The economics of AI systems

The system- and business-level view of AI cost, building on *Inference cost*'s per-call mechanics rather than repeating them: the full cost ledger beyond the token bill, why cost per successful task is the unit that actually matters, how to price a full workflow and an agent loop's long tail, how quality and cost trade off through eval-priced model tiers and routing, how prompt length, caching, batching, and self-hosting act as cost levers at scale, how budgets and quotas bound both spend and blast radius, how to forecast cost as volume grows, how to monitor unit economics in production, and the gate that decides whether an AI feature should ship at all.