Adam Smith Meets AI: When to Buy a Decision Model Instead of Building One
To Make, or to Buy: The Economics of Decision Models
Decision models like Jev output fixed choices, scores, or probabilities, and their output tokens are nearly free, so a batch of 400,000 queries costs about $4.20 versus $13 on GPT-5 nano. That economics creates a "decision hierarchy"—rules and caches as L1, decision models as RAM, LLMs as disk, humans as tape—with Chow's 1970 rule setting the confidence threshold where escalating to the next level beats answering. When errors are expensive, accuracy decides; price barely matters.
In our decision hierarchy, the LLM is selling one product - fewer mistakes. Each level is a supplier with 3 numbers: a price, an error rate, and the cost of each error it makes.