An AI feature that anyone can call is a cost nobody controls. This release makes AI spend a property of the user record: a monthly limit, checked before each call against what the user has already spent this month, and a default model that administrators set once at the level that makes sense.
Quotas
- A limit per user The user record carries a monthly AI limit in euros. A value below zero means no limit.
- Checked before the call The configuration service sums the current month's spend from the AI log for that user and refuses the call with a quota-exceeded error when the limit is reached.
- Usage reported uniformly The OpenAI, Ollama, Vertex AI and Cohere clients were reworked to report their usage through the same service, so the sum is complete whichever provider answered. The Anthropic client gains streaming in the same release.
- Logged at group level User groups gain an AI-logs level, so what is recorded is a policy rather than a per-user choice.
Default models
- Three levels A default model is declared per user, per department or per company, each keyed by provider, model kind and model code.
- Resolved by precedence The user's own default wins, then the department's, then the company's, so a single company-wide choice covers everyone until someone needs a different one.
The spend limit is the control a finance team asks for before an AI feature reaches production. It lives beside the user's other entitlements and is enforced in the same place every call passes through.