maintained during the analysis process, including code names, short definitions, inclusion/exclusion guidance, and exemplar excerpts. Coding decisions, code merges, and theme refinements were documented through analytic memos and versioned revisions of the coding structure. NVivo’s AI-assisted summarization was used only as a supportive condensation tool after human coding decisions had been made; it was not used as a substitute for coding or interpretation.

Because this focus group was exploratory and involved a small qualitative dataset, the study does not claim a formal statistical inter-rater reliability coefficient. Instead, trustworthiness was strengthened through transparent coding documentation, iterative codebook refinement, researcher discussion of theme boundaries, and triangulation with the prototyping findings and framework artefacts.

While the focus group prompts primarily assessed clarity, flow, and perceived usefulness, we interpret the evidence through a proposed organizational-governance lens. In this sense, the framework is designed to support three types of mechanism: (i) more explicit organizational decision points (e.g., approval, scope boundaries, review gates), (ii) more standardized deliverables that can be audited and acted upon (SMART goal, KSA map, KPI set/cadence, review log), and (iii) clearer governance allocation regarding who validates what, and when escalation is required. This framing aligns with the framework’s explicit role separation—AI structures dialogue and outputs, while the organization retains decision rights and governance control. However, these mechanisms should be interpreted as proposed design features supported by expert appraisal rather than as empirically confirmed organizational outcomes.

Accordingly, the focus group instrument was suited to assessing perceived design quality and governance implications, but not to measuring real-world usability, organizational adoption behavior, or downstream coaching impact. Those dimensions require field deployment with actual participants, repeated use over time, and outcome-based measurement.

6.3. Results

From the six focus group participants, an overall expert appraisal of the framework’s relevance, design quality, governance needs, and refinement priorities was collected. The results reported below, therefore, reflect informed perceptions of framework clarity, usefulness, and organizational plausibility rather than broad population-level generalizations.

Participants were unanimous that the research topic is timely and relevant, and several noted the framework’s potential to improve specific, well-scoped tasks such as structuring action plans and summarizing sessions. However, these observations should be read as expert perceptions of design usefulness rather than as evidence of actual superiority in real organizational use.

We examine the framework’s proposed organizational-governance relevance (not only feasibility) through two analytical mechanisms: (i) decision governance, i.e., whether the framework is designed to structure who decides what, when, and based on which evidence (Table 9), and (ii) auditable production, i.e., whether the framework is designed to produce measurable artefacts that could support scale and oversight (Table 11). At this stage, these should be understood as proposed mechanisms judged plausible by expert participants, not as confirmed organizational outcomes. Value is realized when the framework reduces ambiguity, prevents premature scaling, and creates an evidence trail that managers can use for oversight.

In this study, “organizational value” refers to the framework’s capacity to structure decision-making, produce auditable deliverables, and support governed implementation. It does not refer to directly measured organizational performance gains or validated coaching outcomes.

In this research, feasibility and effectiveness are treated as distinct constructs. Feasibility refers to whether the framework can be executed coherently and produce auditable outputs within an organizational workflow. Effectiveness would require evidence that these outputs lead to meaningful improvements in developmental or organizational outcomes. The present results support the former, not the latter.

Accordingly, value in this study is treated as organizational and procedural value—namely, the framework’s capacity to structure decisions, produce auditable outputs, and support governed implementation—not as a direct measure of intervention effectiveness at the employee-performance level.

6.3.1. Organizational Decisions Influenced by the Framework

The focus group evidence suggests that the proposed framework may offer organizationally relevant governance mechanisms primarily by structuring and reshaping decision-making, rather than only improving the feasibility of AI-supported coaching conversations. Participants repeatedly emphasized the need for governance, oversight, and organizational involvement, indicating that the framework’s outputs could serve as triggers for explicit decision rights and decision gates (Table 10). These observations support the plausibility of the framework’s governance design, but they do not by themselves establish empirically demonstrated organizational value.

Across the decision set in Table 10, four decision classes emerged as value-critical:

• Readiness gating (do not proceed without KSA baseline + role clarity),
• Scope boundaries (what the AI coach may and may not do; escalation conditions),
• Evidence cadence (KPI frequency, review gates, and manager checkpoints),
• Accountability & escalation (who signs off, who intervenes, and what triggers escalation).

These decisions are proposed governance mechanisms intended to reduce adoption risk by turning AI coaching into an auditable organizational process rather than an informal tool.

