Job embeddedness is a construct that reflects the extent to which employees are connected to their jobs and work environments through personal, professional and practical linkages (Goliroshan et al., 2021; Pasaribu et al., 2021; Saeed et al., 2020). Factors such as age, tenure and personal commitments influence embeddedness, with older employees often demonstrating greater stability in their employment decisions (Norouzinik et al., 2022; Setyaningrum et al., 2025; Shah et al., 2020b). The concept is commonly described through three dimensions: fit (compatibility between the individual and the organisation), links (the number and quality of formal and informal relationships within the workplace) and sacrifice (the perceived cost of leaving the organisation) (Fuchs, 2022; Rahimnia et al., 2022; Shah et al., 2020a). Drawing on SET, employees who experience high JE perceive strong organisational support and reciprocal benefits, which foster trust, loyalty and a reduced desire to leave. Conservation of Resources Theory further explains this relationship by positing that employees seek to retain valuable resources such as workplace relationships, accumulated expertise and organisational familiarity that would be lost upon departure. Empirical studies have consistently shown that greater embeddedness significantly reduces TI by strengthening employees’ ties to the organisation (Alharbi, 2020; Setthakorn et al., 2024; Van Zoonen et al., 2021). Accordingly, JE is expected to exert a negative effect on TI (H1).

Job satisfaction, on the other hand, represents a positive emotional state arising from the appraisal of one’s job experiences and environment (Andersson et al., 2021; Naz et al., 2020; Rohaida & Zainal, 2022). It captures the degree to which employees’ expectations, needs and values align with their job roles and organisational context (Barkhuizen & Gumede, 2021; Febrianti et al., 2020; Malik et al., 2020). Job embeddedness can enhance JS by creating a sense of belonging, fostering meaningful workplace relationships, aligning personal values with organisational goals and increasing perceived investments in one’s job (Shah et al., 2020a, 2020b; Wibowo et al., 2022). From the perspective of SET, embeddedness facilitates positive exchanges, such as recognition, career opportunities and social support, which in turn heighten satisfaction. Meanwhile, COR theory suggests that embeddedness preserves valuable psychological and social resources, enabling employees to cope more effectively with workplace challenges and uncertainties. Research has shown that higher JE correlates with greater JS and adaptive responses to organisational change (Costas-Gual et al., 2024; Gašić et al., 2025; Hessari et al., 2024; Rashid et al., 2022; Reh et al., 2021; Yang et al., 2020; Yildiz & Yildiz, 2022). Thus, JE is expected to have a positive relationship with JS (H2).

Turnover intention refers to an employee’s conscious and voluntary decision to leave their current position or move to another organisation (Abudaqa et al., 2021; Alkadash, 2020; Dordunu et al., 2020; Mokoena et al., 2022). High turnover rates can disrupt organisational stability, inflate recruitment and training costs, and diminish overall performance (Adeodu et al., 2021; Aranda-Jiménez et al., 2024). Job satisfaction has been identified as a critical predictor of TI, with numerous studies confirming that satisfied employees are more committed and less inclined to seek alternative employment (James, 2020; Mansoor et al., 2024; Srivastava & Mohaley, 2022; Ullah et al., 2024). Within SET, JS is interpreted as an outcome of equitable and mutually beneficial exchanges, which reinforces loyalty. From a COR standpoint, satisfaction represents a psychological resource that employees are reluctant to relinquish, as leaving could result in the loss of personal, professional and emotional benefits associated with their current role. Consequently, JS is anticipated to have a negative effect on TI (H3).

The interplay between JE, JS and TI is multifaceted and mutually reinforcing. Employees who are strongly embedded in their jobs tend to experience greater satisfaction, which subsequently reduces their intention to leave (Muafi & Kusumawati, 2021; Pašić et al., 2023; Siswanti & Muafi, 2025). Social Exchange Theory explains this dynamic by emphasising that embeddedness strengthens the accumulation of social and professional resources, leading employees to reciprocate with loyalty, while COR theory underscores the protective nature of embeddedness and satisfaction in safeguarding valued resources. Prior research demonstrates that JE can reduce TI both directly and indirectly through JS (Fahmi et al., 2023; Tran et al., 2025; Wiyatno et al., 2024), and in some cases, the direct effect of embeddedness on TI may be stronger than its mediated effect (Muafi & Kusumawati, 2021; Siddiqui et al., 2025; Wigren-Kristoferson et al., 2022). Based on this evidence, JS is posited to mediate the relationship between JE and TI, such that higher embeddedness fosters greater satisfaction, thereby diminishing employees’ intention to leave (H4).

This study employs a quantitative explanatory research design to examine the relationships among JE and TI, with JS as the intervening variable (Azami et al., 2023; Lizarralde-Aiastui et al., 2020; Sousa et al., 2023). The explanatory design is aimed at identifying potential causal linkages among the constructs and drawing inferences grounded in statistical evidence. The quantitative approach facilitates objective measurement of variables through structured data collection tools, such as standardised questionnaires, thereby enhancing the generalisability of the findings.

