The rapid advancement of generative AI and autonomous technologies has introduced significant uncertainty in global labor markets. Existing economic models often fail to represent the complex nonlinear relationships between automation capability, labor costs, and workforce adaptability. This limitation highlights the need for more intelligent frameworks capable of both predicting workforce displacement and generating actionable decision outputs. The motivation behind this work is to design an Agentic AI-based system that bridges predictive analytics and decision intelligence, enabling a more comprehensive understanding of labor market evolution under AI influence.

This study proposes a novel Agentic AI-inspired framework for workforce substitution analysis and automation decision support within economic environments. Unlike conventional approaches that focus solely on prediction, the proposed framework integrates forecasting, decision intelligence, and adaptive feedback mechanisms into a unified architecture.

The framework employs a dual-task learning strategy in which regression models estimate workforce substitution timelines, while classification models recommend automation actions, including monitoring, assistance, or full automation. To improve predictive robustness and generalization, a stacking-based ensemble learning mechanism combines multiple machine learning models within a hybrid predictive layer.

A key contribution of this work is the introduction of a decision intelligence layer that transforms predictive outputs into actionable recommendations, enabling the system to move beyond passive prediction toward goal-oriented decision support. In addition, a feedback mechanism allows the framework to continuously incorporate execution outcomes, supporting adaptive learning and policy refinement over time.

The main contributions of this study can be summarized as follows:

1) Development of an Agentic AI-inspired architecture that integrates prediction, decision support, and adaptive feedback within a unified framework.

2) Introduction of a hybrid ensemble learning approach that combines regression and classification models to improve workforce substitution forecasting and automation policy selection.

3) Design of a risk-aware decision intelligence module that translates predictive insights into actionable workforce automation strategies.

4) Comprehensive evaluation through performance analysis, robustness testing, cross-validation, explainability assessment, and scalability experiments.

The proposed framework provides an interpretable and practical foundation for intelligent workforce transformation analysis and AI-assisted economic decision-making.

Although significant progress has been made in artificial intelligence and labor economics, several limitations remain in existing approaches. Most studies treat prediction and decision-making as independent tasks, relying separately on regression or classification models without integration. Moreover, the application of agentic AI concepts in economic forecasting remains limited, reducing the ability of current systems to simulate autonomous decision behavior. In addition, there is a lack of hybrid ensemble frameworks capable of effectively combining multiple machine learning models for workforce analysis. Another key limitation is the absence of decision-intelligence layers that transform predictions into actionable policies. These gaps indicate the need for a unified framework that integrates prediction, classification, and decision-making within a single architecture.

The Data Layer serves as the foundation of the proposed system. This study utilizes the 2026 Intelligence Economy: Labor vs. AI Compute dataset, publicly available through the Kaggle platform and accessible at the Kaggle repository [28]. The dataset provides a macroeconomic simulation of the projected substitution point between human professional labor and AI agentic compute in the 2026 intelligence economy. The sampling unit consists of occupation-technology observations characterized by workforce, automation, and AI-compute-related attributes. The dataset covers the 2026 forecasting horizon and includes variables such as automation risk, AI augmentation potential, labor costs, substitution elasticity, and other economic indicators relevant to workforce transformation. Because the data are generated through a structured economic simulation rather than collected from direct observations or surveys, the dataset is classified as simulated (synthetic) rather than empirical. The public availability of the dataset and its documented feature definitions support the validity, transparency, and reproducibility of the experimental evaluation. The layer ingests structured data and performs feature extraction, normalization, and feature selection to prepare the data for downstream modeling.

3.1.1. Target Variable and Feature Definitions

The primary regression target in this study is Substitution_Year_Est, referred to as the workforce substitution year, which represents the projected calendar year at which AI-based systems become economically competitive with human labor for a given occupation-technology observation. The dataset contains several key explanatory variables. Automation_Risk_Index represents the estimated susceptibility of an occupation to automation and is provided directly as a normalized risk indicator.

AI_Augmentation_Factor measures the extent to which AI can enhance human productivity without necessarily replacing workers and is included as a dataset attribute. Substitution_Elasticity captures the responsiveness of workforce substitution to changes in AI capability and deployment costs, indicating how readily AI may replace human labor in a given occupation. Additional variables include Human_Labor_Cost_hr, which represents the hourly cost of human labor, and Inference_Cost_2026, which reflects the projected cost of AI inference in 2026. During preprocessing, all numerical features were normalized and used as inputs to the regression and classification models. The original dataset fields were retained, while additional interaction features such as AI × Cost, Risk × Elasticity, and Human-AI Gap were derived to capture nonlinear relationships among workforce, economic, and automation factors.

3.1.2. Fitness Function for Data Layer

where:

N missing = number of missing values;

N total = total number of data points;

σ noise 2 = estimated noise variance.

The Model Layer contains three complementary algorithms running in parallel.

CatBoost Regressor:

CatBoost (Categorical Boosting) handles regression tasks such as continuous value prediction and time series forecasting. It natively processes categorical features using ordered boosting and symmetric trees.

Fitness Function:

where:

Extra Trees Regressor:

Extra Trees (Extremely Randomized Trees) provides an additional regression pathway with higher variance reduction through random splits, complementing CatBoost by capturing different data patterns.

