Why LLMs Are Not Enough for Community Banks: The Case for a Private Large Quantitative Model

Large language models are effective at reading, extracting and drafting banking documents, but community bank decisions run on numbers. Drawing on public research (Apple on quantitative reasoning fragility, Thinking Machines Lab on nondeterminism, NIST on confabulation, explainability and privacy), this paper describes five gaps that matter to risk committees and examiners. It proposes pairing language models with deterministic code and a private large quantitative model (LQM) trained on an institution's own records, designed for repeatable, traceable outputs that state confidence or report insufficient data, with bank-owned encryption keys, a governed model gateway and human decision authority. It notes that revised interagency model risk guidance defines a model as a quantitative method and places generative AI outside its scope, so an LQM can be governed with established validation discipline. Status labels (Live, Built, Running, Specified) accompany each claim; the paper does not report accuracy or validation results.