As AI‑generated content becomes commoditized, economic value shifts from final artifacts to the upstream expertise that steers LLM generation. Organizations lack methods to value this expertise and to pinpoint which individual skills or configurations causally improve performance.
This study focuses on generative marketing, where structured expert labels act as skill primitives conditioning LLM‑generated advertisements. Estimating their causal effects is challenging because outcomes are mediated by stochastic LLM behavior, involve high‑dimensional label interactions, and suffer from multimodal confounding.
We propose D‑VAE (Deconfounder Variational Autoencoder), a counterfactual‑inference framework that integrates a multi‑treatment deconfounder with causal representation learning. D‑VAE recovers causally disentangled labeling logic and derives a PAC‑based bound on omitted‑variable bias from single‑treatment confounders $$\epsilon = O\left(\sqrt{\frac{\log(1/\delta)}{n}}\right)$$.
Applied to 29,177 foundation‑makeup ad creatives, D‑VAE outperforms state‑of‑the‑art benchmarks. A field experiment shows that the recommended label configurations increase engagement by 32.7% and are statistically indistinguishable from human experts. A simulation study further confirms causal attribution under complex confounding and treatment interactions.
Overall, the work shifts value assessment from AI‑generated products to the underlying skills, providing a foundation for pricing and optimizing AI‑augmented skill portfolios.
Review