This study adopted a design science research (DSR) approach to develop and validate an AI-driven career guidance system tailored for vocational students. DSR is suitable for building innovative digital artifacts while addressing real-world educational challenges, particularly in bridging skill gaps and reducing decision-related anxiety ( Chavva, 2025; Holmes et al., 2019). The research followed iterative cycles of problem identification, artifact development, and evaluation. The primary focus was to integrate skills mapping, adaptive mentoring, and real-time labor market analytics into a unified guidance platform.

Participants were vocational high school students from three institutions representing diverse socioeconomic and geographical backgrounds. A purposive sampling technique was employed to select students in their final two years of study, as this group experiences the most pressure when making career decisions. A total of 180 students participated, divided into an intervention group (who used the AI-driven platform) and a control group (who received conventional counseling). Additionally, 15 career counselors and 12 industry mentors were engaged to validate system recommendations and provide qualitative feedback. This sampling aligns with prior research emphasizing the importance of multi-stakeholder involvement in career guidance interventions ( Faruque et al., 2025; Charanya et al., 2025).

The system was designed as a mobile-first platform with three integrated modules: •

Skills Mapping Engine – Leveraging supervised machine learning to analyze academic performance, extracurricular activities, and certifications, and then mapping them against in-demand skills. This module adapts predictive models from AI-driven career mapping research ( Selvaraj et al., 2025).

Additional features included an AI-powered resume builder to enhance employability ( Jha et al., 2025) and an explainable AI interface to improve trust and reduce AI-related anxiety among students ( Faruque et al., 2025; Dhilipan et al., 2025).

Figure 2 presents the core architecture of the AI-driven career guidance system. It demonstrates how multiple modules—data collection, processing, recommendation, and feedback—interact dynamically to generate personalized and secure career recommendations for vocational students. In the system’s implementation, data collection incorporated academic records, vocational certifications, self-assessment results, and internship experiences, representing both cognitive and affective student dimensions. These datasets were preprocessed using normalization and embedding methods to ensure data fidelity and cross-compatibility across institutions ( Faruque et al., 2025). The data processing layer employed a multilingual BERT model for text normalization and contextual encoding, enhancing semantic understanding of skill descriptions and labor market requirements ( Selvaraj et al., 2025). The recommendation engine formed the analytical core of the system. It combined a Random Forest Classifier for supervised prediction of optimal career pathways and a collaborative filtering mechanism to tailor suggestions based on students’ previous interactions and preference patterns ( Kurian et al., 2025; Jha et al., 2025). These algorithms integrated both cognitive–technical features (such as GPA and certifications) and affective–psychological factors (such as anxiety and persistence indices), thereby bridging AI analytics with vocational psychology ( Lee et al., 2022; Sunday & Sefotho, 2025). Through the feedback loop, students’ ongoing activity data—such as repeated skill-gap analyses, mentoring interactions, and updated resumes—were continuously fed back into the model for re-training and adaptive recalibration. This iterative structure aligns with the dynamic feedback-driven learning optimization framework proposed by Song, Shin, and Shin (2024), ensuring that recommendations evolve alongside users’ progress. A user-friendly interface enabled students to engage with AI-based mentoring dialogues, view real-time labor demand analytics, and access skill-gap visualizations. The integration layer connected all modules (skills mapping, mentoring, and labor market analytics) under one cohesive ecosystem, allowing counselors and mentors to monitor progress in real time ( Chavva, 2025). To uphold ethical and trustworthy AI principles, the architecture embedded strong data privacy mechanisms—such as anonymization, encrypted storage, and transparent explainability layers. The system utilized SHAP (SHapley Additive Explanations) values to display the top contributing features influencing each recommendation, increasing transparency and reducing AI-related anxiety ( Dhilipan et al., 2025; Holmes et al., 2019). Collectively, this architecture operationalized the integration of skills mapping, adaptive mentoring, and real-time labor market intelligence into a unified vocational guidance framework. The system achieved high predictive accuracy (87%) while reducing students’ career path anxiety by 26.7%, confirming that the interactive structure illustrated in Figure 2 effectively supports both cognitive readiness and emotional resilience in vocational career decision-making ( Faruque et al., 2025; Kurian et al., 2025; Ye et al., 2024).

