Workplace AI coaching needs rules before results

Researchers have proposed a new framework for using artificial intelligence (AI) in organizational coaching, insisting that AI can help companies scale employee development only if human oversight, auditability and clear decision rights are built into the process from the start.

The study, titled Adoption of Artificial Intelligence in Organizational Coaching Processes, was published in AI. It presents a governance-aware framework that integrates AI into coaching workflows through a structured model combining OSCAR, Knowledge-Skills-Abilities targets, Situational Leadership and key performance indicators, while warning that the evidence remains at a prototype and preliminary validation stage rather than proof of real-world organizational impact.

AI coaching needs structure before scale

Existing work has discussed AI coaching capabilities, adoption factors and ethical principles, but has not fully translated those ideas into a practical organizational workflow. The missing piece, according to the researchers, is a framework that explains how AI should be embedded across coaching phases, how employee competencies should be defined, how progress should be monitored and how governance should operate inside the coaching process.

To address that gap, the authors developed a 10-phase coaching cycle. The process begins with company context, including strategy, mission and constraints. It then clarifies the employee's role and work situation before moving into goal setting, present-state analysis, development-level assessment, option generation, action planning, competency mapping, KPI creation and review.

The design is built around OSCAR, a coaching model based on outcome, situation, choices, actions and review. In this framework, OSCAR provides the process spine. It gives AI a structured sequence for moving the employee from goal definition to action and review.

The Knowledge-Skills-Abilities model adds a competency layer. Instead of leaving coaching goals as broad ambitions, the framework pushes the AI system to define what knowledge must be acquired, what skills must be practiced and what abilities must be strengthened. This step is central to making coaching outputs usable for employee development plans.

Situational Leadership adds adaptation. The AI is designed to classify the employee's readiness level based on competence and commitment, then adjust the amount and style of guidance. A beginner may receive more directive support, while a more experienced employee receives a lighter, more delegative approach.

KPIs add measurement and governance. The framework requires the AI to generate indicators that link the coaching plan to organizational goals, including targets, sources, frequency of review and ownership. The authors stress that these KPI structures should not yet be treated as validated performance measures. They are prototype outputs meant to show how progress could be made reviewable.

The framework also rejects a rigid one-way sequence. Review points can send the employee back to earlier phases when goals change, progress stalls, the situation shifts or the chosen action plan proves unrealistic. That iterative design matters because coaching is rarely linear. It unfolds through reflection, adjustment and renewed commitment.

Human oversight remains the safeguard against AI coaching risks

AI coaching must be governed as a people-development system, not treated as a chatbot deployment. The researchers identify several core risks: hallucinated developmental advice, privacy breaches, biased competency judgments, overdependence on automated guidance and unclear escalation rules.

To control those risks, the framework treats AI outputs as drafts. The AI may ask questions, summarize context, propose development levels, generate options, draft action plans and suggest KPIs. But it does not make final decisions. Employees, managers or organizational approvers retain authority over goals, action commitments, KPI approval and escalation.

The framework includes explicit decision rights across the coaching cycle. The company defines the permissible scope of coaching and governance constraints. The employee confirms role context, chooses options and commits to action. Managers or approvers step in when resources, budget, policy or high-stakes decisions are involved. Human coaches or supervisors intervene when the situation exceeds the appropriate scope of AI support.

The study also turns ethical principles into operational controls. It calls for disclosure that the user is interacting with an AI-supported assistant, data minimization, scope restriction, human validation, audit logs, review checkpoints, bias checks and escalation pathways. Sensitive or out-of-scope cases should be routed to human oversight.

The governance layer is critical because organizational coaching can involve personal ambitions, performance struggles, interpersonal conflict and career decisions. An AI system that gives confident but flawed guidance could distort an employee's self-assessment or push them toward inappropriate development priorities. The framework attempts to prevent that by making every major output reviewable and by requiring human approval before outputs are adopted.

The authors also distinguish adoption evidence from effectiveness evidence. People may find AI coaching useful, convenient or easy to access, but that does not prove it improves behavior, competence, goal attainment or organizational performance. The study emphasizes that AI coaching research remains young and that many existing studies rely on short-term evidence, small samples or self-reported perceptions.

The study does not claim that AI-supported coaching outperforms human coaching. It does not include a control group, long-term employee tracking or real-world deployment across an organization. It evaluates whether the framework can be coherently designed, instantiated and judged plausible by experts.

Prototype tests show promise, but real-world proof is still missing

The researchers tested the framework through a multi-phase design that included a structured literature review, framework construction, large language model prototyping, workflow automation and qualitative evaluation through an online focus group.

The prototyping stage used two fictional cases inside a fictional company. One case involved a senior UX manager trying to balance usability research quality with sprint constraints. The other involved a frontend developer transitioning from React to Angular. The researchers tested the framework with multiple AI tools and examined whether each could produce the required coaching artifacts, including SMART goals, development-level classifications, options, action plans, KSA maps, KPI sets and review plans.

The results showed that the framework could be instantiated across tools and scenarios. The AI systems were able to follow the broad coaching sequence, generate development artifacts and create structured outputs. Different tools showed different tendencies. Some outputs were more technical, others more motivational or action-oriented. But the general workflow remained recognizable.

The researchers also built a proof-of-concept automation using n8n with a language model and memory node to simulate stateful coaching sessions. This demonstrated that the framework could be converted into a traceable workflow with saved context and auditable outputs. However, the study is clear that this was not an enterprise-ready deployment. It did not test scalability, failure recovery, access control, security hardening, authentication, monitoring infrastructure or production-level resilience.

The focus group included six experts from coaching, psychology, human resources and learning-organization practice. Their feedback supported the framework's clarity, flow and usefulness for competence development, while also emphasizing governance needs, manager involvement and clear escalation rules. The feedback informed a refined version of the framework that tightened decision gates, specified minimum required fields for deliverables and clarified oversight roles.

The study's limitations can't be ignored. The evaluation used fictional scenarios rather than real employees and the expert sample was small. The work did not measure employee performance, skill retention, behavioral change, KPI attainment or business outcomes. It did not compare AI-supported coaching with human coaching or ordinary workplace development. The authors treat the results as evidence of prototype feasibility and qualitative refinement, not proof of impact.