Authors: Navya Sri Maddukuri, Hari Nagakoteswar Tripurari
Abstract: Digital commerce platforms are transitioning from static, catalog-constrained recommendation systems toward autonomous personalization engines capable of generating individualized offers, marketing messages, and product bundles in real time using large language models and adaptive policy optimization. While this transition promises substantial gains in customer relevance and firm revenue, it raises significant unresolved questions regarding consumer trust, perceived manipulation, autonomy erosion, and algorithmic price discrimination. This study reports a large-scale randomized field experiment (N = 106,600 customers) conducted across a multi-category digital commerce platform, in which customers were randomly assigned to one of five personalization conditions spanning a capability gradient: rule-based recommendations (control), deep-learning recommendations, static-policy generative AI personalization, autonomous adaptive generative AI personalization, and autonomous adaptive generative AI personalization with a transparency layer disclosing personalization rationale and providing opt-out controls. Intention-to-treat estimates reveal that personalization capability is monotonically associated with increased 12-month customer lifetime value (ranging from +$31.40 for deep-learning recommendations to +$79.20 for autonomous adaptive generative personalization, both p < .001 relative to control), but autonomous adaptive personalization simultaneously produces the largest declines in consumer trust (ΔCTI = –0.58, p < .001), the largest increases in perceived manipulation (ΔPMS = +0.91, p < .001), the largest declines in perceived autonomy (ΔPAI = –0.64, p < .001), and a net increase rather than decrease in 90-day churn (+1.84 percentage points, p < .001) — reversing the directional pattern observed for less capable personalization engines. The transparency layer condition recovers approximately 60% of the trust loss and 50% of the manipulation perception increase associated with autonomous adaptive personalization while preserving approximately 90% of its customer lifetime value gain, identifying a Pareto-improving configuration relative to non-transparent autonomous personalization. Heterogeneous treatment effect analyses reveal that price-sensitive and digitally less literate customer segments experience disproportionately higher manipulation perception increases relative to value gains. The paper contributes a Personalization-Manipulation Frontier framework to marketing analytics and information systems governance research, demonstrating that autonomous personalization capability and consumer welfare are not inherently aligned absent deliberate transparency design.
DOI: http://doi.org/10.5281/zenodo.20841354