This study employs a quantitative explanatory research design, which is particularly suitable for examining causal relationships among constructs and providing empirical validation through statistical modelling (Creswell & Creswell, 2018). The design allows for the simultaneous testing of multiple hypotheses while ensuring robustness in the interpretation of structural relationships. Data were collected using a structured questionnaire, with measurement instruments adapted from validated scales in prior studies to ensure conceptual alignment with the study’s objectives.

The target population comprised millennial employees of BPD across Sumatra, Indonesia. A total of 348 respondents were drawn from eight selected BPDs: Bank Aceh, Bank Sumut, Bank Riau Kepri Syariah, Bank Nagari, Bank Bengkulu, Bank Jambi, Bank Sumsel Babel and Bank Lampung. These banks were purposively selected to represent a broad spectrum of geographic, economic and cultural contexts within Sumatra, thereby increasing the representativeness of the findings. The inclusion of banks from both western and eastern parts of the island was intended to capture variations in organisational climate and employee experiences across different socio-cultural settings (Aguilar-Rodríguez et al., 2021; Rožman et al., 2013). Nonetheless, restricting the sample to eight institutions may introduce potential bias if unobserved regional or cultural factors influence the constructs under study. This limitation is acknowledged, and the results should be generalised with caution to other banking contexts in Indonesia.

All constructs were measured using instruments adapted from widely recognised scales, translated into Indonesian and back-translated into English in accordance with Brislin’s (1980) guidelines to ensure linguistic and conceptual equivalence. Cultural adaptation was conducted through expert panel reviews involving academics and HR practitioners, ensuring that the items were contextually relevant to the Indonesian banking sector:

Responses were recorded on a five-point Likert scale ranging from 1 (‘strongly disagree’) to 5 (‘strongly agree’). The psychometric properties of the measures were rigorously evaluated. Discriminant validity was confirmed using both the Fornell–Larcker criterion, where the square root of the Average Variance Extracted (AVE) exceeded inter-construct correlations, with all values below 0.85. Convergent validity was supported by factor loadings above 0.70, while variance inflation factor (VIF) scores below five indicated the absence of multicollinearity.

Data were collected using a mixed-mode approach. Online surveys were distributed via Google Forms, while printed questionnaires were administered in person for participants without reliable internet access. All respondents received a plain-language statement outlining the study’s purpose, assuring anonymity and emphasising voluntary participation. Ethical clearance was obtained prior to data collection, and informed consent was secured from all participants.

To mitigate common method bias (CMB), several procedural remedies were implemented. Firstly, predictor and criterion variables were measured in separate sections, and for a subset of respondents, temporal separation was applied between the measurement of independent and dependent variables. Secondly, anonymity and confidentiality assurances were emphasised to reduce evaluation apprehension. Thirdly, item order was randomised within each construct to minimise response pattern biases. Finally, items were carefully worded to avoid ambiguity or leading language.

Data were analysed using Structural Equation Modelling (SEM) with Partial Least Squares (PLS) implemented in SmartPLS. This method was selected for its ability to model complex relationships, accommodate measurement error and perform well with non-normal data distributions (Hair et al., 2022). Structural Equation Modelling-Partial Least Squares was particularly appropriate given the study’s focus on both direct and indirect relationships, enabling the examination of JS as a mediating variable in the relationship between JE and TI.

This study was conducted in strict adherence to ethical standards as outlined by Universitas Sriwijaya and its institutional review board. Ethical clearance was granted prior to data collection by the Faculty of Economics Review Board. Participants’ privacy was safeguarded throughout the research process, with all responses anonymised before analysis. Informed consent was obtained from all respondents, ensuring they were fully aware of the study’s objectives and that their participation was entirely voluntary.

Respondents involved in this study mostly have an age between 28 years and 30 years, namely 90 employees (25.9%). The next order is respondents aged 31–35 years, as many as 130 employees (37.4%), and 36–44 years, as many as 128 employees (36.8%). Referring to the category of respondents, most of the respondents were female, as many as 188 employees (54%), and male respondents, as many as 160 employees (46%). Educational background Bachelor (S1), as many as 230 people or 66.1% of the total respondents. Respondents with Diploma (D3) education cover 20.1% or 70 people, Postgraduate (S2) reaches 48 people or 13.8%, who generally occupy managerial or strategic roles. Aspects of work units, as many as 57.5% or 200 respondents work in the front office unit; this section includes tellers, customer service representatives who directly interact with customers. The dominant number (57.5%) reflects their important role in bank operations. Meanwhile, 42.5% or 148 respondents work in the back office unit. Employees in this section include operations, finance, technology and administration teams.

The measurement model was assessed through reliability, convergent validity and discriminant validity. Reliability testing indicated that all constructs demonstrated strong internal consistency (Table 1), with Cronbach’s alpha values exceeding 0.90 and composite reliability (CR) values surpassing the recommended threshold of 0.70 (Nunnally & Bernstein, 1994). Specifically, JE (α = 0.957; CR = 0.957), JS (α = 0.985; CR = 0.985) and TI (α = 0.969; CR = 0.970) all exhibited excellent reliability, confirming that the indicators consistently measure their respective latent constructs.

