Artificial Intelligence Readiness in Kenya’s Public Sector: A Machine Learning and Principal Component Analysis of AI Readiness in Embu County

Abstract

Public institutions are increasingly expected to deliver services through digital and AI-enabled platforms, yet the readiness of their workforce to adopt these technologies remains largely unmeasured, particularly in Kenya's devolved governments. This study assesses AI readiness among 150 public servants in Embu County using a structured survey of twenty-four 5-point Likert items covering the workplace environment, digital capability, governance and individual competence, together with a composite AI_Readiness index. The analysis had two layers. First, item-level descriptive statistics (means, standard deviations and interpretive thresholds) profiled the readiness landscape. Second, five regression models were compared: Random Forest (RF), a Back-Propagation Neural Network (BPNN), a Decision Tree (CART), Support Vector Regression with a radial basis function kernel (SVR-RBF) and a PCA-augmented Random Forest. Each model was trained on a 75/25 hold-out split and validated through 5-fold cross-validation (seed = 123), and performance was assessed by MAE, RMSE, Pearson's r, R² and threshold-based accuracy.

Overall readiness was high (M = 4.2). It was driven by strong psychological willingness (M = 4.5) and perceived ICT infrastructure (M = 4.4), but constrained by weaker institutional enablers: resource allocation (M = 3.7), innovation culture (M = 3.8) and AI governance rules (M = 3.9).

On the composite target, the Random Forest was the strongest model overall (r = 0.9742; R² = 0.9491; MAE = 0.0969; RMSE = 0.1340; exact-match accuracy = 97.37%). The BPNN was a very close second (r = 0.9621; R² = 0.9257; MAE = 0.0972; RMSE = 0.1243; exact-match accuracy = 86.84%). Although the BPNN achieved a marginally lower RMSE (0.1243 versus 0.1340), indicating fewer large errors, the Random Forest was selected as the best overall model on the basis of three complementary criteria: (i) a higher Pearson correlation and R², meaning it explained a greater proportion of variance in the composite readiness score; (ii) a substantially higher exact-match and threshold-based accuracy (97.37% versus 86.84%), which is the more relevant metric for a diagnostic readiness dashboard where correct classification of respondents into readiness bands matters more than minimising average squared error; and (iii) greater cross-validated stability, with lower fold-to-fold variability in r and MAE. The BPNN's lower RMSE reflects a smoother error distribution, but its lower classification accuracy and higher instability across folds make it less suitable for practical deployment. The Decision Tree trailed substantially (r = 0.8462; R² = 0.7161; MAE = 0.2045; RMSE = 0.2550).

At the item level, no single model dominated: CART performed best for Resources (r = 0.953 ± 0.075) and Technical Training (r = 0.950), RF for Leadership (r = 0.912 ± 0.023), and regularised radial-basis regression for Digital Decision-Making (r = 0.809 ± 0.046). A consistent structural pattern emerged. Organisational items were tightly inter-correlated and highly predictable (CV r ≈ 0.85–0.95), whereas individual AI confidence, willingness and verification behavior were largely independent of the workplace environment (CV r ≈ 0.10–0.58). Principal component analysis yielded four interpretable dimensions: Infrastructure and Governance; Training and Digital Capability; Skills–Opportunity Constraint; and Leadership versus Individual AI Skill. The findings support ensemble and neural learners as reliable diagnostic tools for appraising AI readiness in devolved government. They also show that Embu County's readiness gap is institutional rather than attitudinal, so county interventions should prioritise resource allocation, a clear AI policy framework and innovation incentives over further general training.