Case study: AI gateway for a Swiss software company

Note: this case study is based on a real engagement with anonymised metrics. The customer reviewed and approved publication.

Context

A 60-person Swiss software company had grown its AI usage across three product teams. Spend was growing faster than revenue. Each team had its own provider key, its own logging, and its own retry behaviour. Nobody knew what AI cost per feature was.

What we did

1. Mapped the existing AI usage across teams, providers, and tasks.

2. Deployed a LiteLLM gateway behind the existing application stack with budgets, fallbacks, and per-team cost attribution.

3. Enabled provider caching for the highest-volume routes.

4. Moved extraction and classification tasks to cheaper models with quality gates.

Outcome

• 38% reduction in monthly AI spend within six weeks.
• Per-feature cost attribution now available to engineering managers.
• AI usage grew 22% in the same period, mostly because teams stopped being afraid of cost.
• One runaway agent detected and stopped by a circuit breaker before it became an invoice.

What we would do differently next time

We would have started with circuit breakers before caching. The runaway agent happened before the cache was deployed, and the manual intervention cost more time than the caching saved.