Eva Ascarza

We develop a framework for learning semantic customer embeddings that combines the semantic knowledge encoded in large language models with behavioral patterns observed in proprietary customer data. The approach integrates sparse, heterogeneous signals—such as purchases and location visits—into unified customer representations that generalize across marketing applications. Using data from a large customer engagement platform, we show that the embeddings support customer segmentation and persona development, natural-language look-alike discovery, targeting for newly introduced products, and promotional targeting.

We develop a capacity-aware targeting framework that accounts for congestion spillovers created when marketing interventions increase demand in capacity-constrained service settings. Applied to ride-hailing data, the approach shows that conventional targeting systematically over-targets customers—sometimes selecting customers whose congestion externalities outweigh their individual benefits—and substantially improves targeting profitability.

We develop a hierarchical Bayesian model that jointly analyzes many randomized experiments to decompose heterogeneity into customer responsiveness, campaign design, and timing. Using large-scale field data, we show that unobserved customer heterogeneity dominates and that repeated exposure leads to intervention fatigue.

We formalize quitting decisions in sales calls as an optimal stopping problem and develop a generative LLM-based stopping agent trained via imitation learning. Applied to outbound sales calls, the agent reallocates time away from failed calls and increases expected sales by up to 37%.

We introduce Multi-Response State Representation learning for offline reinforcement learning. By jointly predicting multiple behavioral signals, the method improves state representations and substantially increases long-term value in a mobile gaming application.

We develop a two-stage causal ML framework that pools hundreds of past experiments to estimate a conditional incrementality function over customers and interventions. Applied to millions of customers, the approach substantially improves targeting and generalizes to new interventions.

We study how differential privacy distorts CATE estimation and targeting. We propose an honest calibration approach using doubly robust scores and sample splitting to improve accuracy without violating privacy constraints. The method is model-agnostic and DP-compatible.

Studying privacy regulations in California and Virginia, we show that stronger privacy protections increase consumer data sharing on a digital platform and in national surveys. Effects are strongest among users previously least willing to share, suggesting increased trust rather than heightened concern.

Standard experiments sample uniformly, ignoring decision objectives. We propose policy-aware experimentation using expected profit loss to prioritize sampling customers who matter most for targeting decisions. The approach improves profitability while dramatically reducing sample size requirements.

We introduce protected heterogeneity as the share of variation in an algorithmic score explained by protected attributes. We develop a variance-based diagnostic, (R^2_{prot}), that is threshold-invariant and interpretable. We further propose residualized scores that preserve individual differentiation while guaranteeing fairness.

Personalization has become the heartbeat of modern marketing. The rapid expansion of individual-level data, the proliferation of personalized communication channels, and advances in experimentation have fundamentally reshaped how firms tailor their marketing strategies. Furthermore, causal inference and machine learning enable firms to understand how the same marketing action can impact individual customers differently. This article formalizes personalization as a causal inference problem embedded in a test-and-learn framework. We review key challenges and solutions related to data limitations, treatment effect heterogeneity, policy evaluation, and ethical considerations, and identify emerging research trends such as generic and double machine learning, direct policy learning, foundation models, and generative AI.

Using a large randomized control trial in a free-to-play mobile game, we study dynamic difficulty adjustment. While lower difficulty reduces short-term purchases, it increases engagement and long-term retention, leading to higher overall monetization. We document substantial heterogeneity and derive implications for personalized product design.

We propose a probabilistic model that fuses click-stream and purchase data within and across customer journeys to better leverage first-party data in infrequent purchase settings. Using a Bayesian nonparametric Pitman–Yor process to capture context heterogeneity, the model addresses the cold start problem and improves personalization and retargeting performance.

Firms often target customers using CATEs estimated on long-term outcomes, but such outcomes accumulate noise and unobserved heterogeneity. We show that this leads to ineffective targeting. We propose using short-term outcomes and surrogate indices, including a separate imputation strategy for churn and purchase intensity, to improve long-run targeting performance.

We propose a Bayesian nonparametric model to detect individual-level routines—repeated behaviors with temporal structure. Applied to ridesharing data, we show that routineness predicts higher future usage, greater resilience to service failures, and heterogeneous responses to pricing and service disruptions.

