0 cumulative citations
View corpus contextEngagement drives performance in AI-enabled factory work but weak self-regulation caps gains; a survey-informed rule system using PLS-SEM and NCA classifies worker performance with 87% accuracy.
AI-Based Decision Support System for Learning Growth Performance Workforce Use Certainty Factor and Constraint Evaluation
openalex
correlational
low evidence
7/10 relevance
Full text
usable extracted full text
DOI
Source
PDF
Using survey data from 200 Indonesian manufacturing workers, the authors find Engagement strongly predicts performance (β=0.519) while Self-Regulation operates as a necessary bottleneck (NCA d=0.189), and they translate these findings into a Certainty-Factor expert system that classifies worker performance with ~87% accuracy on a held-out test.
Citation observations
Cumulative provider counts captured on specific dates; providers are never combined.
This study presents the design and computational evaluation of an AI-based decision support system for workforce performance in an AI-driven environment under constrained conditions. Previous research has largely emphasized technological determinants and adequate conditions, while neglecting the constraint mechanisms that limit achievable performance. To address this gap, this study integrates Partial Least Squares Structural Equation Modeling (PLS-SEM) and Necessary Condition Analysis (NCA) into a unified computational framework. Empirical data from 200 manufacturing workers was used to derive model parameters. The results identified Engagement Level as a key performance driver (β = 0.519), while Self-Regulation Ability emerged as a critical constraint (d = 0.189). These findings were operationalized into a Certainty Factor-based expert system that models performance as a function of positive contribution, negative influence, and constraint thresholds. The proposed model was evaluated using classification metrics, achieving 87% accuracy, 0.85 precision, 0.86 recall, and an F1-score of 0.85. The results demonstrate strong predictive capabilities and confirm that performance is jointly determined by enabling and constraining factors. This study contributes by bridging statistical analysis and AI system design, providing a constraint-aware decision support model for performance evaluation in complex operational environments.
Summary
Main Finding
The authors design and evaluate a constraint-aware AI decision support system that combines PLS-SEM and Necessary Condition Analysis (NCA) to parameterize a Certainty Factor (CF) expert model for predicting workforce performance in AI-enabled manufacturing. Engagement (ENG) is the main performance driver (β = 0.519), Self-Regulation Capability (SRC) is the primary bottleneck (NCA effect d = 0.189) and is implemented as a hard constraint (threshold θ_SRC = 0.50). The CF-based classifier attains strong predictive performance on a held-out test set (accuracy ≈ 87.5%; precision 0.85; recall 0.86; F1 = 0.85).
Key Points
- Hybrid analytical pipeline: PLS-SEM (identifies sufficiency/weights) + NCA (identifies necessary/bottleneck conditions) → empirical parameters for an expert-system CF model.
- Main variables (normalized to [0,1]):
- ENG (Engagement Level) — main positive driver (β = 0.519).
- AIS (AI System Support) — positive driver (β = 0.522) and supports ENG.
- STR (Operational Stress) — negative factor (β = 0.131).
- SRC (Self-Regulation Capability) — hard constraint; NCA d = 0.189; threshold θ_SRC = 0.50.
- DC (Digital Capability) — adjustment/enhancer (β = 0.494 toward ENG; weak NCA d = 0.065).
- CF model formulation:
- MB = 0.519·ENG + 0.522·AIS
- MD = 0.131·STR
- CF = MB − MD
- If DC > sample mean then CF ← CF + 0.494·DC
- If SRC < 0.50 → classify Performance = Low (constraint gate)
- CF thresholds: CF < 0.40 → Low; 0.40 ≤ CF ≤ 0.70 → Medium; CF > 0.70 → High
- Evaluation: Sample N = 200 (training 160, test 40). Confusion matrix on test set produced high diagonal agreement; metrics above.
- Measurement quality: indicator loadings mostly >0.70; Cronbach’s α = 0.683–0.827; composite reliability >0.80; AVE >0.50; HTMT <0.85.
