From People Data to Decision Intelligence: Gaps in Generative Population Simulation and a Consent-First Architecture

Organisations make consequential decisions — product launches, pricing, policy changes — on behalf of populations they cannot observe directly. Two technologies promise to close this gap: people-data infrastructure, which indexes who individuals are, and generative population simulation, which uses large language model (LLM) agents to predict how people will respond. We argue that neither, alone or combined as currently practised, yields reliable decision intelligence. We identify seven gaps: coverage bias toward online, English-speaking, Western populations; unclear consent and provenance of grounding data; reliance on stated rather than revealed behaviour; static rather than temporal modelling; individuals simulated without social networks; the absence of decision-level evaluation standards; and a missing link between predicted reactions and the decision a firm must take. We propose a four-layer architecture — a consented, predictive people layer; a twin engine; a simulation runtime; and a decision layer — and an evaluation agenda based on back-testing against historical outcomes. We show that because decisions are comparisons, shared model bias and individual noise largely cancel, leaving option-specific cultural bias as the error that matters most. This is a position paper. It reports no new experiments; figures in Section 6 are explicitly assumed planning targets, not findings. AI writing tools assisted in drafting; the author reviewed and takes responsibility for the content.