This supports the study’s design stance that AI can structure dialogue and generate artifacts, but humans retain decision authority across the coaching journey.

(1)

Skill approval and intake decision (go/no-go to proceed): Participants stressed that the coaching process must begin with a validated focus on the right skill. They recommended that skills should either be selected from a company-validated set or, if proposed by the employee, should require a justification and SMART framing, followed by explicit organizational validation before proceeding to downstream planning.

In decision terms, the framework converts “skill selection” from a personal preference into an organizational gate that ensures strategic alignment and reduces the risk of investing effort in low-priority development topics.

(2)

Scope boundaries decision (what the AI is allowed to coach on): The framework also changes decisions by making the scope of coaching explicit and governable. Focus group feedback highlighted the importance of establishing guardrails so that AI remains within the selected development scope and can flag when conversation drifts beyond agreed boundaries.

This transforms scope from an implicit assumption into a concrete governance decision that can be reviewed and enforced.

(3)

Timeline realism decision (approving SMART timelines based on competence evidence): Participants cautioned against defining timelines before establishing whether the target competence level is realistic for the employee’s current knowledge state. One recommendation was to include a technical knowledge assessment before fixing timelines and milestones.

Accordingly, the framework operationalizes timeline setting as an evidence-based decision: timelines are approved only after competence gaps are assessed, and the SMART objective is grounded in feasibility, reducing planning risk and later rework.

(4)

KPI cadence and review-gates decision (when to measure, review, and re-enter earlier phases): Organizational control is further embedded through decisions about review cadence and corrective gates. Participants recommended “regular check-ins” and KPI readings, with explicit allowance to re-enter earlier framework phases when evidence indicates misalignment or insufficient progress. This reframes evaluation from an end-point audit to a continuous governance mechanism (Plan–Do–Check–Act logic), strengthening accountability and adaptability. This point is theoretically important as well as operationally useful. It indicates that practitioners did not interpret the framework as a rigid sequence, but as a structured cycle in which review evidence can legitimately trigger a return to earlier phases. This feedback directly informed the revised interpretation of the framework as iterative and adaptive rather than merely linear.

(5)

Escalation and oversight decision (when managers intervene; budget/effort thresholds): Finally, focus group input emphasized that responsible AI coaching requires clear escalation rules. Participants proposed explicit guardrails that specify when managers should intervene, including budget and effort thresholds, and recommended the presence of a designated approver to validate development proposals.

These decisions institutionalize oversight and ensure that AI-supported coaching remains aligned with organizational constraints, ethics, and resource planning.

Taken together, these five decision categories suggest that one proposed source of the framework’s organizational relevance lies in converting coaching into a governed decision process with explicit decision rights, triggers, and accountability mechanisms, rather than treating AI coaching as a self-contained, employee-only interaction.

Table 15. Organizational decisions changed by the framework.

Decision (What Changes)	Who Decides (Decision Rights)	Framework Trigger/Artifact (What Prompts the Decision)	Decision Rule (Measurable/Auditable)	Focus Group Evidence
D1. Skill approval/intake (Go/No-Go to proceed)	Organization (designated approver); employee proposes, organization validates	Skill selection + justification, SMART draft (Step 3)	Proceed only if (i) skill is from the company-approved list OR (ii) employee-proposed skill includes justification + SMART framing and is explicitly validated by the organization; otherwise, stop/revise	Practitioners recommended company-validated skills or employee-proposed skills requiring justification + SMART and organizational validation; emphasis on a designated approver and organizational involvement
D2. Scope boundaries (what the AI is allowed to coach on)	Manager/organization sets boundaries; AI enforces within-session; employee operates within scope	Scope statement + guardrails attached to the coaching goal (Step 3–4)	Proceed only if scope is explicitly defined; when conversation/output drifts beyond scope, flag and return to scope (or trigger escalation if repeated)	Need for guardrails so AI stays within the selected scope and flags drift
D3. Timeline realism (approve SMART-T only after competence evidence)	Manager/approver validates feasibility; employee/AI proposes	Pre-assessment of current competence/knowledge + SMART goal (Step 3–4)	Approve timeline only if competence evidence exists (e.g., assessment performed and gap stated); otherwise, revise timeline/milestones	Recommendation to add a technical knowledge assessment before fixing timelines
D4. KPI cadence & review gates (when to measure, review, and re-enter earlier phases)	Manager/organization sets cadence and gate rules; employee executes; AI supports reminders/structure.	KPI sheet + review plan (Step 7–10)	KPI plan must specify target, source, frequency, and owner; reviews occur at agreed cadence; if KPI readings show insufficient progress, trigger a review gate: adjust plan and re-enter earlier phases (goal/plan/skill selection)	Practitioners emphasized regular check-ins and KPI readings and the ability to re-enter earlier phases; KPI fields include target, source, frequency, and owner
D5. Escalation & oversight (when managers intervene; budget/effort thresholds)	Manager/organization (oversight); employee escalates; AI flags thresholds	Escalation rule + thresholds attached to plan and review log (Step 9–10)	Escalate to manager/approver when (i) KPI stagnates for N cycles, (ii) effort/budget threshold exceeded, or (iii) scope/risk concerns arise; escalation outcome must be recorded in the review log	Need for guardrails specifying when managers should intervene, incl. budget/effort thresholds; call for a designated approver
D6. Decision authority (governance principle that applies across all decisions)	Humans retain final authority; AI supports structure and drafts artifacts	All framework outputs (SMART goal, KSA map, KPI sheet, review log)	AI output is treated as a draft; final acceptance requires human validation (employee + approver/manager as appropriate)	Framework stance that AI provides structure, but humans retain decision authority

