How a request is protected, from sign-in to answer.

1 Sign-in.

People sign in through your identity provider with single sign-on and multi-factor authentication. Each request carries the user's role and tenant.

2 Policy.

The company AI policy is read for the request. If it cannot be read, the request stops and the user is told why.

3 Input checks.

A classifier checks for prompt injection. Personal data is replaced with labels before any model receives it, including names written in Indian scripts.

4 Data access.

The agent reaches only the data and tools the user is allowed to reach. Tenant isolation is enforced at the data layer.

5 Model call.

Only approved models are used: open-weight models on your hardware, or providers under zero-retention terms. Routine steps go to smaller models.

6 Output and approval.

Guardrails check the answer. Where a step needs a decision, the workflow waits for a named approver.

7 Record.

The step is recorded with its prompt, the context it used and the person who asked. The record is signed and can be exported.

Encryption. Data is encrypted in transit and at rest, with keys held per tenant.