Data collection was carried out in three stages: •

Pre-intervention stage: Students completed a career path anxiety scale and submitted their academic and extracurricular records.

This procedure aligns with best practices in AI-based educational research where both behavioral data and affective outcomes are captured ( Swarnim et al., 2025; Rane et al., 2024).

The analysis combined quantitative and qualitative approaches: •

Quantitative data were analyzed using paired-sample t-tests and ANOVA to examine changes in career path anxiety between groups. Predictive performance of the AI model was measured using precision, recall, and F1-scores, consistent with AI-based career guidance studies ( Selvaraj et al., 2025).

We would like to clarify that although the authors’ primary affiliation is Universitas Negeri Yogyakarta, the research was conducted entirely in South Kalimantan Province (in three vocational high schools). In accordance with local administrative and regulatory requirements, research permits and ethical clearance were therefore obtained from the nearest authorized public university, Universitas Lambung Mangkurat. This practice aligns with regional research regulations that require formal approval from a locally recognized institution. This clarification has now been explicitly stated in the manuscript .

All procedures followed comply with ethical standards in research involving human participants and were approved by the Institute for Research and Community Service (Lembaga Penelitian dan Pengabdian kepada Masyarakat), Universitas Lambung Mangkurat (Approval No. 124/UN8.2/PG/2025). Written consent was obtained from students, parents (for underage students), and participating schools, namely three vocational schools in South Kalimantan: SMK Isfi Banjarmasin, SMK Telkom Banjarbaru, and SMK NU Banjarmasin, as stated in permit number 424/3357/SMK.NU/Disdikbud/2025. Data privacy and security were ensured through anonymization and secure storage of student records, in accordance with recommendations on the ethical adoption of artificial intelligence in education ( Rane et al., 2024; Holmes et al., 2019). The system was designed with transparency mechanisms, including explainable AI outputs, to address potential ethical concerns and enhance user trust ( Dhilipan et al., 2025). Participation was entirely voluntary, and students retained the right to withdraw from the study at any time without any academic consequences.

The model development process constituted the backbone of the AI-driven career guidance system, designed to translate heterogeneous data sources into actionable and personalized career recommendations. This stage was not merely technical but also conceptual, bridging vocational education psychology with real-time labor market intelligence. To achieve this, the development followed five major sub-stages: data preprocessing, feature engineering, model architecture design, training and validation, and explainability integration. a.

Data Preprocessing

The input data came from three principal sources: •

Student Data – including academic performance records, vocational certifications, internship experience, and results from self-assessment surveys measuring career interests and anxiety levels.

Preprocessing included: •

Data Cleaning – resolving missing data through imputation (mean for numerical, mode for categorical), and removing inconsistent entries.

This ensured the dataset was both high fidelity and cross-compatible across structured and unstructured formats ( Faruque et al., 2025). b.

Feature Engineering

The system’s predictive strength depended on strategic feature construction. Features were categorized into: •

Cognitive-Technical Features: GPA, vocational certifications, internship duration, domain-specific assessments.

Dimensionality reduction was conducted via Principal Component Analysis (PCA) and t-SNE visualization, which reduced noise and enhanced clustering interpretability. Importantly, synthetic features were engineered—for example, a “career readiness index” combining academic, skill, and psychological factors into a composite score. c.

Model Architecture Design

The architecture adopted a hybrid model to balance precision, adaptability, and interpretability: •

Unsupervised Layer: K-Means clustering grouped students with similar skill-anxiety profiles, enabling peer benchmarking.

This layered design allowed the system to not only recommend careers but also explain why a given pathway matched the student’s current skills and psychological state. d.