lists the organizational decisions changed by the framework.

The table presents a proposed operationalization of governance-by-design: each framework artefact is intended to trigger a governed decision with clear decision rights and auditable rules. These governance decisions are also the points at which key risks can be managed in practice. Scope boundaries help contain inappropriate or out-of-domain recommendations; review gates create opportunities to detect implausible, stale, or biased outputs; escalation rules define when AI support is no longer sufficient; and retained human decision authority reduces the risk of overdependence on automated coaching suggestions. In this sense, the framework’s governance structure is not only procedural, but also risk-mitigating by design.

6.3.2. Measurable Deliverables Produced by the Framework

Beyond shaping organizational decisions, the focus group evidence and the prototyping protocol indicate that the framework is designed to support organizationally relevant deliverables by producing a repeatable set of auditable outputs rather than open-ended conversational responses. The framework explicitly specifies the key artefacts it produces (e.g., SMART goal, action plan, KPI sheet, review log), and the prototyping stage operationalized “successful instantiation” as the ability of the AI-supported implementation to generate these deliverables in a complete and usable form. These findings support the plausibility of the framework’s governance design, but they do not yet demonstrate realized organizational value in practice. The framework explicitly specifies the key artifacts it produces (e.g., SMART goal, action plan, KPI sheet, review log), and the prototyping stage operationalized “successful instantiation” as the ability of the AI-supported implementation to generate these deliverables in a complete and usable form.

In this sense, reliability is interpreted narrowly as deliverable completeness and field-level auditability (i.e., whether the framework output contains the information needed for organizational validation, governance, and follow-up), rather than as a direct measure of coaching effectiveness. This notion of reliability is intentionally narrower than longitudinal effectiveness. In the present study, reliability refers to whether the framework consistently produces complete, auditable, and decision-ready outputs. It does not yet refer to whether those outputs translate into sustained behavioral change, improved retention, or measurable progression over extended coaching cycles.

Table 16. Framework deliverables as measurable organizational outputs.

Deliverable	Minimum Required Fields (Measurable/Auditable)	Acceptance Rule (Binary Pass/Fail)	Generated in Framework Step(s)	Evidence Basis
SMART goal specification	Specific outcome; measurable indicator; target; timebox; relevance/alignment statement; baseline (if available)	Pass if all SMART fields are explicitly stated; Fail if any field is missing/ambiguous.	Step 3	Artifact list includes SMART goal; required in a successful instantiation set.
Competence target (D-level rationale) and justification	Current competence state; target state; rationale for gap; evidence/assessment input (e.g., quick knowledge check)	Pass if current vs. target and rationale are explicit; Fail otherwise	Step 4	“D-level and justification” included as framework output
Options set + development-path decision	At least 2–3 options; selection rationale; constraints considered (time, cost, access); decision owner	Pass if a clear choice is made with rationale and constraints; Fail if output remains exploratory only.	Step 5	Framework specifies options + decision rationale as outputs
Action plan (development plan)	Activities; sequencing; timeline/milestones; required resources; responsible parties (who does what)	Pass if plan includes activities + timeline + responsibilities; Fail if missing any element.	Step 6	The action plan is listed as a key artifact, included in “successful instantiation” deliverables.
KSA map (Knowledge–Skills–Attitudes mapping)	KSA elements linked to target competence; mapping to planned activities; gaps to be addressed	Pass if KSA map links competence target to activities; Fail if KSA is generic or unlinked.	Step 7	KSA map is listed as an artifact; included in “successful instantiation.”
KPI sheet (measurement plan + cadence)	KPI definition(s); target; source; frequency; owner; measurement method	Pass if the KPI sheet includes target/source/frequency/owner for each KPI; Fail otherwise	Step 8	Step output requires a KPI sheet including target/source/frequency/owner; KPI deliverable included in “successful instantiation”
Review plan and review log (governance record)	Review dates; KPI readings; decisions taken (continue/adjust/escalate); next actions; approver/manager sign-off when required	Pass if at least one review cycle is specified and the log structure supports audit; Fail if no governance record is produced.	Steps 9–10	Review and follow-up are explicit steps; review/reporting is included in the successful instantiation set.

