<![CDATA[Small taxi fleet operators operating in emerging markets generate rich GPS telemetry through every completed trip, yet that data is rarely fed back into operational decisions. This study evaluates three AI-augmented capabilities, demand forecasting, predictive fare modeling, and driver performance scoring, deployed within Digikab, a SaaS platform built on GPS-enabled Android devices, across 47 independent fleet operators managing 284 vehicles over a six-month observation window. Using a quasi-experimental design with 29 AI-enabled operators and 18 controls on the same platform, this study applies gradient-boosted tree models (XGBoost/LightGBM) for zone-level demand forecasting, achieving a mean absolute error (MAE) of 2.14 trips per zone per hour (RMSE = 3.87, R² = 0.81) against held-out validation data. Demand elasticity estimation for fare optimization is formalized through an arc elasticity framework applied to operator-specific tariff history. Driver performance is decomposed into five weighted behavioral dimensions derived exclusively from the existing trip record, with coaching feedback generated via a large language model (LLM). Operators in the AI-enabled cohort demonstrated a 14.3% improvement in revenue per driver-hour (p = 0.003), a 22.4% reduction in idle time proportion (p = 0.007), an 8.9 percentage-point increase in trip completion rate (p = 0.014), and a 19.0% reduction in average passenger wait time (p = 0.012). The cold-start challenge for new operators is addressed through a progressive blending architecture that phases per-operator training signal in over a minimum three-month accumulation period. Findings demonstrate that meaningful predictive fleet intelligence is achievable at trip-history volumes in the thousands rather than millions, provided model architecture, output design, and operator interface are co-designed for the small-fleet context.]]>
Copyrights © 2026