Training and Validation

Training used a 10-fold cross-validation strategy with stratified sampling to ensure representation across anxiety levels and skill profiles. Evaluation metrics included accuracy, precision, recall, F1-score, and ROC-AUC. •

The model achieved Accuracy = 0.87, Precision = 0.85, Recall = 0.83, and F1-score = 0.84.

Additionally, the NLP module was tested using BLEU and ROUGE scores against gold-standard counselor-student dialogues. The chatbot achieved a BLEU score of 0.71, indicating semantically faithful guidance. e.

Explainability and Trust Integration

Since vocational students and educators may distrust “black-box AI,” an explainable AI (XAI) layer was incorporated. Using SHAP (SHapley Additive exPlanations) values, the system could highlight the top three factors influencing each recommendation—for example, “Recommended career: Network Technician (based on: Cisco Certification, low anxiety score, regional job growth index).” This aligns with dynamic assessment principles ( Jeltova et al., 2007) and Vygotsky’s zone of proximal development ( Vygotsky, 1978), where feedback is both instructional and developmental. The explainability mechanism ensured that students perceived recommendations as personalized, transparent, and pedagogically meaningful.

Figure 3 visualizes how various AI-driven components such as data-driven insights, real-time feedback, and continuous learning converge to produce a personalized career assessment. This integration ensures that recommendations are both technically valid and psychologically responsive to student needs. In practical implementation, the data-driven insights were generated by aggregating three layers of information: (1) students’ academic and certification records representing cognitive competence, (2) self-assessment and anxiety indices representing psychological readiness, and (3) real-time labor market analytics capturing industry demand patterns. These data streams were processed through supervised and unsupervised learning models to identify both the current skill position of each student and the optimal career trajectory most aligned with their competencies and interests ( Selvaraj et al., 2025; Kurian et al., 2025). The algorithmic layer used Random Forest classification to predict career suitability while K-Means clustering grouped students with similar skill–anxiety profiles, facilitating peer-based benchmarking and adaptive goal setting. The real-time feedback mechanism played a central role in translating algorithmic outputs into actionable guidance. As students interacted with the platform—through chatbot mentoring, skill-gap analysis, and labor market exploration—their new data were continuously looped back into the model. This enabled the system to re-evaluate previous recommendations and dynamically adjust the career pathways, consistent with the feedback-driven optimization framework proposed by Song, Shin, and Shin (2024). For example, if a student’s anxiety index decreased after multiple mentoring sessions, the model recalibrated their predicted confidence level and expanded the range of recommended job roles or internships. The continuous learning component relied on reinforcement through both user behavior and mentor validation. Each confirmed recommendation (e.g., when a student accepted a career suggestion or completed a targeted skill course) functioned as positive feedback, while ignored or rejected suggestions triggered model recalibration. This adaptive loop allowed the system to evolve contextually, reflecting the individual progress of each student and collective behavioral trends across the cohort ( Faruque et al., 2025; Ojeda-Ramirez et al., 2023). Psychologically, this design was grounded in Social Cognitive Career Theory (SCCT), where students’ self-efficacy and outcome expectations influence their decision-making confidence ( Lent, Brown, & Hackett, 2002). By combining AI analytics with adaptive mentoring, the system supported students’ zone of proximal development ( Vygotsky, 1978), helping them transition from anxiety-driven indecision to confident, self-regulated career planning. Implementatively, Figure 3 represents this holistic cycle: data insights inform recommendations, feedback refines decisions, and continuous learning personalizes guidance. The result is not merely a static report but a living career assessment system one that evolves with the student’s development and mirrors real-time labor market shifts. This integrative feedback architecture directly contributed to the observed 26.7% reduction in career path anxiety and 87% model accuracy in matching students to suitable career pathways, validating both the technical soundness and psychological relevance of the proposed AI framework ( Ye et al., 2024; Sunday & Sefotho, 2025; Chavva, 2025). f.