lists the minimum required fields and acceptance rules.

We therefore define reliability in this study as follows: for a given step, the framework produces an output that meets the minimum required fields and acceptance rules (Table 16) without relying on tacit coach knowledge. This is a framework-operationalization criterion, not a substitute for expert rating, inter-rater agreement analysis, or downstream performance measurement. This matters for scaling conceptually: SMEs typically lack spare managerial bandwidth, and auditable outputs (SMART goal, KSA map, KPI cadence, review log) are proposed as a governance trail that could make delegated oversight more feasible.

6.3.3. Synthesis: Decision-Gating and Auditable Deliverables as Value Mechanisms

Analyzing the achieved results allowed us to identify a set of contributions and perspectives that informed a detailed expert appraisal of the proposed artefact’s design quality, governance plausibility, and refinement needs (

Table 17. Topics discussed by the focus group and additional remarks that have arisen.

Topic	Remark
The phases were considered clear, logically organized, and reflective of a natural coaching conversation. However, participants requested more explicit interdependence between phases and smoother “gates” (micro-checkpoints) to reduce mechanical transitions. They also recommended adding a technical-knowledge assessment before fixing timelines, so SMART “T” estimates are grounded in current competence.	Future directions.
Stakeholders stressed the need for stronger organizational involvement—beyond a single touchpoint—to validate need, feasibility, timelines, and alignment with the department’s mission. They recommended clearer guardrails (e.g., when managers should intervene, budget/effort thresholds, and how the organization constrains AI recommendations) to assure strategic fit.	Future directions.
The framework makes the company a first-class actor: it supplies mission, vision, values, departments, roles, and coaching policies (e.g., time budget), and validates what will (and will not) be coached. This creates pre-validated coaching paths aligned with strategy. A designated approver in the company validates development proposals.	Confirmation of our research.
The framework introduces a good goal-setting flow: (a) Predefined, company-validated skills the employee can pick (via referral by the company or the employee’s initiative), and (b) Employee-proposed skills, where the employee must justify relevance, draft a SMART objective, and submit for organizational validation before proceeding. Access to skills is limited to the person’s role.	Confirmation of our research.
The framework emphasizes AI’s role in raising situational awareness for the coachee.	Confirmation of our research.
The framework adds AI-generated technical knowledge tests tailored to the selected skill (grounded in the Knowledge component of KSA), plus deeper commitment probing (motivation, effort, alignment with work). The output maps to situational leadership levels D1–D4, on a firmer diagnostic basis.	Confirmation of our research.
The framework constrains proposals to methods previously attempted and proven by the organization, so options are not only personalized but also operationally feasible and aligned with rules/resources.	Confirmation of our research.
The framework cross-references objective data with employee profile and industry/role requirements to produce a more customized KSA map.	Confirmation of our research.
The framework formalizes regular check-ins, uses KPI readings to track progress, and explicitly allows re-entry to earlier phases when adjustments are needed—tightening plan-do-check-act discipline.	Confirmation of our research.
It will be necessary to specify that the company recognizes thematic areas (to ensure strategic alignment), but not the confidential content of one-to-one coaching. Also need to instruct the AI to keep conversations within the selected scope and to flag drift.	Future directions

).

To increase transparency in the qualitative reporting,

Table 18. Representative insights from the focus group (anonymized/paraphrased).

