Predicting demand for new products is an important and challenging problem in marketing, especially for product categories where brand is a key driver of choice. In these settings, observed product attributes do not explain choice patterns well, which makes predicting sales of a new product as a function of marketing mix variables intractable. To address this problem, I develop a scalable framework that enriches structural demand models with large language models (LLMs) to predict consumer preferences for new brands. After estimating preferences for existing brands using a structural model, I use an LLM to learn a mapping from text descriptions of the brand and consumer to these estimated preferences. When fine-tuned in this way, I show that the tuned LLM is able to generalize to previously unseen brands that were excluded from the training sample. In contrast, conventional models based on text embeddings return predictions with zero or even negative correlation with the actual utilities. My fine-tuned LLM achieves the first informative predictions of consumer preferences of new brands from raw text, with a 25% smaller mean squared error and correlation between predictions and held out preferences of 0.34. Additionally, I combine causal estimates of the price effect from instrumental variables methods with the LLM predictions to enable pricing-related counterfactuals. More broadly, this approach illustrates how new kinds of questions can be answered by using the capabilities of modern LLMs to systematically combine the richness of qualitative data with the precision of quantitative data.

We develop a framework for improving the effectiveness of marketing communications by aligning a language model to the causal signal in A/B tests. The core idea is that randomized experiments reveal not just which content performs best, but directions of improvement that can serve as supervision for a generative model. Using within-experiment comparisons, we fine-tune a language model to create an improver architecture that transforms lower-performing content into higher-performing content. The resulting model can be composed with any baseline decision-maker, e.g. a human expert, to suggest improvements to new content. We validate our framework through a large-scale field experiment in email marketing spanning 36 new campaigns and 283 million impressions. Subject lines improved by our aligned model achieve click-through rates 33% higher than those from unassisted human experts, while a general-purpose language model with 30 times as many parameters as our fine-tuned model does not improve outcomes. This suggests that alignment to experimental data adds value beyond model scale alone and remains relevant even as general-purpose models become more capable. Our findings provide the first field evidence that aligning language models to A/B tests improves outcomes in real marketing campaigns and show that integrating AI with experimentation can train systems that make better future decisions.

Do human differences persist and scale when decisions are delegated to AI agents? We study an experimental marketplace in which individuals author instructions for buyer-and seller-side agents that negotiate on their behalf. We compare these AI agentic interactions to standard human-to-human negotiations in the same setting. First, contrary to predictions of more homogenous outcomes, agentic interactions lead to, if anything, greater dispersion in outcomes compared to human-mediated interactions. Second, crossing agents across counterparties reveals systematic dispersion in outcomes that tracks the identity and characteristics of the human creators; who designs the agent matters as much as, and often more than, shared information or code. Canonical behavioral frictions reappear in agentic form: personality traits shape agent behavior and selection on principal characteristics yields sorting. Despite AI agents not having access to the human principal's characteristics, demographics such as gender and personality variables have substantial explanatory power for outcomes, in ways that are sometimes reversed from human-to-human interactions. Moreover, we uncover significant variation in "machine fluency" - the ability to instruct an AI agent to effectively align with one's objective function - that is predicted by principals' individual types, suggesting a new source of heterogeneity and inequality in economic outcomes. These results indicate that the agentic economy inherits, transforms, and may even amplify, human heterogeneity. Finally, we highlight a new type of information asymmetry in principal-agent relationships and the potential for specification hazard, and discuss broader implications for welfare, inequality, and market power in economies increasingly transacted through machines shaped by human intent.

Many business scenarios require a decision maker to balance multiple competing objectives. Typical decision support algorithms for such scenarios require calibrating the relative weights on the objectives, but this is brittle and sensitive to misspecification. As a result, suggestions from learned policies demonstrate pathologies, appear unrealistic, and are rejected by decision makers. An alternative approach is to imitate the decisions of managers (i.e. the mapping from state to action) in the historical data, but this replicates the mistakes that managers made in the past. We develop a new approach that imputes state values that make the past decisions of managers optimal. Policies trained on this modified data perform well when deployed, which we demonstrate in an application to scheduling shift workers to match forecasted demand in fast food and retail stores. More broadly, this method is applicable to reinforcement learning problems with underspecified reward, where policy learning is infeasible because the mapping from the state to the optimal action is not uniquely identified. In many of these settings, including our case with multiple weighted rewards, inferring the state that would have made an action optimal is easier than explicitly learning the reward and allows us to learn a superior policy compared to imitation learning.

Consumers' choice behavior is driven by measurable, objective attributes of products as well as unmeasured attributes (perceived convenience, quality, sportiness, etc.). Past approaches of modeling the effect of unmeasured attributes on decisions have entailed eliciting some components of a Fishbein multiattribute utility model from consumers and inferring the remaining components from observed choices. These approaches have either elicited perceptual ratings on each product in the choice set along some pre-specified dimensions and inferred the importance of each dimension, or have elicited ratings of the importance of each dimension and inferred the perceived level of each dimension. The former approach doesn't scale well to many products, as it requires rating every product on every dimension, and the latter approach still relies on an analyst selecting the correct dimensions to rate. Notably, it will fail if consumer perceptions are high-dimensional. I propose an alternative method that merges choice data (e.g. from conjoint) with responses to open-ended questions on consumers' needs (e.g. from interviews or focus groups). I use these stated needs to identify consumer attitudes in the Fishbein model, which implies that the estimated perceptions are the extent to which needs are satisfied. I show how to apply causal abstraction, a recently developed method from the NLP literature, to interpretably estimate the extent to which each product satisfies these needs. This result shows how to operationalize the vast literature on laddering by integrating it into models of choice in an interpretable way. It is also informative for decisions on changing product attributes to better meet consumer needs or for crafting marketing content strategies that emphasize these needs.