Fitness Function:

where:

CatBoost Classifier:

CatBoost Classifier handles categorical label assignment and anomaly detection using ordered boosting to prevent target leakage.

Fitness Function:

where:

Precision = TP/(TP + FP).

Recall = TP/(TP + FN).

Accuracy = (TP + TN)/(TP + TN + FP + FN).

Combined Model Layer Fitness:

where:

w 1 , w 2 , w 3 are algorithm weights with ∑ w i = 1 .

The Fusion Layer aggregates outputs from the three base algorithms using a stacking ensemble approach.

Stacking Architecture:

Base learners (Level 0): CatBoost Regressor, Extra Trees Regressor, CatBoost Classifier.

Meta-learner (Level 1): Logistic Regression (classification) or Ridge Regression (regression).

Fitness Function for Stacking Meta-Model:

where:

F i = fitness of base algorithm i.

w i = optimal stacking weights.

λ = diversity coefficient (0.2).

Fitness Function for Meta-Learner:

Cross-Validation Fitness (5-fold):

The Decision & Action Layer translates fused predictions into concrete actions. The Decision Engine performs inference, applies predefined policies, and validates outcomes. The Action Executor triggers responses through API calls, workflows, or notifications.

Decision Class Generation:

The original dataset does not provide categorical automation recommendations. Therefore, the decision classes used by the classification component were derived from the workforce substitution year estimate (Substitution_Year_Est). To transform the continuous workforce substitution forecast into actionable recommendations, occupations were grouped into three automation-readiness categories using temporal thresholds. Occupations with the earliest projected substitution years were assigned to the Automate class, indicating that AI deployment is economically feasible in the near term. Occupations with intermediate substitution timelines were assigned to the Assist class, indicating that AI should augment human workers while maintaining human oversight. Occupations with later projected substitution timelines were assigned to the Monitor class, indicating that full automation is not immediately justified, and that continued observation is recommended. This policy-based categorization enables the proposed framework to convert predictive outputs into interpretable and actionable workforce automation decisions.

Fitness Function for Decision Engine:

where:

τ delay = average decision latency (seconds).

Fitness Function for Action Executor:

The total fitness combines all layer-specific functions:

where:

∑ α i =1 (layer weights from cross-validation).

ϵ = cost penalty coefficient (0.01 - 0.05).

C coast = normalized computational cost.

Optimization Objective:

Subject to:

CatBoost: learning rate ϵ [0.01, 0.3], depth ϵ [3, 10];

Extra Trees: n_estimators ϵ [50, 500], min_samples_split ϵ [2, 20].

A feedback loop from the Action Executor back to the Data Layer enables continuous improvement. Action outcomes are captured as new data points for retraining.

Feedback Fitness:

Positive F f b indicates system improvement over time.

Figure 2 presents a comparative bar chart of R² scores across the three models. Table 2 provides a detailed comparison of MAE, RMSE, and R² values.

Figure 2. Model performance comparison (R² score).

Table 2.Model performance comparison.

As shown in Table 2, The proposed Hybrid model achieves an R² of 0.8574, outperforming all individual baselines: Random Forest (0.6502), CatBoost (0.6379), Extra Trees (0.6246), XGBoost (0.6163), and LightGBM (0.5943). This represents a 31.9% to 44.2% improvement in explained variance. The Hybrid model also reduces MAE by 36.6% and RMSE by 36.8% compared to the best individual model (Random Forest), demonstrating the superior predictive power of the stacking ensemble approach.

Figure 3 displays the rest heals of the Hybrid model are centered at zero (median = 0.00) with a symmetric distribution, indicating no systematic bias. The residual range (−1.5 to 1.7) is substantially narrower than all individual models (which range from −2.9 to 2.8). The interquartile range (IQR = 1.1) is also the smallest among all models, with the next best being Random Forest (IQR = 1.4). These results confirm that the proposed Hybrid model provides more stable, accurate, and reliable predictions compared to individual baseline models.

Figure 3. Hybrid model residual distribution.

Table 3. Residual distribution statistics.

As shown in Table 3, The Hybrid model demonstrates the narrowest residual range (−1.5 to 1.7) compared to all individual baseline models: Extra Trees (−2.4 to 2.6), CatBoost (−2.8 to 2.5), XGBoost (−2.5 to 2.7), LightGBM (−2.9 to 2.8), and Random Forest (−2.3 to 2.4), indicating more stable and reliable predictions.

Figure 4 presents the learning curve of the CatBoost Regressor, showing RMSE as a function of training iterations. The curve demonstrates rapid initial convergence, with RMSE decreasing sharply within the first few iterations before stabilizing. This indicates efficient learning without significant overfitting.

Figure 4. All models learning curve.

Figure 5 presents a scatter plot comparing actual years against predicted years. The alignment between actual and predicted values along the diagonal indicates strong predictive performance. Points clustering near the identity line (y = x) demonstrate that the model effectively captures temporal trends in the data.

Figure 5. Actual vs. Predicted year.

