1. Introduction
Artificial Intelligence (AI) is emerging as a pivotal technology with the potential to revolutionize productivity and reshape the employment landscape across various economic sectors. Since the commercial release of advanced models like ChatGPT in late 2022, there has been heightened anticipation of a transformative shift comparable to the advent of the internet. However, this optimism is tempered by the reality of persistently low productivity growth in many advanced economies, which raises critical questions about the actual economic impact of AI and the mechanisms through which it influences productivity and employment [
1].
AI’s economic impact is multifaceted and complex. Theoretically, AI can be seen both as a complement and a substitute for human labor. When viewed as a complement, AI enhances human productivity by augmenting decision-making processes and operational efficiency [
2]. Conversely, when viewed as a substitute, AI can automate tasks traditionally performed by humans, potentially leading to job displacement but also contributing to productivity gains [
3,
4]. Empirical evidence supports both perspectives, with studies indicating that AI-using firms often experience positive productivity effects without necessarily observing significant negative impacts on overall employment [
5].
The challenge of measuring AI’s economic impact is compounded by its intangible nature and rapid technological evolution. Traditional economic metrics and national accounting frameworks struggle to capture the full value generated by AI, often leading to an underestimation of its contributions [
6]. This measurement difficulty contributes to what is known as the productivity paradox—where the rapid advancements in technology do not immediately translate into observable productivity gains [
7]. Recent advancements in AI have explored the integration of machine learning models to better understand user sentiment and predict outcomes in various domains [
8].
To tackle these challenges, this study aims to address these challenges by leveraging advanced machine learning models and Bayesian Network Analysis to analyze firm-level data and provide a more nuanced understanding of AI’s impact on productivity and employment. Specifically, the research employs Random Forest and Gradient Boosting Machine (GBM) models to identify key predictors of productivity changes and assess their relative importance. By designing the questionnaire to gather granular data on AI tools usage, integration, and organizational factors, we address the issue of AI’s intangibility by converting it into measurable constructs, such as AI_Integration_Level and AI_Tools_Complexity”. Additionally, Bayesian Network Analysis is used to explore the probabilistic dependencies between various features, offering a comprehensive view of the dynamics at play. The Bayesian Network was constructed using the HillClimbSearch algorithm with Bayesian Information Criterion (BIC) to determine the optimal structure, balancing model complexity, and goodness-of-fit. Parameters were estimated via Maximum Likelihood Estimation (MLE), ensuring accurate representation of conditional probability distributions. Inference was conducted using Variable Elimination, a method suitable for capturing the intricate dependencies in the data. The performance of the models was rigorously evaluated using 5-fold cross-validation to mitigate overfitting and ensure robust performance metrics.
The research hypotheses are grounded in existing literature and empirical findings. First, it is hypothesized that AI integration significantly enhances productivity at the firm level, with AI integration being the most critical predictor of productivity change (Hypothesis 1). This is supported by studies showing substantial productivity gains in firms adopting AI technologies [
9]. Second, it is hypothesized that the impact of AI on employment is complex, with positive productivity effects at the firm level not necessarily translating into negative employment effects at the aggregate level (Hypothesis 2). This hypothesis is informed by research indicating that AI can complement human labor, leading to new job creation and task augmentation rather than straightforward job displacement [
3,
4]. Third, it is hypothesized that the benefits of AI integration are moderated by factors such as AI complexity, areas of AI utilization, and employee characteristics (Hypothesis 3). This is supported by evidence that the impact of AI varies significantly across different contexts and applications [
1,
5].
By employing sophisticated analytical techniques and building on a robust theoretical foundation, this study seeks to contribute to the ongoing discourse on AI’s economic impact. The findings will provide valuable insights for policymakers and business leaders, helping them to harness AI’s potential for economic growth while mitigating potential adverse effects on employment. Through detailed feature importance analysis and the exploration of probabilistic dependencies, this research aims to offer a comprehensive understanding of AI’s role in shaping the future of work and productivity.
The key contributions of this paper are multifaceted. First, it utilizes a unique dataset derived from a survey of 233 employees across various industries, providing valuable empirical insights into the ways AI tools impact productivity. The study goes beyond simplistic measures by considering the complexity of AI integration and its interaction with demographic factors such as age. Second, the research applies a diverse range of analytical techniques, including logistic regression with interaction terms, Random Forest, XGBoost, and Bayesian Network Analysis. This multi-method approach allows for a more nuanced and robust exploration of how AI influences employee productivity.
