Skip to main content
Every assistant save starts a background compliance scan through the platform AI gateway. The scan evaluates the system prompt, greeting, flow instructions, call-screening identity, and any audio greeting transcript as untrusted content.

​
When an assistant is blocked

An assistant is blocked only for concrete fraud or abuse risk: third-party impersonation (bank, government, unrelated brand), caller-ID spoofing, credential theft, or similar scam intent. Greeting drift, name mismatches between prompt and greeting, missing AI disclosure, language switches, and ordinary business personas are not blocking reasons. Blocked assistants cannot answer or start:
  • phone or WhatsApp voice calls;
  • web calls and web chat;
  • automated WhatsApp or other messaging replies;
  • automated email replies;
  • requests that start calls through API v1 or MCP.
The assistant editor shows the score, flags, and reason. Its previous active state is remembered.

​
Fix and rescan

Edit the flagged prompt or greeting and save the assistant. A new scan starts immediately; a one-minute safety sweep also picks up queued scans. If the new snapshot is clean, the assistant is unblocked automatically and returns to its previous active state. Old scan results never apply to newer content. Editing a declined assistant creates a new decision snapshot, so it can be checked again.

​
Request manual review

From the assistant editor, choose Request manual review and explain why the current content should be approved. A platform administrator receives an email and can see the exact prompt snapshot, evidence, score, scan history, and your explanation. You can also use:
  • GET /assistants/{id}/compliance-review
  • POST /assistants/{id}/compliance-review with { "reason": "..." }
  • MCP tools get_assistant_compliance_review and request_assistant_compliance_review
Manual approval unblocks the exact reviewed snapshot. If the assistant changes before the decision, the stale decision is rejected and the latest scan remains authoritative.