Out of Unstructured Data, Atlas! Mapping Strategic Landscapes with Generative AI

Abstract

The innovation literature has demonstrated the value of maps for decision-making. We present a generative AI based approach-Atlas-that constructs spatial and semantically interpretable maps from unstructured text. Leveraging a transformer-based variational autoencoder fine-tuned on patent data, Atlas implements five key mapping desiderata from cartography literature-i) orthogonality, ii) semantic isometry, iii) proportional scaling, iv) representation expressiveness, and v) extensibility for information overlay. We demonstrate the utility of the maps produced by Atlas across three business domains: (1) illustrating how firms shift focus across technologically interpretable dimensions; (2) identifying regions of an innovation space characterized by concentrated IP licensing activity; and (3) highlighting regions of elevated litigation risk as well as efficient, traversable paths around them. Our contribution is a novel and widely applicable operationalization of a regularized latent-space map derived from text data, guided by effective map-design desiderata and yielding a managerially useful, strategy-informative representation. Several decision-support experiments demonstrate that our approach provides superior strategic guidance compared with standard techniques commonly applied to text data.

Keywords: generative AI, managerial decision support, large language models, patents, interpretability, economics of innovation