Third, the findings highlight the critical role of AI integration into organizational workflows, showing that merely adopting AI tools is insufficient without strategic and comprehensive integration. Moreover, this study uncovers generational differences in adaptability to AI tools, with younger employees experiencing greater productivity gains compared to their older counterparts. These generational insights suggest that adaptability to AI technologies may vary significantly across age groups. Finally, the paper offers practical recommendations for policymakers and business leaders, advocating for targeted training programs and the establishment of ethical frameworks to maximize AI’s economic potential.
The remainder of this paper is structured as follows:
Section 2 reviews the existing literature on AI’s impact on productivity, focusing on both theoretical perspectives and empirical studies.
Section 3 outlines the research methodology, including data collection, preprocessing, and the machine learning models employed in the analysis.
Section 4 presents the findings, highlighting the results of the logistic regression, Random Forest, XGBoost, and Bayesian Network Analysis.
Section 5 discusses the broader implications of these findings, particularly concerning AI’s potential to enhance productivity and its differential effects across employee demographics. Finally,
Section 6 concludes with recommendations for future research and strategies to optimize AI’s role in the workplace.
4. Results
To understand the demographics, AI usage, and organizational impacts among respondents, a detailed questionnaire was administered. The detailed distribution of responses for each question (
Table 1) helps in understanding the demographic and professional background of the respondents, as well as their experiences and perceptions related to AI implementation.
Table 3 provides descriptive statistics for the numerical variables, offering insights into the central tendencies and variability within the dataset. The variable AI_Integration_Level refers to the extent to which AI tools are embedded into the organizational workflows. The variable AI_Tools_Complexity refers to the sophistication and functionality of the AI tools employed within the organization. The variables were measured on a continuous numerical scale, with higher values representing a more comprehensive and advanced integration of AI within the organization. These numerical summaries help contextualize the findings and validate the robustness of the subsequent analyses.
4.1. Logistic Regression Model
To focus on the predictors of significant productivity changes, the target variable, ‘Productivity_Change_Percentage’, was re-encoded into a binary outcome, ‘Productivity_Change_Binary’. This binary outcome was defined as 1 for notable productivity change (≥40%) and 0 for lesser changes (<40%). This threshold was chosen to distinguish between minor and substantial productivity improvements, thus providing a clearer understanding of the factors contributing to significant productivity enhancements.
Feature selection was performed using LassoCV, a regularization technique, to identify significant predictors. LassoCV was chosen for its ability to handle multicollinearity and select the most relevant features, thereby improving the model’s performance. Interaction terms were included to capture the combined effect of AI tools usage with integration levels and tool complexity. The LassoCV feature selection identified the following key predictors: Age, Innovation and Competitiveness Improvement, AI Tools Usage * AI Integration Level, and AI Tools Usage * AI Tools Complexity.
A logistic regression model was fit using the selected features (
Table 4). The model’s performance was evaluated using classification metrics, including precision, recall, F1-score, and the ROC AUC score.
We also included interaction terms and fit the second logistic regression model. The interaction terms between AI Tools Usage and AI Integration Level and AI Tools Usage and AI Tools Complexity were included to explore potential multiplicative effects. The rationale was that the productivity impact of AI tools might not only depend on their usage or complexity alone but also on how these tools are integrated within the organization. High levels of integration can enhance the utility and effectiveness of AI tools, amplifying their impact on productivity.
Table 5 presents the logistic regression results with interaction terms, including the odds ratios for the significant predictors. The model showed a good fit with a pseudo-R-squared value of 0.2687 and a log-likelihood of −99.453. Notably, the interaction term between AI Tools Usage and AI Integration Level demonstrated a significant positive association (β = 0.4319,
p < 0.001), indicating that increased usage and higher integration levels collectively enhance productivity.
Main findings
Age: A negative coefficient (β = −0.4520, p < 0.001) suggests that older age groups are associated with lower productivity changes.
AI Tools Usage * AI Integration Level: This interaction term had a positive coefficient (β = 0.4319, p < 0.001), indicating that the combined effect of frequent AI tool usage and high integration levels significantly increases the likelihood of productivity improvement.
AI Tools Usage * AI Tools Complexity: Although this interaction term was positive (β = 0.0840), it was not statistically significant (p = 0.264).