Iterative Refinement

Finally, a human-in-the-loop cycle was employed: teachers, career counselors, and students evaluated early outputs and fed corrective input back into the model. This participatory refinement ensured cultural and contextual alignment with the needs of Indonesian vocational schools ( Indrawati & Kuncoro, 2021). This study produced several significant findings that demonstrate the effectiveness of the AI-driven career guidance system in reducing vocational students’ career path anxiety, enhancing career clarity, and strengthening their metacognitive engagement with future planning. This ultimately resulted in an accessible system as shown in Figure 4 below.

Figure 4 displays the main interface of the AI-Guidance web system designed to support vocational students in exploring suitable career pathways. The homepage introduces users to the core function of the system through options such as Start Assessment and Ask AI Assistant, enabling personalized career guidance. It also presents key features including Career Assessment, Psychometric Test, AI Recommendations, and Online Mentoring, which work together to provide comprehensive and AI-based career support. The interface is designed to be simple, user-friendly, and accessible for students.

Figure 5 illustrates the end-to-end process of the AI-based career guidance system, starting from the assessment, proceeding to career recommendation acceptance, and finally generating a personalized career roadmap. This process is built upon Holland’s RIASEC theory, which classifies individuals based on six personality types—Realistic, Investigative, Artistic, Social, Enterprising, and Conventional—to help align personal traits with suitable career environments ( Holland, 1997). In the first stage, the system analyzes the user’s RIASEC profile generated from the assessment, which identifies dominant personality types and interest areas. Based on these results, the system recommends specific careers that match the user’s profile. For example, a high score in the Social and Investigative categories aligns with careers like nursing or healthcare services, as seen in the recommendation section of the interface. This is consistent with Holland’s assumption that individuals experience greater satisfaction and success when working in environments that match their personality types. In the second stage, users receive AI-driven career recommendations that include compatibility scores, required skills, potential salary ranges, and relevant industry sectors. This step not only enhances self-awareness but also supports informed decision-making by connecting personal interests to labor market demands. Such a data-driven recommendation approach is aligned with personalized career counseling models that integrate psychological theory and technology ( Nauta, 2010).

In the final stage, the system produces a career roadmap that outlines progressive steps for vocational students, beginning with graduation preparation, moving through workforce transition, and advancing to early career development. This roadmap integrates skill development pathways, including internships, certifications, training programs, and mentorship, which support long-term career growth. This staged guidance approach is aligned with developmental career theories emphasizing continuous career planning and skill acquisition ( Super, 1990).

Table 1. Career Anxiety Pre- and Post-Test presents the differences in career anxiety scores before and after the intervention. In the intervention group, mean anxiety scores dropped significantly from 73.1 to 53.6, while the control group experienced only a minor decrease from 72.4 to 70.2. A paired-sample t-test confirmed that the reduction in the intervention group was statistically significant (p < 0.001). These findings support previous research showing that AI-based interventions can reduce uncertainty and anxiety in decision-making contexts ( Kurian et al., 2025; Selvaraj et al., 2025).

Figure 6 illustrates the operational workflow of the AI Core within the proposed career guidance system, highlighting how its interconnected components — Personalized Career Advice, Real-Time Feedback, Skill Gap Identification, and Market Trend Analysis — collectively contribute to measurable student outcomes such as Enhanced Career Path , Improved Skills , and Informed Decisions. In practice, the AI Core functioned as the central computational engine integrating data inputs from students’ academic records, psychological assessments, and labor market analytics. This core module synthesized multiple data streams using supervised and unsupervised learning algorithms (Random Forest and K-Means) to generate Personalized Career Advice tailored to each student’s competency profile and anxiety level ( Selvaraj et al., 2025; Faruque et al., 2025). Each recommendation was accompanied by explainable AI outputs through SHAP-based reasoning, allowing students and counselors to understand why a certain pathway was suggested ( Dhilipan et al., 2025). The Real-Time Feedback mechanism allowed continuous interaction between the student and the system. As learners updated their achievements, completed skill modules, or participated in mentoring sessions, the AI Core recalibrated its recommendations to reflect progress and changing psychological states. This iterative loop mirrors the feedback-driven learning model described by Song, Shin, and Shin (2024), ensuring that guidance remains current, adaptive, and reflective of students’ evolving readiness. The Skill Gap Identification module compared students’ competencies against aggregated labor market requirements derived from the Real-Time Labor Market Intelligence database. This enabled the AI system to pinpoint missing or underdeveloped skills, recommending specific courses, certifications, or microlearning opportunities. Over 65% of students in the experimental group repeated this skill-gap analysis at least three times, showing that it encouraged metacognitive engagement and self-directed improvement. As a result, students demonstrated measurable skill enhancement and greater confidence in aligning their competencies with industry expectations ( Ye et al., 2024; Kurian et al., 2025). Simultaneously, the Market Trend Analysis component employed natural language processing to analyze thousands of job postings and sector reports. This real-time data informed the system’s recommendations, ensuring that students’ choices aligned with emerging career opportunities rather than outdated occupational data. By continuously updating skill–demand mappings, the AI Core empowered students to make Informed Career Decisions, reducing uncertainty and anxiety associated with labor market unpredictability ( Budhwar et al., 2023; Chavva, 2025).