Theme	Representative Insight	Implication for Framework Refinement
Phase clarity and flow	“The phases follow a natural coaching logic, but the transitions need smoother gates so the process feels less mechanical.”	Add micro-checkpoints and clearer transition logic between phases.
Timeline realism and competence evidence	“SMART timelines should not be fixed before checking whether the target competence level is realistic for the employee’s current knowledge state.”	Add a technical-knowledge or competence-baseline check before timeline approval.
Governance and organizational involvement	“The organization needs a stronger role in validating scope, feasibility, timelines, and intervention thresholds.”	Strengthen approver role, scope boundaries, and governance gates.
Scope boundaries and confidentiality	“The company may validate the thematic area, but not the confidential content of one-to-one coaching conversations.”	Differentiate organizational governance of the topic/scope from private coaching content.
AI empathy and tone	“The AI can be useful, but tone sometimes feels impersonal or overly effusive.”	Add tone/style parameterization and preserve coachee centrality.
Review, escalation, and iteration	“KPI reviews should allow return to earlier phases and should define when manager or human-coach intervention becomes necessary.”	Make review gates and escalation logic more explicit in the framework.

summarizes representative insights associated with the main themes that emerged from the focus group. Because the sample is small and professionally specialized, the insights are presented in anonymized and, where necessary, lightly paraphrased form.

Perceptions of AI empathy were mixed. In some cases, the AI engaged helpfully; in others, it felt impersonal or overly effusive. Participants suggested parameterizing tone and style to reflect corporate identity and ensuring the coachee remains the protagonist (the AI should not dominate the conversation). The automatic session summary was valued for organizing next steps.

Evaluators encouraged testing “negative cases” (e.g., unmotivated or resistant employees) and personalized profiles to reflect heterogeneous motivational triggers and performance drivers across populations.

In Figure 6, a heat map illustrates the references for the different evaluation prompts, comparing these references.

The focus group study provided actionable expert feedback indicating that the framework is clear, useful, and timely as a design proposal, while also revealing essential refinements for organizational governance, competence realism, conversational flow, and AI persona/role.

These insights substantively informed Version 2, strengthening the framework’s practical fit for learning-organization contexts and its responsible use of AI in coaching.

Overall, practitioner feedback indicates that the framework may create value not merely through conversational quality, but through its capacity to convert coaching into governed organizational action. The framework operationalizes this by (i) introducing approval and scope gates (skill validation, guardrails, escalation rules), (ii) producing auditable deliverables (SMART goal, KSA map, KPI sheet with cadence and ownership, review log), and (iii) enabling structured re-entry/adjustment when KPIs show drift.

7. Discussion

The results reported in Section 5 and Section 6 support three main interpretive points. First, the prototyping stage suggests that the proposed framework is operationally feasible. Across multiple LLM services and two use cases, it was possible to instantiate the framework and generate the expected artefact stack. Second, the focus group results indicate that expert participants perceived the framework as clear, governance-aware, and potentially useful for structuring coaching-related decision processes. Third, the combined evidence suggests that the framework is better understood at the present stage as a design artefact with organizational plausibility than as a validated intervention with demonstrated impact.

At the same time, these findings must be interpreted within the limits of the study design. The prototype runs relied on fictional scenarios and free-tier LLM services, while the qualitative validation relied on a small expert-oriented sample. Accordingly, the current study supports claims about feasibility, coherence, design quality, and governance relevance, but not about long-term coaching outcomes, behavioral change, or organizational performance impact.

The results also clarify an important conceptual point: the framework’s contribution lies not only in structuring AI-supported coaching conversations, but in making key artefacts and decision points explicit. This strengthens its relevance for organizational governance, especially where review cadence, scope boundaries, escalation, and auditable deliverables matter. However, these should still be understood as proposed and preliminarily appraised mechanisms rather than empirically confirmed organizational outcomes.

A further implication concerns the distinction between generated structure and validated impact. In the present study, KPI sheets, review cadence, and action plans were generated successfully as part of the framework outputs. Still, they were not empirically tested as predictors of behavioral change, competency development, or business performance.

Likewise, the comparison across ChatGPT, Gemini, and DeepSeek remained descriptive and framework-oriented, not comparative in the experimental sense. As a result, the current paper supports the plausibility of the framework as a structured coaching architecture, but not yet its added value over baseline approaches such as unguided AI usage, framework-free prompting, or alternative coaching arrangements.