An inherent risk of algorithmic personalization is disproportionate targeting of protected groups. Removing protected attributes does not eliminate bias due to correlated proxies. We propose BEAT, a bias-eliminating adapted trees approach that ensures balanced allocation while preserving personalization value. Using simulations and an online experiment, we demonstrate both group and individual fairness.

We propose a probabilistic machine learning approach to overcome the cold start problem in CRM by efficiently learning about new customers. Our method combines collaborative filtering with causal inference to estimate individual treatment effects with limited data.

We identify common pitfalls in applying AI to marketing and provide practical guidance for improving marketing AI effectiveness.

Companies often target customers with the highest risk of churning, assuming they are the best candidates for retention. Combining field experiments with machine learning, we show that high-risk customers are not necessarily the most responsive to retention interventions. Instead, targeting based on heterogeneity in treatment response significantly improves retention effectiveness.

In today's turbulent business environment, customer retention presents a significant challenge for many service companies. Although a large body of research focuses on predicting customer churn, several equally important aspects of retention management have received less attention. We draw on prior research and current practice to provide insights into managing retention and to identify areas for future research. We advocate a broad perspective on retention that goes beyond a binary retain/churn view, propose a flexible definition of retention, discuss alternative metrics for measuring retention, and present an integrated framework for managing retention that leverages new data sources and methodologies such as machine learning. We identify trade-offs between reactive and proactive retention, between short- and long-term remedies, and between discrete campaigns and continuous processes, and outline a research agenda to better align academic work with managerial needs.

We study hybrid settings in which firms face both observed (overt) and unobserved (silent) churn. We develop a hidden Markov model that separates the two churn types. Applying the model to a daily deals site and a performing arts organization, we find that overt churners are often highly active before leaving, while silent churners disengage early and rarely respond to firm communications.

CRM campaigns traditionally focus on targeted customers. Using a randomized field experiment in a mobile telecom setting, we show that such campaigns generate spillover effects through social networks, increasing usage and reducing churn among non-targeted but connected customers. We estimate a social multiplier of 1.28 and show that spillovers are driven by increased communication between targeted customers and their connections.

Facing increasing customer churn, many service firms have begun recommending pricing plans to their customers. Using a large-scale field experiment, we examine the effectiveness of such proactive retention campaigns. We find that encouraging customers to switch to cost-minimizing plans can increase rather than decrease churn. We propose two mechanisms—reduced inertia and increased salience of past usage—and provide evidence supporting both explanations. We further assess the revenue implications and derive recommendations for targeting such campaigns.

As firms become more customer-centric, concepts such as customer equity come to the fore. Any serious attempt to quantify customer equity requires modeling techniques that can provide accurate multiperiod forecasts of customer behavior. Although a number of researchers have explored the problem of modeling customer churn in contractual settings, there is surprisingly limited research on the modeling of usage while under contract. The present work contributes to the existing literature by developing an integrated model of usage and retention in contractual settings. The proposed method fully leverages the interdependencies between these two behaviors even when they occur on different time scales. We propose a model in which usage and renewal are modeled simultaneously by assuming that both behaviors reflect a common latent variable that evolves over time. We capture the dynamics in the latent variable using a hidden Markov model with a heterogeneous transition matrix and allow for unobserved heterogeneity in the associated usage process.

In many service industries, firms introduce three-part tariffs to replace or complement existing two-part tariffs. Behavioral research suggests that the attributes of a pricing plan may affect behavior beyond their direct cost implications. Evidence suggests that customers value free units above and beyond what might be expected from the change in their budget constraint. We examine a market in which three-part tariffs were introduced for the first time and analyze tariff choice and usage behavior for customers who switch from two-part to three-part tariffs. We show that switchers significantly "overuse" relative to what would be predicted from the change in the budget constraint alone. We develop a joint discrete–continuous model of tariff choice and usage that allows for a higher valuation of free units. The results indicate that the majority of three-part-tariff users value minutes under a three-part tariff more than under a two-part tariff, generating substantial revenue implications for the firm.