Figure 7 illustrates the AI Career Assistant feature within the AI-based career guidance system. This interface provides an interactive virtual assistant designed to support students in exploring and planning their career pathways. The AI assistant integrates personalized guidance by offering deep career analysis, real-time labor market insights, skill development recommendations, and career learning roadmaps. Through the chat-based interface displayed on the right side of the screen, users can directly ask career-related questions, such as suitable job options, required skills, internship opportunities, or CV and interview tips. The assistant is powered by advanced AI technology, enabling it to deliver adaptive and data-driven responses tailored to each user’s career profile and interests. This feature enhances accessibility to individualized career counseling, especially for vocational students who may have limited access to professional human counselors. By providing continuous, on-demand guidance, the AI Career Assistant serves as a practical digital mentoring tool that supports self-directed career development in a user-friendly manner.

The system’s AI model was evaluated using supervised learning with labeled career trajectories and skill datasets. As shown in Table 2. AI Model Performance, the model achieved strong performance with an accuracy of 87% and an F1-score of 0.84. This level of performance is comparable with advanced AI-based career recommender systems in other contexts ( Selvaraj et al., 2025), highlighting the robustness of the model.

Engagement data indicated that students interacted with the AI platform beyond initial use. As shown in Table 3, Student Activity on the AI Platform, 65% of students conducted the skill-gap analysis at least three times, suggesting sustained metacognitive engagement with the feedback provided by the system. In addition, 79% of participants engaged in adaptive mentoring sessions more than twice, reflecting active utilization of personalized guidance features. A high proportion of students (87%) downloaded career roadmap reports, indicating strong interest in structured career planning support, while 54% used the AI-based resume builder to enhance employability preparation. These findings suggest that students were not passive recipients of system recommendations but demonstrated self-regulated learning behaviors, consistent with prior research highlighting the role of adaptive digital guidance in fostering learner agency ( Ojeda-Ramirez et al., 2023).

Survey results revealed high levels of user satisfaction with the AI-driven career guidance system across all evaluated aspects. As presented in Table 4. Student Satisfaction Scores, adaptive mentoring support received the highest mean score (M = 4.60, SD = 0.39), indicating strong perceived value of personalized guidance. The relevance of career recommendations was also rated highly (M = 4.52, SD = 0.41), suggesting that students considered the system’s outputs well aligned with their skills and career goals. Ease of use (M = 4.48, SD = 0.47) and access to real-time labor market information (M = 4.44, SD = 0.50) further reflect positive user experiences, while transparency of AI decision-making achieved a strong mean score (M = 4.36, SD = 0.55). Overall, these findings indicate that students perceived the system as both useful and trustworthy. The results are consistent with prior research on explainable artificial intelligence (XAI), which highlights that transparent AI-driven recommendations enhance user trust, acceptance, and sustained system adoption ( Faruque et al., 2025).

Semi-structured interviews with 20 students and 5 guidance teachers revealed three overarching themes that reinforce the quantitative findings.

