Most founders track how many queries their AI agent handled last month but almost none know what each one actually cost once retries, tool calls and context bloat are added in, and that gap is where margins quietly go negative. Fully-loaded cost per AI agent query includes input tokens, output tokens, retries, tool calls and context window overhead, not just the base model API price. A single customer running complex multi-step tasks can cost 5 to 15 times more to serve than a simple-query customer on the same flat subscription price. Retries and tool calls are the two costs founders miss most often, and both scale with task complexity rather than with revenue. Founders should track cost per interaction by customer segment weekly, not monthly, because margin collapse usually shows up in the heaviest users weeks before it hits the aggregate number. The article cites Replit's CEO Amjad Masad discussing the company having to rework its Replit Agent pricing after usage-based inference costs on power users outpaced what flat-tier pricing could support, eventually moving toward consumption-based billing. Cost to serve for an AI agent is the full stack of everything a single interaction touches before it returns an answer, comprising input tokens, output tokens, context tokens, retries and tool calls.