- Preprocessing: missing-value screening, outlier removal (z ±3.29), min–max normalization, Harman’s single-factor test for common-method bias.
Data & Methods
- Data: Survey responses from 200 manufacturing workers in Indonesia (2021–2024) using AI-enabled systems. Performance labels validated via independent supervisor assessments.
- Methods:
- Measurement model assessment (reliability/validity) and structural model estimation via PLS-SEM to obtain path coefficients and effect sizes.
- Necessary Condition Analysis (CR-FDH) to identify minimum levels (bottlenecks) and compute NCA effect sizes; threshold θ_SRC = 0.50 chosen by CR-FDH and ROC/Youden optimization (J = 0.42).
- Conversion of PLS-SEM βs and NCA ds into CF expert-system parameters (empirically derived weights rather than expert elicitation).
- Simple rule-based CF classifier (algorithmic steps described above). Train/test split 80/20; evaluation with accuracy, precision, recall, F1.
- Limitations noted by authors (implicit/explicit):
- Single-country, manufacturing-sector sample limits external generalizability.
- Relies on survey/self-report indicators augmented by supervisor labels; potential measurement issues.
- CF design choices (e.g., DC additive rule, discrete CF thresholds) have pragmatic justification but are somewhat ad hoc and may need further validation.
Implications for AI Economics
- Complementarities and bottlenecks:
- The study quantifies that technological inputs (AI system support, digital capability) and human behavioral factors (engagement) are complementary but constrained by necessary human capabilities (self-regulation). Economic models of AI adoption should therefore include both sufficiency (drivers) and necessity (bottlenecks) conditions to predict realized productivity gains.
- Investment prioritization and returns to training:
- Because SRC acts as a hard constraint, investments in technology (AI tools) may yield limited returns unless accompanied by investments in worker self-regulation and related human-capital interventions. Cost–benefit analyses of AI deployment should incorporate the marginal gains conditional on raising necessary human-capability thresholds.
- Policy and organizational strategy:
- Policies that subsidize technology without parallel support for workforce behavioral readiness (training, job design, workload management) risk underdelivering on productivity—models can be adapted to simulate the joint effects and identify thresholds where tech investments become productive.
- Labor-market modeling:
- Modeling labor productivity and wage returns in AI-intensive settings should include non-linear threshold effects (necessary conditions) rather than only linear complementarities; ignoring bottlenecks can overstate elasticities of output with respect to AI capital.
- Measurement and forecasting:
- Deployable constraint-aware predictive models (as demonstrated) can improve organizational forecasting of productivity improvements from AI adoption by flagging when human constraints are binding—enabling better timing and sequencing of investments.
- Research agenda for AI economics:
- Incorporate NCA-like necessary-condition thinking into structural micro- and macro-econometric models of technology diffusion and productivity.
- Evaluate heterogeneous returns across firms/sectors by estimating binding constraints (e.g., SRC equivalents) and simulate policy levers (training vs. capital subsidies).
- Extend cost-effectiveness analyses to account for rule-based decision systems (CF) and empirically derived thresholds, and test welfare implications of different deployment strategies.
Suggested next steps to apply these insights in economic analysis: replicate across sectors/countries, integrate cost parameters for interventions that raise necessary conditions (e.g., training), and embed constraint-aware productivity functions into firm-level production functions or general-equilibrium models to assess aggregate impacts.
Assessment
Paper Typecorrelational
Evidence Strengthlow — All empirical evidence is observational and cross-sectional from a single convenience/sector sample; PLS-SEM provides associative rather than causal inference, NCA reports necessary-condition associations but does not establish causality, and model parameters/thresholds are estimated on the same dataset used to develop the rule system with a small held-out test (n=40), limiting external validity and robustness to sampling or unobserved confounding.
