We spent a long stretch encouraging engineers to use AI development tools, and usage eventually grew fast enough that cost became a leadership problem. Although a spending cap looked like the obvious answer, the underlying arithmetic showed that a limit could easily solve the wrong problem.
This is a candid case study of designing AI cost guardrails while adoption was still evolving. I'll explain why we tried visibility and efficiency measures before hard limits, how the arithmetic showed that a daily cap was a circuit-breaker rather than a budget control, and how we designed a tiered exception model. The most important decision is whether a cap is meant to bound total spend, catch runaway usage, or address both.
This talk has been presented at TechLead Conf London 2026: Adopting AI in Orgs Edition, check out the latest edition of this Tech Conference.



















