Skip to main content

Trust & Governance

The more capable the assistant, the more a client will ask "how do we keep this from going wrong." This chapter is your answer.

Content Filtering & Safety​

ai12z uses a layered security architecture: pre-checks (reCAPTCHA v3, session question limits, rate/IP safeguards), an independent LLM safety layer that can return a customized refusal message across four categories (Self-Harm/Crisis, Jailbreak/Prompt Injection, Active Threat/Violence Emergency, and a general Default/Other Policy Violation category), and admin-configurable alerting. This is already fully documented for this platform — don't duplicate it, link to it:

  • Filtered Content Responses — how to customize each category's response message and set up email alerts (which include the visitor's IP address, conversation ID, and full transcript — treat the reporting mailbox accordingly)
  • The Testing Checklist's Pre-Launch Security section already includes this as a pre-launch checklist item

Logs & Debugging​

Every answer is traceable: logs cover orchestration, queries, webhooks, and actions, with an Analyze capability that helps explain why the assistant answered the way it did — inspecting metadata, directives, sources, timing, and token usage. See Logs for the full reference. This is what turns "the bot said something wrong" from a mystery into a fixable, specific finding — the same content-debugging discipline that matters for any client who wants to trust what's live on their site.

Test Plans​

Rather than manually re-checking answer quality after every content or prompt change, QA Copilot Tests (also referred to as Test Plans) let a team define multi-step test scenarios, run them on a schedule, and get notified automatically if something fails — catching a regression before a visitor does, not after.

Controlled / Compliance-First Answer Mode​

For regulated or compliance-sensitive clients, a direct generated answer isn't always the right default. ai12z supports a more controlled mode: instead of generating a direct answer, presenting the most relevant approved resources and explaining why each is relevant — giving the organization more control over accuracy and risk while still helping the visitor find trusted information fast. The underlying mechanism is tag-based content restriction — see How to Use Tags for Content Filtering in ai12z — combined with careful system-prompt design.

Worked example: Healthcare and Life Sciences. Rather than answering a clinical question directly, the assistant can surface the most relevant approved resources with an explanation of why each is useful. For life sciences specifically, this extends to guiding healthcare professionals to approved drug and device information without relying on uncontrolled model knowledge, detecting and routing adverse events into systems like Veeva or Salesforce, and pushing conversation activity into those systems for follow-up. This is one instance of a general pattern — any client in a regulated space (financial services, legal, government) can apply the same controlled-answer approach.

Last updated on