Security at Claro

A single environment flag halts all AI processing across every org instantly. No per-tenant action needed.

The full pipeline runs and logs what would happen, but nothing is sent or executed. Review AI behaviour before going live.

Every ticket is classified with a confidence score. Below the org's configured threshold, it routes to human review and is never guessed.

Spam, explicit escalation requests, and general inquiries are caught immediately after classification. No response is generated.

Customer history is scanned for serial returners, repeated damage claims, and refund abuse before any write action proceeds.

External risk signals (chargebacks, fraud flags) are fetched per customer. Any high or critical flag sends the ticket to human review.

Merchants define custom conditional rules evaluated in priority order before any action is approved.

The order ID referenced in an action is cross-checked against what was extracted from the ticket. Mismatches block execution.

Every monetary action has a configurable auto-approve ceiling. Above it: human approval required. A second ceiling blocks entirely.

Caps on org-wide executions per time window, and per-customer limits within a rolling period. Both block silently if the cap is hit.

The very first time any action executes for an org, a human must approve, regardless of amount or confidence.

Every response is checked for PII, toxicity, blocked patterns, and merchant-defined rules before dispatch. Hard violations block delivery.