This includes the absence of a direct comparison with traditional human coaching: without a control condition, the present study cannot determine whether the framework performs better, worse, or differently in terms of effectiveness, efficiency, or user satisfaction.

Future research should therefore move in three directions. First, field-based deployment studies are needed to assess real-world usability with organizational users rather than fictional scenarios. Second, longitudinal studies are needed to evaluate behavioral change, competency development, KPI attainment, and progression across repeated coaching cycles. Third, a stronger comparative and technical evaluation is needed to assess output quality, inter-rater agreement, deployment robustness, and governance performance under real conditions.

7.1. Empirical Validation Agenda

A logical next step for this research is a small-scale controlled study designed to test the added value of the framework beyond fictional scenarios. One suitable design would compare at least two conditions: (i) participants using the proposed framework to complete a coaching-planning task; and (ii) participants using either unguided AI prompting or a business-as-usual/human-only planning approach. A stronger design could add a third condition based on hybrid human–AI coaching.

Such a study could assess measurable differences in at least five dimensions: (i) artefact completeness, (ii) internal coherence of the resulting plan, (iii) quality and usability of generated KPIs, (iv) perceived usefulness and trust, and (v) expert-rated actionability. Where feasible, repeated measures across review cycles could also assess whether KPI readings, review logs, and action plans correspond to actual competency progression or behavioral change over time.

This kind of controlled validation would allow the framework to be evaluated not only as a coherent design artefact, but also as a practically useful intervention whose outputs can be compared against alternative approaches. It would be especially valuable for testing whether the KPI layer adds genuine monitoring value rather than merely producing well-structured documentation.

7.2. Managerial Implications: A Staged Implementation Roadmap

For organizations considering the adoption of AI-supported coaching, the findings of this study suggest a staged implementation approach rather than immediate large-scale deployment. The framework developed here is especially suitable for pilot-based introduction because it combines structured coaching logic with auditable artefacts, decision gates, and review mechanisms. Based on the framework design, prototyping results, and focus group feedback, a practical implementation roadmap for companies can be organized into four stages, as listed in

Table 19. Suggested managerial implementation roadmap for AI-supported coaching.

Stage	Managerial Objective	Key Actions	Minimum Outputs/Controls
1. Pilot definition	Start with a bounded and manageable use case	Select one development area, one employee segment, and one limited coaching objective; define expected use of the framework.	Approved pilot scope; selected skill/topic; initial SMART goal template
2. Governance setup	Establish decision rights and organizational guardrails	Define approver role, scope boundaries, validation rules, escalation thresholds, and human override responsibilities	Scope statement; approval rule; escalation criteria; role/responsibility map
3. Monitored execution	Run coaching cycles with auditable follow-up	Use the framework to generate action plans, KSA maps, KPI sheets, and review logs; schedule review cadence	Action plan; KSA map; KPI sheet (target/source/frequency/owner); review log
4. Ethical safeguards	Protect confidentiality, fairness, and responsible use	Implement disclosure, data minimization, restricted access, human validation, and escalation for sensitive or ambiguous cases	Disclosure statement; minimum-data rule; human validation checkpoint; escalation pathway
5. Controlled scale-up	Expand only if pilot evidence supports it	Review pilot outputs, governance performance, and user feedback before extending the scope	Scale-up decision; revised governance rules; updated monitoring cadence

: pilot definition, governance setup, monitored execution, and controlled scale-up.

In the pilot phase, organizations should begin with a narrow scope rather than with broad enterprise-wide use. A suitable starting point is one validated development area, one limited employee segment, and one clearly bounded coaching objective. At this stage, the goal is not to automate coaching in full, but to test whether the framework produces usable outputs such as SMART goals, action plans, KSA maps, KPI sheets, and review logs under real organizational constraints.

In the governance setup phase, organizations should define who approves the coaching topic, who sets scope boundaries, who reviews outputs, and when escalation to a manager, approver, or human coach becomes mandatory. This is particularly important because the study indicates that AI-supported coaching becomes more organizationally plausible when decision rights, review cadence, and escalation thresholds are made explicit rather than left informal.

In the monitored execution phase, organizations should require KPI sheets and review logs for every coaching cycle. At minimum, KPI monitoring should specify target, source, frequency, and owner, while review logs should capture progress, drift, decisions taken, and next actions. This allows organizations to move from one-off coaching interactions to a controlled cycle of review, adjustment, and accountability.