The inclusion of interaction terms revealed important insights into how AI tools usage, when combined with high integration levels, can substantially enhance productivity. This underscores the importance of not only adopting AI tools but also ensuring their comprehensive integration within organizational workflows. The model’s (with interaction terms) overall accuracy was 80%, with a macro average F1-score of 0.73. The ROC AUC score of 0.837 indicates a strong discriminative ability of the model. However, the relatively lower recall for the positive class (0.52) suggests that further refinement is needed to improve the model’s sensitivity.
The model achieved an overall accuracy of 80%, with a ROC AUC score of 0.837, indicating good discriminative ability.
The findings suggest that age and the interaction between AI Tools Usage and AI Integration Level are significant predictors of notable productivity changes. Specifically, younger employees and those working in environments where AI tools are heavily used and well-integrated are more likely to experience significant productivity gains. These insights highlight the importance of targeted training and integration strategies to maximize the benefits of AI adoption.
Further, LassoCV and RidgeCV were employed to handle multicollinearity and select the most relevant features. The selected features by Lasso included Age, Innovation and Competitiveness Improvement, Communication and Collaboration Changes, AI Tools Usage * AI Integration Level, AI Tools Usage * AI Tools Complexity, and AI Tools Usage Squared. Ridge selected a more comprehensive set of features, including various demographic and organizational attributes. The final combined set of features from both Lasso and Ridge included Age, Innovation and Competitiveness Improvement, AI Tools Usage * AI Integration Level, AI Tools Usage * AI Tools Complexity, AI Tools Usage Squared, and several additional features from the Ridge selection.
The logistic regression model with interaction terms and polynomial features for the numerical variables, validated through 5-fold cross-validation, exhibits satisfactory performance with a 0.7512 ± 0.0369 accuracy and a 0.7692 ± 0.0409 ROC AUC score. The model is proficient in distinguishing between the two classes, though enhancements in predicting notable productivity changes are needed.
The complexity and variety of the dataset variables justify the use of advanced ensemble methods like Random Forest and XGBoost. Our dataset includes diverse features such as ‘Age’, ‘Gender’, ‘Education’, ‘AI_Tools_Usage’, ‘AI_Integration_Level’, two interaction terms, and two polynomial features. While logistic regression is useful for identifying key predictors and understanding direct relationships, it has limitations in capturing complex non-linear interactions and dependencies between variables.
The logistic regression analysis highlighted significant predictors like ‘Age’ and the interaction between ‘AI Tools Usage’ and ‘AI Integration Level’, but its linear nature restricts its ability to uncover more intricate patterns and relationships. While innovation remains a theoretically important factor in productivity gains, the specific dynamics captured in this study may be more closely tied to the direct influence of AI tools usage and integration.
To further improve predictive performance, we implemented Random Forests and XGBoost models, evaluated through cross-validation. Random Forests and other tree-based methods, like XGBoost, inherently capture interactions between features due to their hierarchical nature. This means they can handle interactions without explicitly requiring the interaction terms to be manually created. The interpretability techniques SHAP and LIME were employed to gain deeper insights into model predictions, providing transparency and understanding of feature contributions.
4.2. Random Forest and XGBoost
Random Forest and XGBoost are powerful ensemble learning methods that offer several advantages for our analysis:
Handling Non-Linearity and Interactions: Both models can naturally capture non-linear relationships and interactions between variables without the need for explicit feature engineering. This is important given the interaction terms and polynomial features in our dataset, such as ‘AI_Tools_Usage * AI_Integration_Level’ and ‘AI_Tools_Usage_Squared’.
Feature Importance: Random Forest and XGBoost provide insights into feature importance, helping to identify which variables and interactions have the most significant impact on productivity changes. This aligns with our goal of understanding the key factors driving productivity.
The Random Forest model was configured with 200 estimators and a random state of 42. The choice of 200 estimators strikes a balance between computational efficiency and model performance, as increasing the number of trees typically enhances the model’s robustness and generalization capabilities but also raises computational costs. The random state ensures reproducibility of the results. The best parameters that resulted from hypertuning were bootstrap = True, max_depth = None, min_samples_leaf = 2, min_samples_split = 2, and n_estimators = 200. This configuration was evaluated using 5-fold Stratified Cross-Validation (CV) to maintain class balance across folds, providing a reliable performance estimate and minimizing the risk of overfitting by ensuring the model is tested on all subsets of the data.
