Workload Economics of Applied AI in Industrial Operations: Why demos are cheap, products are not, and the difference is decided in week one

A generative-AI feature that costs cents in a demonstration can cost orders of magnitude more in production, without any change to the model, the price list or the feature itself. The difference is workload: how often the system calls a model, how much context each call carries, and which model serves each step. All three are architectural decisions taken at design time, and all three multiply rather than add. This note states the workload relationship explicitly, works through an illustrative agent-deployment scenario with its assumptions declared rather than buried, and sets out three design levers — retrieval, agent discipline and model routing — with the costs each one introduces. It closes with four questions that can be answered before a system is built, while answering them is still cheap. The intended reader is the person accountable for the budget rather than the code: an operations, maintenance or transformation manager deciding whether a proposed system is affordable at scale. Worked examples are drawn from industrial maintenance, where the data conditions — inconsistent free text over a governed asset taxonomy — make the retrieval argument concrete. The reasoning is not specific to that domain.