Per-team AI budgets that actually work
Unlimited model access is easy to grant and hard to walk back. How to attribute AI spend per team and set limits that do not block work.
AI spend behaves differently from cloud spend: it scales with usage in near real time, and a single new feature can multiply token volume overnight. Monthly invoice review is too slow a control loop.
Attribution first
You cannot budget what you cannot attribute. Route model traffic through a gateway that stamps every request with team, application, and environment, then report cost per team daily.
Set soft limits before hard limits
- Warn at 70% of a team's monthly allowance.
- Throttle non-interactive workloads before user-facing ones.
- Reserve hard cut-offs for runaway automation, never for a human in a support queue.
Make the cheaper path the default
Most quality loss comes from the wrong default, not the cheaper model. Route routine classification and extraction to small models, escalate to frontier models on low confidence, and cache repeat prompts.
Report unit economics, not totals
Cost per resolved ticket or per generated report tells a product team something actionable. Total monthly spend only tells them to feel bad.
The governance angle
Budgets and access control are the same control plane. Once every request is attributed and policy-checked, "who spent this" and "who was allowed to do this" become the same question — which is exactly what an auditor wants to see.