AI safety & conduct
AI safety & conduct is where you set the content-moderation guardrails that apply to every agent in your project - conversational and guided alike. These filters inspect what users send in and what your bot sends back, blocking unsafe content before it reaches a person. Because Configuration is project-level, whatever you enable here protects the whole bot, not a single agent.
This page is part of Configuration. Settings here apply to all your agents and guided agents. For an overview of how Configuration fits together, see Configuration and the Harness overview.
Note: The AI safety & conduct sub-page only appears if it's enabled for your account. If you don't see it under Others in the left menu, your project doesn't have safety filters provisioned - talk to your account team.
Step 1: Open the AI safety & conduct page
In your bot, go to AI Agent → Configuration and click AI safety & conduct (under Others in the left menu).
The page splits filters into a foundational, always-on layer and an advanced layer you toggle per check:
- Standard content safety - a foundational, always-on content-moderation check. It runs by default and can't be turned off.
- Customer query filters (advanced) - individual checks on what the user sends in: toggle Banned Topics, Violence, Sexual content, and similar to block unsafe inputs before the bot acts on them.
- AI response filters (advanced) - individual checks on what the bot sends back: Toxicity, Bias, and Sensitivity (sensitive-data leakage) in the generated reply.
Step 2: Choose your advanced filters
Standard content safety is on for everyone - you don't configure it. Decide which advanced checks you need on top of it:
| Layer | Filter | Catches |
|---|---|---|
| Customer query filters | Banned Topics | User messages about topics you've ruled out of scope. |
| Customer query filters | Violence | Violent or harmful user inputs. |
| Customer query filters | Sexual content | Sexually explicit user inputs. |
| AI response filters | Toxicity | Hostile, abusive, or offensive wording in the bot's reply. |
| AI response filters | Bias | Biased or discriminatory framing in the bot's reply. |
| AI response filters | Sensitivity | Sensitive-data leakage in the bot's reply. |
Turn on a filter by toggling it. Query filters guard the input path; response filters guard the output path - most regulated use-cases want coverage on both sides.
Step 3: Mind the latency trade-off
Every advanced filter you enable adds a moderation pass, and each pass costs time.
Note: The page warns: "Enabling filters adds latency. For time-sensitive use-cases like VoiceAI agents, consider disabling them." For voice channels, weigh filter coverage against your turn-latency budget - a moderation pass that's invisible on chat can become an audible pause on a call.
Tip - voice vs chat. If the same project serves both chat and voice, decide which layers are non-negotiable for compliance and keep only those on for voice. You can run a fuller filter set on chat where the extra latency doesn't hurt the experience.
Best practices
- Turn safety filters on before you ship, not after a customer hits one. A filter that blocks unsafe content in production is worth far more than one you added in a post-incident review.
- Test adversarial prompts in the Testing Lab before each release. Send the kinds of inputs you most want blocked (banned topics, hostile language, attempts to extract sensitive data) and confirm the filter catches them.
- Pair filters with safe landings. For regulated industries (finance, healthcare), back these checks with legal-approved fallback messages so a blocked turn has somewhere safe to land instead of a generic error.
- Cover both directions for high-stakes bots. Query filters stop unsafe inputs; response filters stop unsafe outputs. For anything handling regulated or personal data, enable both.
- Re-measure latency after each change. Enabling a filter shifts your latency profile - especially on voice. Re-test turn timing whenever you add or remove a check.
Next: Voice settings