I built an AI system to do a job I was tired of doing. It worked and grew. Understanding what it cost became another engineering problem.

This talk follows the mistakes behind an expensive agent: using a powerful model for every task, carrying unnecessary context, asking for long answers, and missing opportunities to reuse stable input. We work through the decisions that affect the bill, from measuring cost per task to choosing models, reducing unnecessary tokens and making prompt caching useful.

We also look at the checks needed to keep quality from slipping and to notice unexpected spending. The examples come from building an autonomous system. The aim is to give developers a practical way to examine their own costs and decide which changes are worth making.

For engineers, architects and technical leaders building agentic systems. The session connects an increasingly practical concern, the AI bill, to decisions the audience can influence: model choice, context, output length, caching and spending checks. A personal sequence of mistakes gives each technique a reason to matter. Attendees leave with an order for examining their own costs and checking that savings preserve quality. A fit for AI engineering, platform engineering and FinOps programmes.