The performance of the Agentic Decision Engine was evaluated using a confusion matrix for three-class classification (classes 0, 1, and 2). Figure 6 and Table 4 present the confusion matrix.

Figure 6.Agent Decision confusion matrix.

Table 4. Confusion matrix results.

The confusion matrix reveals strong classification performance with high diagonal values. Overall accuracy is calculated as:

Table 5 presents the class-wise performance metrics.

Table 5. Class-wise performance metrics.

Figure 7 illustrates the distribution of action classes executed by the Action Executor. The distribution appears relatively balanced across classes, indicating that the system does not exhibit bias toward any particular action type.

Figure 7.Action class distribution.

Figure 8 provides a comprehensive bar chart comparing all three models across MAE, RMSE, and R² metrics. The visual comparison reinforces the quantitative findings, with the Hybrid model consistently outperforming both individual models.

Figure 8. Model performance comparison (MAE, RMSE, R²).

Figure 9 presents a box plot comparison of residual distributions for Extra Trees, CatBoost, and the Hybrid model. The Hybrid model exhibits the smallest interquartile range (Q3-Q1 = 1.1) and the narrowest overall whiskers, confirming its superior predictive stability and lower error variance compared to the individual models.

Figure 9. Residual distribution box plot.

Table 6 presents the results of the paired t-test and Wilcoxon signed-rank test conducted to compare the proposed Hybrid model with the baseline approaches. The obtained p-values for all comparisons exceed the conventional significance threshold (α = 0.05). Therefore, the null hypothesis cannot be rejected, and the observed performance differences cannot be considered statistically significant under the current experimental setting.

Nevertheless, the proposed Hybrid model consistently achieved superior predictive performance across all evaluation metrics. In particular, the model attained the highest R² score (0.8574) and the lowest MAE (0.5646) and RMSE (0.6656) among all evaluated methods. These improvements indicate meaningful practical benefits in terms of prediction accuracy, robustness, and decision-support capability. Consequently, while statistical significance was not established, the results suggest that the proposed framework offers practical value for workforce substitution forecasting and automation decision-making applications.

Table 6.Statistical significance test results.

We employed Repeated K-Fold cross-validation to assess model stability and generalization capability. Table 7 presents the cross-validation configuration and results.

Table 7.Cross-validation configuration and results.

The consistent performance across 50 train-test splits (10 folds × 5 repeats) with a narrow standard deviation of 0.0394 demonstrates that the model is stable and generalizes well to unseen data. The 95% confidence interval provides statistical assurance of the model’s reliability.

We applied Grid Search and Optuna-based Bayesian optimization to tune model hyperparameters for optimal performance. Table 8 presents optimization results.

Table 8. Hyperparameter optimization results.

The optimization results confirm that the default hyperparameters used in the proposed Hybrid model are near-optimal for this dataset. The Optuna-based Bayesian optimization achieved a best R² of 0.6500 for CatBoost, which is consistent with the default configuration (0.6379), validating our parameter selection.

To interpret model predictions and enhance transparency, we employed SHAP (Shapley Additive explanations) and LIME (Local Interpretable Model-agnostic Explanations) as shown in Table 9.

Table 9. Feature importance rankings for substitution year prediction.

The analysis reveals that automation risk, AI augmentation potential, and substitution elasticity are the primary drivers of workforce substitution timing, providing actionable insights for policymakers and business leaders.

We evaluated model robustness under three real-world scenarios: noise injection, missing data, and data drift. Table 10 presents the results.

The model maintains R² > 0.62 under all tested conditions. Under 10% Gaussian noise, the R² degraded by less than 0.01. Under 10% missing data, the model-maintained R² > 0.63. Under moderate data drift (drift strength = 1.0), the R² remained above 0.63. These results demonstrate strong robustness to real-world data imperfections, making the model suitable for practical deployment.

Table 10.Robustness analysis results.

We conducted scalability experiments to measure training time, memory consumption, and response speed across different dataset sizes. Table 11 presents the results.

Table 11. Scalability results.

Key Findings:

Training time scales approximately linearly with dataset size (R² = 0.998 for linear fit);

Memory usage remains low (<0.5 MB) even for the full dataset;

Inference speed is under 0.1 seconds per 100 samples, enabling real-time predictions;

These results confirm that the proposed model is computationally efficient and suitable for production deployment in real-time decision support systems.

Table 12 compares our proposed method with existing state-of-the-art approaches in agentic AI and labor market analysis.

Table 12. Comparison with published studies.

Key Differentiators of Our Method:

Provides quantitative predictions (R² = 0.8574) rather than categorical assessments;

Achieves high classification accuracy (95.4%) for actionable decisions;

Unifies regression and classification within a single agentic framework;

Includes comprehensive validation (statistical tests, robustness, scalability);

Our method provides both accurate numerical predictions and reliable classification, making it suitable for real-world workforce transformation analysis.

Table 13 presents the Precision-Recall curves for the multi-class classification task, complementing the ROC analysis.

Table 13. Precision recall curves for three-class classification.

The high Average Precision scores (all > 0.94) confirm that the classifier maintains both high precision and high recall across all action classes, with particularly strong performance for Class 2 (Automate), achieving AP = 0.98.