Most organizations have an AI cost policy. Far fewer can tell whether it’s working. In this session, the Harness team shows what happened when we stopped treating our AI cost policy as a document and ran it as a live, enforced limit on our own AI tooling bill.
We’ll share why we built AI Budgets, how we set it up, what worked, what pushed back, and what we’d do differently. You’ll see the results: average weekly AI spend per developer fell from $184 to $115, and cost per million tokens dropped about 55% the week governance went live. Then we’ll walk through how weekly budgets, graduated alerts, and self-service approval requests keep developers moving while giving FinOps and engineering leadership a ceiling everyone can see.
You’ll leave with a practical playbook for putting guardrails on AI spend in your own organization.
Key takeaways
- A policy you can’t see working isn’t a control. Only 13% of organizations can tell whether their AI cost policy is holding. Attendees learn how to close that gap.
- Visibility comes before enforcement. Attributing spend by user, provider, model, and token type first is what makes a budget credible and fair.
- Cost per million tokens tells you more than total spend. It measures how well AI was used, not just how much, and it dropped about 55% in our own deployment.
- Guardrails enable adoption instead of blocking it. A warn-then-hold structure with self-service requests means no one has to be the “AI police,” and engineering can say yes to more AI.
- Cost governance and delivery governance can share one engine. Budgets can become pipeline gates, so cost problems get caught before they ship.
Who should attend
FinOps practitioners, engineering and platform leaders, and finance partners who own or are being asked about AI spend. This is the same audience the blog’s “Who this is for” section already speaks to.
We’ll share why we built AI Budgets, how we set it up, what worked, what pushed back, and what we’d do differently.