Similarly, the XGBoost model was configured with 100 estimators to maintain consistency with the Random Forest model and to leverage the strength of ensemble methods in boosting performance through multiple iterations. The use_label_encoder = False parameter was set to bypass the default label encoder in XGBoost, facilitating a direct use of the preprocessed labels and preventing potential encoding issues. The eval_metric = ‘logloss’ was chosen to align the evaluation with logistic regression settings, as log-loss provides a robust metric for binary classification problems by penalizing false classifications proportionally to their confidence. The best parameters resulted from hypertuning were ‘colsample_bytree’: 0.6, ‘gamma’: 0.1, ‘learning_rate’: 0.01, ‘max_depth’: 3, ‘n_estimators’: 100, and ‘subsample’: 0.6. This setup ensures that the model is optimized not just for accuracy but also for the confidence of predictions, enhancing its overall reliability and interpretability. The 5-fold Stratified Cross-Validation for the XGBoost model similarly ensures class balance and provides a comprehensive evaluation of the model’s performance, reducing the likelihood of overfitting and ensuring generalizability across different subsets of the data.
The Random Forest classifier achieved a Cross-Validated ROC AUC of ROC AUC: 0.8114 ± 0.0627. This indicates that the model is relatively stable across different subsets of the data, with the average ROC AUC indicating good discriminatory ability, though the standard deviation suggests some variability.
The XGBoost classifier achieved a Cross-Validated ROC AUC of 0.8098 ± 0.0556. This indicates that the model performs well across different subsets of the data, with the average ROC AUC showing good discriminatory ability and a relatively low standard deviation indicating consistent performance.
Both Random Forest and XGBoost classifiers show moderate to good performance in predicting the binary productivity change outcome. The Random Forest model has a slightly higher accuracy and cross-validated ROC AUC score compared to the XGBoost model, indicating it might perform better on this dataset. However, both models exhibit some variability in performance, as indicated by the standard deviations of the cross-validated ROC AUC scores.
Both models effectively capture non-linear relationships and interactions between features (
Figure 1 and
Figure 2). The high importance of interaction terms and polynomial features underscores this capability.
The consistent importance of features like AI_Integration_Level, AI_Tools_Usage, and their interactions across both models highlights their critical roles in driving productivity changes. The importance of strategic factors (e.g., Innovation_and_Competitiveness_Improvement, Future_Preparedness) indicates that organizations’ strategic approaches to AI integration significantly impact productivity outcomes.
While not as critical as the top features, demographic variables like Age, Years_Using_AI, and Years_with_Company still contribute to the model’s predictive power. This suggests that personal and professional backgrounds also play a role in productivity changes.
4.3. Interpretation of LIME Values for Random Forest Model
In our analysis, we utilized LIME (Local Interpretable Model-agnostic Explanations) to interpret the model’s predictions and gain insights into the contribution of each feature. This technique offered detailed explanations for individual predictions, enhancing our understanding of how different factors influenced the outcomes of our machine learning models.
Based on the aggregated LIME values, the features that have the most significant impact on the Random Forest model’s predictions are presented in
Figure 3, and the features for the XGBoost model are presented in
Figure 4.
For the Random Forest model, the most influential features include the interaction between AI tools usage and AI integration level, AI integration level, and the squared term of AI tools usage. This indicates that both the extent of AI integration and the intensity of AI tools usage, especially when combined, play important roles in predicting productivity changes. Similarly, for the XGBoost model, the key features identified are the same interaction term, innovation and competitiveness improvement, and human resources industry, among others. This consistency across models underscores the importance of how extensively and intensively AI tools are used within the organization, as well as the perceived improvements in innovation and competitiveness. These insights suggest that organizations should focus on the comprehensive integration of AI tools and monitor their usage to maximize productivity benefits.
LIME’s detailed feature importance with specific thresholds (e.g., “AI_Integration_Level > 0.74”) provides a more nuanced understanding of feature impacts compared to the aggregate nature of model-derived importance.
4.4. Bayesian Network Modeling
Logistic regression identified key predictors like ‘Age’ and the interaction between ‘AI Tools Usage’ and ‘AI Integration Level’, but its linear nature limits its ability to capture more complex relationships. Random Forest and XGBoost models, while effective, highlighted the importance of non-linear interactions and feature importance, but their interpretability can be limited. The Bayesian Network approach addresses these limitations by explicitly modeling the probabilistic dependencies among all variables. It allows us to understand not just the direct effects of variables like ‘Age’ and ‘AI Tools Usage’, but also their indirect effects and interactions with other factors.