In the ethical safeguards phase, companies should implement minimum governance controls before scaling: disclosure that the interaction is AI-supported, data minimization, role-bound access to coaching artefacts, human validation of outputs, and explicit escalation for sensitive, ambiguous, or out-of-scope cases. In practical terms, AI should be treated as a bounded facilitative agent that structures dialogue and drafts outputs, while humans retain final authority over decisions and developmental commitments.

Taken together, this staged roadmap suggests that the most realistic managerial use of the framework is as a governed pilot that can gradually mature into a broader organizational capability if review evidence, user experience, and governance capacity support further scaling.

8. Conclusions

This work proposed and iteratively refined a conceptual framework for integrating Artificial Intelligence (AI) into organizational coaching to support the development of technical and behavioral competencies aligned with business strategy.

The study combined literature review, framework design, prototype demonstration, and preliminary qualitative validation. At the current stage, the contribution is best understood as the presentation and early appraisal of a structured, governance-aware framework rather than as conclusive evidence of coaching effectiveness or organizational impact. Across design, operational trials, and stakeholder feedback, the framework demonstrated prototype-level feasibility and conceptual usefulness as an instrument to structure conversations, guide action planning, and support monitored review.

The final version of the framework consolidates 10 interrelated phases that together enable a guided, iterative, and measurable development journey. Although these phases are presented in an ordered form for clarity, the framework is designed to support recursive movement across phases whenever review evidence, contextual change, or developmental needs make re-entry necessary. This structure promotes personalization, consistency, and strategic alignment at both individual and organizational levels.

Conceptually, the framework is grounded primarily in coaching theory through OSCAR, while KSA, Situational Leadership, and KPI logic operate as bounded supporting constructs. This clarification is important because it positions the study not as a broad synthesis of unrelated theories, but as an OSCAR-centric extension for organizational AI-supported coaching.

Preliminary validation combined: (i) operational application with a custom GPT-based assistant (ChatGPT Plus) to simulate coaching sessions and assess flow, completeness, and actionability; and (ii) an automation proof-of-concept using n8n with LLaMA 3 to demonstrate traceable, auditable follow-up sessions in a controlled workflow setting. Together, these demonstrations showed that the framework is practical and adaptable at the prototype level, and that it can be translated into an automatable workflow architecture. However, they do not establish real-world coaching impact, organizational effectiveness, or broad deployment validity. At the same time, the governance and organizational-value claims advanced in this study should be interpreted carefully. The present evidence supports the plausibility and expert appraisal of the framework’s proposed governance mechanisms—such as decision gates, auditable deliverables, escalation rules, and review structures—but it does not yet demonstrate that these mechanisms produce confirmed organizational outcomes in real deployment settings.

A focus group with practitioners provided preliminary qualitative validation of the framework’s clarity, governance relevance, and perceived usefulness, while also recommending deeper question tiers and stronger organizational controls. Because this phase relied on a small expert-oriented sample, the resulting evidence should be interpreted as design appraisal rather than as proof of usability or impact in real organizational settings. These insights directly informed the refinements in future research.

This research suggests that AI can serve as a plausible design basis for more structured and scalable coaching support when embedded in a governed organizational process. When embedded in a structured approach, AI assists in goal clarification, progress tracking, competency diagnosis, and KPI-based monitoring, thereby extending access to development opportunities beyond what is economically feasible with human-only coaching.

Human-centered guardrails remain essential. In the revised framework, these guardrails are not treated only as normative principles, but as explicit executable controls linked to disclosure, validation, scope restriction, logging, bias review, and escalation. Current foundation models (e.g., ChatGPT, Gemini, DeepSeek) still lack the depth of empathy and contextual flexibility expected from professional coaches; without explicit mechanisms to bind outputs to evolving strategy, there is a risk of strategic drift. The framework, therefore, emphasizes coachee autonomy and organizational oversight and recommends adherence to International Coaching Federation (ICF) ethical guidelines to preserve trust and integrity.

Organizational integration is proposed as a mechanism that may strengthen governance quality, implementation readiness, and practical plausibility. Focus group feedback highlighted the need for the organization to validate needs, timelines, scope, and alignment, as well as to incorporate technical assessments where appropriate steps that enhance the realism and impact of the development plans produced through AI-supported conversations.

From a practical standpoint, the study also proposes a staged implementation roadmap for companies, emphasizing pilot-based adoption, explicit governance setup, KPI-based monitoring, and minimum ethical safeguards before broader organizational scale-up.

