Machine-learning personalization lifted Telkomsel’s data-package take-up by 6.6%, delivering IDR 141.6m in incremental revenue in a three-month trial; a CatBoost model flagged monthly data revenue and usage frequency as the top predictors of purchases.

Main Finding

A Customer Value Management (CVM) program at PT Telkomsel, supported by a CatBoost Gradient Boosting Decision Tree model trained on 5.7 million customer transaction records (Jan–Mar 2023), improved personalized targeting: the model achieved 86% accuracy (for predicting potential lapsers/purchase propensity), the CVM-driven test-and-learn campaign raised take-up rate by 6.55%, and generated an incremental revenue uplift of IDR 141.6 million. The strongest predictors of purchase were monthly data-package revenue, frequency of data usage within specific price ranges, and total monthly data revenue.

Key Points

• Objective: Measure impact of CVM + machine learning on increasing purchases of Telkomsel cellular data packages and identify behavioral drivers of purchase decisions.
• Data scale: 5.7 million customer transaction records over a 3-month window (Jan–Mar 2023).
• Modeling approach: Supervised classification using CatBoost (GBDT). Model used to predict lapsing/purchase propensity and drive personalized offers via Telkomsel’s CVM/DMP and MyTelkomsel app.
• Performance: CatBoost model reported 86% accuracy on the prediction task.
• Field outcome: Personalization campaign (test-and-learn) based on model output produced a 6.55% higher take-up rate and IDR 141.6 million in additional revenue during the campaign period.
• Key behavioral drivers: monthly revenue from data packages, frequency of data usage in specific price bands, and total monthly data revenue (behavioral transactional variables dominated).
• Validation & process: Machine learning predictions were supplemented/triangulated with Focus Group Discussions (FGDs) with subject-matter experts.
• Limitations: single firm (Telkomsel), short observation window (3 months), and reliance on one ML algorithm (CatBoost) — limits generalizability and long-run inference.

Data & Methods

• Data
  • Source: Telkomsel customer transaction logs and behavioral records (integrated via DMP / MyTelkomsel).
  • Size & period: ~5.7 million customer data points, January–March 2023.
  • Key features used (reported): monthly data-package revenue, frequency of data use within price ranges, total monthly data revenue; likely additional behavioral/transactional covariates from the DMP.
• Methods
  • Modeling: Supervised classification with CatBoost (Gradient Boosting Decision Tree).
  • Objective variable: purchasing propensity / identifying potential lapsers (binary classification for targeting).
  • Evaluation: Model accuracy reported at 86%; business evaluation via a test-and-learn campaign measuring take-up rate and revenue uplift.
  • Complementary methods: Focus Group Discussions with experts to interpret drivers and support implementation decisions.
• Field experiment / deployment
  • Personalized offers were delivered through the MyTelkomsel app as part of the CVM program.
  • Impact assessed by comparing take-up and revenue between personalized (treatment) and control treatment groups (test-and-learn framework).
• Reporting gaps / methodological notes
  • Paper reports accuracy and business uplift but does not detail other ML metrics (AUC, precision/recall, calibration), feature engineering specifics, treatment assignment protocol, or costs of model development/deployment.
  • Short follow-up period — longer-term retention/LTV effects not observed.

Implications for AI Economics

• Firm-level productivity and ROI
  • Evidence that ML-driven personalization can materially raise conversion rates (+6.55%) and produce measurable short-term revenue gains (IDR 141.6m in this campaign). For firms with large customer bases, small percentage improvements can scale into substantial revenues.
  • Important to weigh incremental revenue against development, data infrastructure, and campaign delivery costs to assess ROI; the paper reports uplift but not net ROI.
• Targeting efficiency and resource allocation
  • Behavioral transaction variables (spend and usage frequency) are powerful predictors, implying firms can improve targeting efficiency by prioritizing high-signal behavioral features rather than only demographics.
  • CVM enables finer segmentation and potentially lower marketing waste via more relevant offers.
• Competition, pricing, and market dynamics
  • Widespread adoption of ML-based CVM can intensify competition on personalized offers and dynamic pricing; firms that deploy it effectively may secure a data-driven competitive advantage.
  • May induce heterogeneous pricing and offer strategies across consumers, with distributional implications (e.g., personalized discounts vs. price discrimination).
• Consumer welfare and regulatory considerations
  • Personalization improves product fit, but raises issues around transparency, privacy, and fairness. Regulators and firms should consider consent, data governance, and potential segmentation-related exclusionary effects.
• Research design and causal inference
  • The test-and-learn field assessment is the correct applied-economics approach to identify causal impacts of personalization. Future work should report randomized assignment details, longer horizons (LTV), and heterogeneous treatment effects.
  • Researchers should complement predictive performance (accuracy) with causal estimands (ATE, uplift, retention effects) and cost-benefit analyses.
• Generalizability and future research directions
  • Single-firm, short-horizon evidence is promising but not definitive. Replication across firms, longer panels, and robustness to algorithm choice (compare GBDT vs. neural nets, uplift models, causal forests) are needed to understand external validity and long-run impacts.
  • Combining predictive models with causal uplift/targeting models could further optimize marketing spend (target those with highest incremental response rather than highest propensity).
• Policy and labor implications
  • Automation of segmentation and campaign management may change marketing labor needs (more analytics roles, fewer manual segmentation tasks).
  • Anticipate shift toward continuous model monitoring, fairness auditing, and consumer-protection compliance as ML personalization scales in telecoms and other industries.