The next analysis involved the Bayesian model to understand predictors of productivity change. A Bayesian Network was constructed using the HillClimbSearch algorithm with the Bayesian Information Criterion (BIC) as the scoring metric. This approach iteratively explores possible network structures to maximize data fit while balancing model complexity and goodness-of-fit by penalizing overly complex models. Parameters were estimated through Maximum Likelihood Estimation (MLE), ensuring that the conditional probability distributions (CPDs) accurately reflect the observed relationships. Inference was performed using Variable Elimination, which enables exact probabilistic reasoning within the network. This method effectively handles the complex dependencies among ordinal variables, binary encodings, and interaction terms in the dataset. To evaluate the model’s performance, a 5-fold cross-validation approach was employed, mitigating overfitting risk and providing reliable metrics.
The learned structure of the Bayesian Network revealed significant relationships between variables. Notably, ‘AI_Tools_Usage’ demonstrated a direct influence on P (Productivity_Change_Binary|AI_Tools_usage). Additionally, interactions were observed between ‘Innovation_and_Competitiveness_Improvement’ and ‘Job_Opportunities_Creation’ P(Opportunities_Creation|Innovation_and_Competitiveness_Improvement), highlighting the complex interdependencies in the data.
Notably, the interaction between AI tools usage and AI integration level emerged as a critical predictor, indicating that the combined effect of these two factors significantly influences productivity outcomes. This finding is supported by the high conditional probabilities and the frequent appearance of AI-related features in the learned structure. Additionally, features such as ethical policy implementation and future preparedness were closely linked, suggesting that organizations with well-developed ethical considerations are better prepared for future challenges and are likely to experience positive productivity changes.
Demographic and categorical variables also played a significant role in shaping productivity outcomes. For example, the analysis showed that the residence of employees (such as those in Romania, Greece, and Canada) and industry sectors (like IT, telecommunications, and environmental conservation) influenced productivity changes. This highlights the importance of geographical and sectoral contexts in the implementation of AI tools. The network structure also pointed to the interconnectedness of various industry sectors, with the IT industry frequently linked to other sectors like public service and health service, emphasizing the widespread impact of IT on different areas. Overall, the Bayesian network provided a comprehensive view of how different factors interact and contribute to productivity changes, emphasizing the multifaceted nature of AI integration in the workplace.
To evaluate the predictive capability of the network, it was queried to determine the probability of productivity change given specific evidence. Based on the feature importance and LIME values from the XGBoost and Random Forest models, several queries were constructed to explore different scenarios (
Table 6).
Queries involving high levels of AI tools usage and AI integration show a notable likelihood of positive productivity change. For instance, when AI tools usage is high and ethical considerations are extensively addressed (Query 7), there is a significant 72.0% probability of productivity change. Similarly, a high AI integration level coupled with maximum innovation and competitiveness improvement (Query 6) yields a 73.1% probability of productivity change. These findings underscore the importance of comprehensive AI adoption and strong ethical frameworks in driving productivity enhancements within organizations. Additionally, the interaction between AI integration and innovation appears important, as seen in the moderate probabilities of productivity change even with substantial AI tools usage and future preparedness (Query 10), emphasizing the need for balanced and well-integrated AI strategies.
Conversely, scenarios with minimal AI integration and low innovation improvement exhibit low probabilities of productivity change. Query 9, for instance, demonstrates a starkly low 3.4% probability of productivity change when AI integration is minimal and innovation improvement is low. This highlights the potential stagnation in productivity when AI tools and innovative practices are underutilized. Moreover, the findings suggest that even with high education levels and moderate job opportunity creation (Query 3), the probability of productivity change remains relatively low at 27.9%, indicating that factors like AI integration and ethical considerations may play more important roles in driving productivity. Overall, the results advocate for robust AI integration, ethical policy implementation, and continuous innovation as critical levers for enhancing productivity in modern workplaces.
The Bayesian Network model effectively captures the intricate relationships between variables, providing a robust framework for predicting productivity outcomes. These insights can guide organizations in optimizing their AI adoption strategies by focusing on key factors such as AI tools usage, innovation, ethical policies, education, company culture, and AI training.
