Machine learning can forecast firms' self-reported AI/IoT success: Random Forests cut prediction error versus regularized linear models, and perceived value realization is the dominant signal—rising sharply in the mid-range and plateauing at high levels. Results are predictive and interpretable but based on self-reports and do not establish causation.

Main Finding

A Random Forest model outperforms a regularized linear baseline (Elastic Net) and a simple Decision Tree in predicting firm-reported AI/IoT success (0–100 scale) using Czech and Slovak enterprise survey data (n = 1,250). The single strongest predictive signal is value realization (ai_iot_advantage_share); readiness and performance measures increase predicted success, while reported barriers reduce it. Nonlinear diagnostics reveal a mid-range threshold effect and saturation at high advantage-attribution values. Results are presented as an interpretable screening tool for firms and managers, not as causal evidence.

Key Points

• Outcome: reported AI/IoT success measured on a bounded 0–100 scale.
• Data: enterprise survey from Slovakia and the Czech Republic, n = 1,250.
• Models compared: Elastic Net (regularized linear baseline), Decision Tree, Random Forest.
• Evaluation: consistent out-of-sample evaluation framework (train/test procedures).
• Best performer: Random Forest achieves the strongest out-of-sample predictive performance and reduces absolute prediction errors relative to Elastic Net for the majority of test observations.
• Error characteristics: overall improvement with Random Forest, but diagnostics reveal a small tail of extreme prediction errors.
• Feature importance: ai_iot_advantage_share (share of value/advantage attributed to AI/IoT) is the most stable and informative predictor across model families.
• Nonlinear patterns:
  • Threshold-like transition in predicted success around the mid-range of advantage attribution.
  • Saturation pattern at higher advantage-attribution values (diminishing marginal predicted gains).
• Other predictors: readiness and performance-related variables are positively associated with predicted success; higher reported barriers are negatively associated.
• Interpretation: models used for interpretable screening and managerial reflection; authors explicitly avoid causal claims.

Data & Methods

• Data source: firm-level enterprise survey responses from Czech Republic and Slovakia (n = 1,250).
• Outcome variable: self-reported AI/IoT success score bounded between 0 and 100.
• Feature set: variables capturing perceived AI/IoT advantages (ai_iot_advantage_share), readiness indicators, performance measures, and barrier indicators, among others.
• Modeling approaches:
  • Elastic Net: regularized linear baseline to capture linear associations and do variable selection/shrinkage.
  • Decision Tree: single-tree nonlinear baseline for simple partitioning interpretation.
  • Random Forest: ensemble of trees to capture complex nonlinearities and interactions.
• Evaluation protocol: consistent out-of-sample testing (train/test splits or cross-validation) to compare predictive generalization.
• Model interpretability tools:
  • Permutation feature importance for ranking predictors.
  • Partial Dependence Plots (PDP) and Individual Conditional Expectation (ICE) curves to inspect average and instance-level nonlinear effects, respectively.
• Diagnostics: evaluation of absolute errors across observations, investigation of error tails and model behavior across predictor ranges.

Implications for AI Economics

• Value realization (the share of advantage firms attribute to AI/IoT) is the single most informative predictive signal for reported success. For researchers and policymakers, this suggests measuring realized business value is critical when studying AI adoption outcomes.
• Nonlinear relationships matter: threshold and saturation effects imply that moderate increases in perceived advantage can precipitate larger changes in reported success up to a point, after which gains level off. Policy or managerial interventions may be most effective when targeted at firms around the mid-range of perceived advantage.
• Readiness and performance indicators predict higher reported success, while barriers predict lower success—supporting targeted readiness-building and barrier-reduction policies (e.g., skills, integration support, financing).
• Use case: Random Forest-based screening tools can help identify firms likely to report high or low AI/IoT success for tailored advisory or subsidy programs, but designers should account for a small tail of extreme prediction errors (robustness/uncertainty measures are advisable).
• Limits and cautions:
  • Predictive, not causal: associations should not be interpreted as causal effects. Experimental or quasi-experimental designs are needed to establish causality.
  • External validity: results are specific to Czech and Slovak firms and the survey instrument; applicability to other contexts requires validation.
  • Data needs: richer longitudinal or administrative outcome data would strengthen causal and economic inference about AI value realization.
• Recommended next steps for AI economics research: combine predictive screening with causal identification (e.g., matched studies, randomized trials, instrumental variables), improve measurement of realized value, and explore model ensembles/calibration to mitigate extreme-error tails.