The present work prioritized framework design and pilot-level validation. As such, it did not include outcome-based measurement of behavioral change, performance improvement, learning transfer, or longitudinal developmental progression, nor did it include quantitative/statistical analysis of impact or controlled comparison with alternative coaching approaches. This is a significant boundary condition because feasibility and effectiveness are not equivalent: the present study shows that the framework can be instantiated coherently and can generate auditable deliverables, but it does not establish that these outputs improve real-world coaching outcomes. Therefore, the current contribution is best understood as demonstrating conceptual, operational, and organizational plausibility rather than effectiveness.

As limitations, it should be noted that prototype evaluations used free LLM tiers and that the qualitative validation relied on a small, expert-oriented focus group. This sampling strategy was appropriate for design critique and governance-oriented appraisal, but it limits generalizability and does not capture the experiences of end-users in real organizational settings. Accordingly, the findings from the focus group should be read as perceptions of framework quality and organizational plausibility rather than as evidence of actual usability, behavioral change, or performance impact. Future versions will have to use deeper technical checks at this level. A further limitation is that the prototyping stage did not include a formal expert-rating benchmark with anonymized outputs, standardized scoring rubrics, or inter-rater agreement analysis. Accordingly, the results show that the framework can be operationalized and that structured outputs can be generated, but they do not establish comparative model quality or robust output-performance rankings. In addition, the study did not include an empirical ablation design to isolate the contribution of each framework component separately. The present contribution is therefore a theoretically justified and operationally tested layered architecture.

In contrast, the relative marginal contribution of OSCAR, KSA, Situational Leadership, and KPI logic remains a priority for future comparative research. A further limitation is that the background review in Phase 1 was not reported as a fully reproducible PRISMA-compliant systematic review with explicit search strings, screening counts, and a formal flow diagram. Future work should strengthen this phase through a complete review protocol and selection-flow reporting. Another limitation is that the n8n/LLaMA-3 automation layer remained a proof-of-concept.

Although it demonstrated workflow traceability and stateful orchestration, the study did not evaluate scalability under concurrent use, robustness under failure conditions, formal exception handling, security hardening, or deployment architecture constraints. Accordingly, the automation results should be interpreted as evidence of workflow feasibility rather than enterprise readiness.

A further limitation is that the manuscript proposes governance mechanisms and organizational-value pathways, but does not empirically test whether these mechanisms translate into realized organizational outcomes under real deployment conditions. A further limitation is that the prototyping comparison relied on fixed prompting conditions and artefact-based assessment, but did not include blinded inter-rater scoring or repeated-trial variance analysis across all model outputs. Therefore, the comparison supports replicable framework instantiation rather than full benchmarking of model performance.

As future work, immediate priorities include controlled empirical studies comparing framework-guided use with baseline approaches such as unguided AI prompting, business-as-usual development practice, human-only coaching, or hybrid human–AI coaching arrangements. Such studies should examine not only artefact quality and expert-rated actionability, but also whether the KPI structures generated by the framework correspond to real developmental progress, behavioral change, and organizationally relevant outcomes over time. Additional priorities include head-to-head evaluations against business-as-usual coaching, stronger linkage of KSAs and KPIs to business performance, and the operationalization of ethical governance through audit trails, bias testing, and escalation protocols.

Further work should extend the present governance layer into full deployment-grade ethical implementation, including ICF-aligned disclosure protocols, formal data-flow documentation, auditable review logs, structured bias/harms testing, role-based access controls, and DPIA-style accountability documentation. Thus, the present contribution should be read as ethics-informed and operationally governance-aware rather than as a complete compliance implementation of ICF or EU ethical requirements.

Other aspects relevant for future work include deployment studies in varied organizational contexts (size, industry, coaching culture), outcome-based evaluation, and a dedicated technical assessment of the automation layer. This technical agenda should include concurrent-session testing, workflow resilience under failure conditions, explicit error-handling and fallback logic, authentication and access control, observability and audit pipelines, prompt/version governance, and cost-performance analysis for deployment at organizational scale.

Overall, the framework offers a structured, ethically grounded, and organization-aware pathway to leverage AI in coaching, complementing (not replacing) human expertise and expanding access to evidence-informed development at scale. The synthesis of design logic, stakeholder feedback, and operational piloting provides a solid foundation for larger-scale, outcome-oriented trials and for embedding AI-augmented coaching into learning-organization practices.