This policy is a draft, not yet approved by independent legal counsel.
The Service routes AI requests to approved upstream providers. The Service does not train or host foundation models. Model eligibility is determined per tenant plan.
Inputs are classified by sensitivity. Inputs containing regulated data are routed only to models approved for that classification.
Outputs are filtered through the output guard before delivery. The guard checks for prohibited content, policy violations, and brand-term compliance.
Models must not be used to generate legal advice, medical diagnoses, or financial guarantees. The Service provides AI assistance, not professional advice.
If an upstream model is unavailable, the Service falls back to an approved alternative or returns a controlled error. No silent degradation occurs.
Every AI receipt records the model used, provider, request ID, and response digest. Tenants can audit model usage through the receipt system.