In order to validate the robustness and generalizability of the Bayesian Network model constructed to predict productivity changes based on various organizational and individual factors, a k-fold cross-validation approach was employed. A 5-fold cross-validation (k = 5) was implemented using the KFold method from scikit-learn, ensuring that the dataset was split into 5 equal parts with shuffling enabled (random_state = 1). For each fold, the training subset was used to learn the structure and parameters of the Bayesian Network using the HillClimbSearch algorithm and BicScore for scoring. The MaximumLikelihoodEstimator was employed for parameter learning, and inference was performed using the VariableElimination method. State names for each variable were obtained, and evidence values were mapped to valid states within the range for each variable to ensure accurate inference. The cross-validation ROC results across the 5 folds were averaged to provide a comprehensive evaluation of the model’s performance. The ROC AUC score of 0.7970 ± 0.0832 and the accuracy of 0.7817 ± 0.0694 demonstrate a strong ability of the model to discriminate between significant and non-significant productivity changes, highlighting the model’s overall discriminative power.
These results validate the Bayesian Network model as a robust predictive tool for assessing productivity changes based on the provided evidence. The high recall and ROC AUC scores are particularly noteworthy, suggesting the model’s potential utility in applications where identifying significant productivity changes is critical.
4.5. ROC Curve Comparison for Predictive Models
To evaluate the performance of the machine learning models used in this study, we generated ROC curves for Logistic Regression, Random Forest, XGBoost, and Bayesian Network (
Figure 5). ROC curves provide a visual representation of the models’ ability to discriminate between productivity changes and non-changes by plotting the True Positive Rate (Sensitivity) against the False Positive Rate (1—Specificity) across various threshold values. The Area Under the Curve (AUC) serves as a summary metric of the model’s performance.
As illustrated in
Figure 5, the Random Forest model outperformed the others with an AUC of 0.94, indicating excellent performance in distinguishing between significant and non-significant productivity changes. XGBoost followed closely behind with an AUC of 0.92, confirming its ability to model complex relationships in the data. The Bayesian Network model achieved an AUC of 0.86, which reflects strong performance but is lower than the tree-based models. Logistic Regression, while effective, had the lowest AUC at 0.84, consistent with its more linear assumptions about the relationships between features and productivity changes.
The close proximity of the ROC curves for Random Forest and XGBoost models emphasizes the strength of ensemble methods in capturing non-linear interactions and complex feature dependencies in the dataset. Notably, the Bayesian Network model also exhibited a good ability to predict productivity changes, highlighting its strength in modeling probabilistic dependencies between variables. Logistic Regression, though slightly outperformed, still showed an adequate level of discrimination, particularly given its simplicity compared to the other methods.
The comparative analysis suggests that Random Forest and XGBoost, with their higher AUCs, are better suited for predicting significant productivity changes in this context, especially when complex interactions between AI tools usage and organizational factors are present.
5. Discussion
The findings of this study provide important insights into the economic impacts of AI on productivity across various organizational contexts. Several key observations emerge from the analysis, highlighting the significance of AI integration and its interplay with other organizational factors in driving productivity changes.
5.1. Key Findings and Their Implications
The most critical finding is that the interaction between AI tools usage and AI integration level significantly enhances productivity. The logistic regression model with interaction terms demonstrated that high levels of AI tools usage, combined with thorough integration within organizational workflows, result in substantial productivity improvements. This underscores the importance of not only adopting AI tools but also ensuring their comprehensive integration within organizational systems. The positive coefficient (β = 0.4319) for this interaction term signifies that increased usage and higher integration levels collectively enhance productivity.
Additionally, the age of employees emerged as a significant predictor of productivity changes, with older age groups associated with lower productivity improvements. This suggests that younger employees might be more adaptable to AI tools or that there might be generational differences in how AI technologies are utilized and embraced in the workplace. These insights highlight the need for targeted training programs that cater to different age groups to maximize AI’s benefits.
5.2. Comparison with Previous Studies
While this study contributes valuable insights into the impact of AI tools on employee productivity, it is important to situate the findings within the broader context of existing research. A significant body of literature has examined the ways AI technologies influence productivity across various industries and sectors. Authors such as Brynjolfsson and McAfee (2014) [
12] have extensively studied the “productivity paradox” in AI adoption, where rapid advancements in technology have not immediately translated into observable productivity gains in many sectors. In contrast, more recent studies, such as those by Czarnitzki et al. (2023) [
5], found evidence of AI-driven productivity growth at the firm level, particularly in technology-intensive industries.
