Language models are useful precisely where data runs out: they carry common sense about how people reason, object and justify decisions. Early on, that lets a simulation be useful before the domain has produced enough evidence of its own.
The risk is obvious. So the generative layer is constrained by a probabilistic skeleton — logistic transitions between funnel states, elasticity to price, information and brand perception, plus behavioural cloning on real conversations. If a segment historically never reacts the way the model finds linguistically plausible, the empirical evidence wins. Over time the balance shifts from cognitive inference to observed behaviour.