Table 2 presents the results of the factor loadings and convergent validity (AVE) for the three constructs. Overall, all items demonstrated satisfactory outer loadings, with values above the recommended threshold of 0.70, indicating that each indicator adequately reflects its underlying construct. The AVE values for JE (0.742), JS (0.712) and TI (0.803) all exceeded the 0.50 benchmark (Fornell & Larcker, 1981), thereby confirming strong convergent validity across constructs.

Reliability testing further reinforced these findings, as both Cronbach’s alpha and CR values were well above the conventional 0.70 cut-off (Nunnally & Bernstein, 1994). This demonstrates that the indicators within each construct possess high internal consistency and measurement stability. Nonetheless, a closer inspection of the factor loadings reveals that some items of JS (e.g. JS6 = 0.788; JS8 = 0.811) were relatively lower compared to other indicators within the same construct. Although these values remain above the minimum threshold and thus acceptable, they suggest potential redundancy or weaker alignment with the latent variable. Future studies may consider refining or eliminating weaker items to enhance the overall psychometric robustness of the scale.

Taken together, the results confirm that the measurement model exhibits strong reliability and convergent validity while also highlighting areas for refinement in subsequent applications. This provides a solid foundation for advancing to the structural model analysis.

The most important or main evaluation of goodness of fit is the Standardised Root Mean Square Residual (SRMR) index (Table 3). If the SRMR value < 0.08, it means the model fits. The test results show SRMR of 0.018, so that SRMR < 0.08. Another index is the Normed Fit Index (NFI), where if NFI > 0.9 means the model is a good fit. The results of the test show an NFI value of 0.810, which is in the marginal fit category, or still acceptable. Thus, in general, it can be concluded that the model has a good fit, so there is a match between the empirical model and the empirical model.

The structural model results provide compelling evidence for the relationships proposed in this study. The explanatory power of the model, reflected in the R² values, is moderate and meaningful. Job satisfaction (JS) is explained by JE at 33.1%, while TI is explained by both JE and JS at 48.2%. These values suggest that while JE alone contributes significantly to satisfaction, the combination of JE and JS plays an even stronger role in shaping employees’ TIs. This finding aligns with prior PLS-SEM research, where R² values in the moderate range are considered both acceptable and theoretically informative (Hair et al., 2019).

The analysis of effect sizes (f²) provides deeper insight into the relative strength of the structural paths. Job embeddedness exerts a strong influence on JS (f² = 0.495), reaffirming the view that when employees feel embedded in their roles and communities, their overall satisfaction is substantially enhanced. In turn, JS shows a medium-to-large effect on TI (f² = 0.289), underscoring its central role as a mechanism that links JE to employees’ intentions to stay or leave. Interestingly, the direct effect of JE on TI is relatively weak (f² = 0.083), which suggests that embeddedness alone is insufficient to retain employees unless it is accompanied by higher satisfaction. This pattern of results strengthens the argument for a partial mediation, where JS acts as the primary pathway through which JE reduces TI. Theoretically, this adds nuance to the SET and COR by showing that resources embedded in the job context translate into lower turnover only when they are converted into positive psychological states such as satisfaction.

The predictive relevance analysis (Q²) further demonstrates the robustness of the model. Both JS (0.215) and TI (0.312) yielded positive Q² values, confirming that the model is not only explanatory but also predictive. This predictive accuracy is important in organisational contexts, as it indicates that managers can rely on JB and JS as reliable indicators when designing retention strategies. From a practical perspective, these findings highlight the need for organisations, particularly in the non-Western banking context studied here, to focus on initiatives that foster embeddedness while simultaneously nurturing satisfaction. For millennial employees, who often prioritise career development, work-life balance and meaningful engagement, such practices may be especially effective in reducing TI.

Overall, the structural model findings reinforce both theoretical and practical contributions. They demonstrate that JE is a powerful antecedent of satisfaction and that satisfaction is the critical lever for reducing turnover. By situating these insights within SET and COR, the study advances scholarly understanding while also offering actionable guidance for organisations seeking to retain valuable talent in an increasingly competitive environment.

While this study provides valuable theoretical and practical insights, several limitations must be acknowledged. Firstly, the cross-sectional design restricts the ability to establish causality between the examined constructs. Future studies employing longitudinal or experimental approaches would allow for stronger inferences about causal pathways and the temporal stability of these relationships. Secondly, the cultural and contextual focus on millennial employees in Regional Development Banks across Sumatra limits the generalisability of the findings. Cultural norms, economic structures and labour market dynamics may shape how JE and satisfaction interact with TI. Replication in different sectors, regions and countries is therefore necessary to assess the universality of the proposed model. Thirdly, while JS was modelled as a mediator, other organisational and psychological factors such as organisational culture, leadership style, perceived organisational support or employee engagement may also play mediating or moderating roles. Incorporating such variables in future research could provide a more comprehensive understanding of the mechanisms underlying TI. Finally, the generational scope of this study was limited to millennials; the findings may not generalise to Gen Z or older cohorts. Addressing these limitations will help refine the JE framework and extend its applicability across diverse organisational, generational and cultural contexts.