The results align with the theoretical perspectives proposed by [
2,
3], which suggest that AI can enhance productivity by improving decision-making and operational efficiencies. The shift from mid-skill to high-skill and managerial positions reported by [
13] also supports our findings that AI integration fosters higher productivity, particularly in more complex and strategic roles.
However, our study presents a more nuanced view compared to the mixed empirical results reported in the previous literature. For instance, while [
19] found significant increases in patents and trademarks associated with AI but no increase in sales per worker, our findings highlight the critical role of AI integration in realizing productivity gains. This suggests that the benefits of AI may not solely depend on innovation outputs but also on how well AI tools are embedded within organizational processes.
Unlike some studies that suggest AI’s productivity gains are mainly concentrated in specific high-tech industries (Calvino and Fontanelli, 2023) [
9], our research reveals that AI’s positive impact extends across a diverse range of sectors. This suggests that AI’s influence on productivity is not limited to technology-heavy fields but can be observed in traditional industries as well, provided that AI tools are well-integrated into daily operations.
The “productivity paradox” has been a central theme in AI research, where advancements in AI technology often do not translate into immediate productivity improvements at the macroeconomic level. Studies by Parteka and Kordalska (2023) [
21] have discussed this phenomenon in depth, pointing to the slow diffusion of AI technologies across industries and the time lag before benefits materialize. Our study addresses this by focusing on firm-level data and examining productivity changes that occur once AI tools are fully integrated into workflows. The interaction terms between AI tools usage and AI integration in our models demonstrate that productivity gains are realized when there is comprehensive, rather than superficial, AI adoption.
This study applies Bayesian Network Analysis to explore probabilistic dependencies and predict AI’s impact on employee productivity, highlighting the value of robust forecasting in AI-driven environments. This aligns with other authors’ approaches [
30].
While there is substantial literature on AI and productivity, this study makes original contributions by using advanced analytical techniques such as Bayesian Network Analysis and machine learning models, including Random Forest and XGBoost, to explore complex interdependencies. Unlike many studies that rely solely on traditional econometric methods, our approach captures non-linear relationships between variables, revealing that the interaction between AI tools usage and organizational AI integration levels is a critical driver of productivity.
Additionally, our study highlights generational differences in AI adaptability—an area that remains underexplored in the current literature. As our analysis shows, younger employees experience greater productivity gains from AI tools compared to their older counterparts. This suggests that future research should examine not only the technical aspects of AI adoption but also the demographic factors that influence how AI impacts productivity across various employee groups.
5.3. Strengths and Limitations
One of the strengths of this study is the comprehensive dataset obtained from a diverse sample of employees across various industries. The use of advanced modeling techniques, such as logistic regression with interaction terms, Random Forest, and XGBoost, provides robust insights into the factors driving productivity changes. The inclusion of interpretability techniques like SHAP and LIME further enhances the transparency and understanding of model predictions.
However, the study also has limitations. The reliance on self-reported data from the questionnaire may introduce biases related to respondents’ perceptions and experiences. Despite efforts to ensure a representative sample, there may be inherent biases in the data that could affect the generalizability of the findings. Additionally, the cross-sectional nature of the data limits the ability to infer causal relationships between AI usage and productivity changes. While the sample size of 233 responses provides a solid basis for statistical analysis, it may not fully represent the broader workforce. Certain industries or employee groups might be underrepresented, limiting the ability to generalize the findings to other sectors or populations. Additionally, the global diversity of respondents introduces potential regional variations in AI adoption and impact, which may not be fully accounted for in this analysis. Finally, the study does not delve deeply into the potential ethical concerns or organizational challenges associated with AI integration, such as data privacy, transparency, or employee resistance. These factors could significantly affect the success of AI implementation and its overall productivity outcomes, highlighting a need for future research to address these complexities.
5.4. Unexpected Outcomes and Inconclusive Results
Some unexpected outcomes include the relatively low probability of productivity change even with high education levels and moderate job opportunity creation. This suggests that factors like AI integration and ethical considerations may play more important roles in driving productivity than initially anticipated. Furthermore, the relatively lower recall for the positive class (0.52) in the logistic regression model indicates that there may be other unobserved factors influencing productivity changes that were not captured in the study.
The Bayesian Network analysis provided additional insights by capturing the intricate relationships between variables and offering a robust framework for predicting productivity outcomes. Queries involving high levels of AI tools usage and integration consistently showed high probabilities of productivity change, emphasizing the importance of a well-rounded AI strategy.