As summarized in Table 5, Key Themes from Qualitative Interviews, the first theme, empowerment through personalization, reflects students’ increased confidence resulting from career recommendations tailored to their individual skills and profiles. The second theme, authentic assessment, highlights how the integration of real-time labor market data enhanced the perceived relevance and tangibility of career guidance, enabling students to better understand current industry demands. The third theme, psychological readiness, captures students’ experiences of reduced anxiety and improved emotional preparedness for making future career decisions. Collectively, these qualitative insights suggest that the AI-driven system supported both cognitive and affective dimensions of career development. These findings are consistent with the principles of authentic assessment, which emphasize real-world relevance and learner agency ( Gulikers et al., 2004), and further support the role of contextualized technical and vocational education and training (TVET) in promoting sustainable career development ( Ye et al., 2024). Taken together, the results demonstrate that the AI-driven career guidance system not only achieved high technical accuracy in career mapping but also had meaningful psychological and educational impacts. The intervention group experienced a 26.7% reduction in career path anxiety, significant engagement with metacognitive tools such as skill-gap analysis, and strong satisfaction with adaptive mentoring features. Qualitative data provided further evidence that students felt empowered, reassured, and better aligned with real-world labor market conditions. These combined quantitative and qualitative findings emphasize that integrating skills mapping, adaptive mentoring, and real-time labor market intelligence in vocational education can simultaneously improve technical accuracy and psychological readiness—a dual outcome rarely captured in existing studies.

The outcomes of this study carry significant implications for multiple stakeholders in vocational education and workforce development. a.

For Vocational Counselors and Teachers

The AI-driven career guidance system can serve as a pedagogical support tool, reducing the workload of school counselors who often handle large caseloads with limited resources. By providing automated yet adaptive recommendations, teachers can focus more on mentoring and socioemotional support, rather than administrative career matching tasks. The explainability features also allow educators to understand the rationale behind system suggestions, making it easier to integrate into counseling practices.

While the study provides valuable insights into the development and implementation of an AI-driven career guidance system, several limitations should be acknowledged. a.

Sample and Generalizability

The study involved 220 vocational students from three institutions, which, although sufficient for exploratory validation, may not fully represent the diversity of vocational learners across regions or educational contexts. Future studies should include larger and more heterogeneous samples, incorporating students from different geographic, cultural, and socioeconomic backgrounds to increase external validity. b.

Short-Term Evaluation

The system was evaluated over a three-month period, primarily focusing on immediate impacts such as reductions in career path anxiety and improvements in self-efficacy. However, the long-term effects on employment outcomes, career persistence, and adaptability remain unexamined. Longitudinal studies are necessary to determine the sustained impact of AI-guided career pathways. c.

AI Model Transparency and Bias

Although explainable AI methods were integrated, the system still carries the risk of algorithmic bias, especially when trained on historical labor market data that may reflect existing inequalities ( Williamson et al., 2020). Future work should explore bias mitigation strategies and the co-design of algorithms with stakeholders to ensure fairness and inclusivity. d.

Integration with Human Mentorship

The current system relies on adaptive digital mentoring but does not fully explore synergies between AI recommendations and human mentorship practices. Future iterations should study hybrid approaches that combine the efficiency of AI with the empathy and contextual sensitivity of human counselors. e.

Technical and Infrastructural Constraints

In under-resourced schools, especially in remote or 3T (terdepan, terluar, tertinggal) regions, limited internet connectivity and digital literacy may affect system adoption. Future work should investigate offline-compatible, mobile-first solutions and capacity-building programs for both students and educators. f.

Scalability and Policy Integration

While the system demonstrates promising potential for institutional use, its scalability at the national level has not yet been tested. Future research should explore integration with government education databases, labor market information systems, and TVET policy frameworks ( Ye et al., 2024; Chavva, 2025). g.

Future Work Directions

Building on these limitations, several directions for future research and development are proposed: •

Conduct longitudinal tracking of student cohorts to evaluate the impact on career stability and labor market integration.