Methods Rigormedium — The authors use appropriate and modern tools for survey-based analysis (PLS-SEM measurement checks, NCA for bottlenecks, and a held-out test set for predictive evaluation). However, there are important weaknesses: reliance on self-report measures, limited description of sampling frame beyond 'simple random sampling', potential circularity in deriving CF weights from the same sample used to evaluate the CF classifier, small test sample (n=40), and no sensitivity checks, out-of-sample validation beyond a single split, or strategies to address omitted variable bias.
SampleSurvey of 200 manufacturing workers in Indonesia (data collected 2021–2024) who operate AI-enabled digital systems; constructs measured via adapted Likert-scale items (ENG, AIS, STR, SRC, DC, PERF); 160 observations used to fit/derive model parameters and thresholds, 40 held out for test evaluation; ground-truth performance labels provided by supervisors.
Themesproductivity human_ai_collab
IdentificationNo exogenous causal identification; the paper infers directional relationships from cross-sectional survey data using PLS-SEM path coefficients and complements those associations with Necessary Condition Analysis (NCA) to identify bottleneck thresholds; derived coefficients and NCA effect sizes are then mapped into a Certainty Factor (CF) rule-based classifier and evaluated via an 80/20 train-test split using supervisor-assigned labels.
GeneralizabilitySingle-country (Indonesia) and single-sector (manufacturing) sample limits transferability to other industries and economies, Sample size modest (n=200) with small test set (n=40), increasing sampling variability, Measures largely self-reported and performance labels are supervisory assessments (potential subjectivity and common-method concerns), ‘AI System Support’ is an aggregate survey construct that may mask wide heterogeneity in types/levels of AI technology used, Thresholds and CF weights are empirically tuned to this sample and may not generalize without external validation
Claims (8)
| Claim | Direction | Outcome | Confidence & Evidence | Details |
|---|---|---|---|---|
| Among the 200 Indonesian manufacturing workers studied, Engagement Level was the strongest reported positive predictor of workforce performance (β = 0.519, p < 0.001). Organizational Efficiency | positive | Workforce performance |
Reading fidelity
high
Study strength
medium
|
n=200
β = 0.519
|
| Self-Regulation Capability was positively associated with workforce performance (β = 0.345, p < 0.001). Organizational Efficiency | positive | Workforce performance |
Reading fidelity
high
Study strength
medium
|
n=200
β = 0.345
|
| AI System Support was positively associated with Engagement Level (β = 0.522, p < 0.001). Worker Satisfaction | positive | Employee engagement |
Reading fidelity
high
Study strength
medium
|
n=200
β = 0.522
|
| Digital Capability was positively associated with Engagement Level (β = 0.494, p < 0.001). Worker Satisfaction | positive | Employee engagement |
Reading fidelity
high
Study strength
medium
|
n=200
β = 0.494
|
| Self-Regulation Capability was identified as the primary bottleneck condition: high performance was reported as unattainable below a critical self-regulation threshold, regardless of other enabling factors. Organizational Efficiency | negative | Attainment of high workforce performance |
Reading fidelity
high
Study strength
medium
|
n=200
d = 0.189
|
| The model used a normalized Self-Regulation Capability threshold of 0.50 as a gate: observations below this threshold were classified as having low performance. Organizational Efficiency | negative | Performance classification |
Reading fidelity
high
Study strength
medium
|
n=200
θ_SRC = 0.50
|
| The Certainty Factor decision-support model achieved 87.5% accuracy on its held-out test set. Decision Quality | positive | Performance-category classification accuracy |
Reading fidelity
high
Study strength
medium
|
n=40
87.5% accuracy
|
| On the held-out test set, the model achieved precision of 0.85, recall of 0.86, and an F1-score of 0.85 when classifying low, medium, and high performance. Decision Quality | positive | Precision, recall, and F1-score of performance classification |
Reading fidelity
high
Study strength
medium
|
n=40
precision = 0.85; recall = 0.86; F1-score = 0.85
|