Security & Approval Design — Rules Enforced in Code
The AI is capable, but it operates inside guardrails it cannot cross: tiered approval gates, hard ad-spend caps, KMS-encrypted credentials, and a full audit trail — all enforced in code, not AI judgment.
- Approval gates: negative-review replies, ad budget changes, and pricing moves require owner approval in Telegram by default; in managed service, the owner may explicitly whitelist specific actions, including negative-review replies
- Spend caps: ad spend limits are enforced by the system; the AI has no code path to exceed them
- Review-reply checks: guidance asks the AI to avoid admissions of liability, compensation promises, naming staff, and arguments. A person should check the exact draft and facts before approving a public reply, or review the proposed scope before enabling auto-publication; the free demo does not programmatically enforce these content rules
- Audit trail: every action, approval, and failure is logged; failures alert with screenshots, never silently