> ## Documentation Index
> Fetch the complete documentation index at: https://docs.bland.ai/llms.txt
> Use this file to discover all available pages before exploring further.

# Guard Rails

> Automated monitoring that runs continuously on every AI response to catch critical failures before they reach your customers.

Guard rails continuously analyze every AI and user response during production calls. They catch severe failure modes like discrimination, self-harm encouragement, or compliance violations before they cause damage.

## Why Guard Rails Matter

When running thousands of AI calls, edge cases aren't a question of *if* but *when*. Manual review doesn't scale. If you're running 50,000 calls a week, you can't have humans listening to every conversation in real time.

**Legal exposure adds up fast:**

* A single discriminatory statement during a hiring call can trigger an EEOC complaint
* TCPA violations carry statutory damages of \$500 to \$1,500 per call
* Class action lawyers actively monitor for these violations

**Brand damage is hard to recover from:**

* A call where your AI mishandles someone in distress can go viral
* These failures often happen in edge cases your testing didn't cover

Guard rails also help with stakeholder management. Your legal team gets confidence that even in scenarios nobody anticipated during testing, there's still a safety net catching dangerous outputs.

### Two Types of Guard Rails

> [accordion] Require Guard Rails (Time-based)
>
> The guard rail **must** be triggered within a specified time window, otherwise the configured action runs.
>
> Used for TCPA disclosures that must happen early in the call:
>
> * **AI disclosure** - Must disclose the caller is speaking with an AI
> * **Self introduction** - Must identify who is calling
> * **Recording disclosure** - Must disclose the call is being recorded
>
> The default time window is 30 seconds. You can change it per guard rail.

> [accordion] Forbid Guard Rails (Continuous)
>
> The conversation content must **not** trigger the guard rail. These continuously monitor the conversation.
>
> Used for:
>
> * TCPA opt-out handling
> * Custom guard rails
> * Content policy violations

## Apply guard rails

You can apply guard rails in two places:

1. **An agent**: open the agent and select **Guardrails**, the second tab at the top of the builder, beside **Agent**. See [Agent builder](https://docs.bland.ai/agents/agent-builder.md#guardrails). To set them through the API, include `settings.guardrails` when you [create an agent version](https://docs.bland.ai/api-v2/post/agents-id-versions.md).
2. **Any outbound call via API**: include the `guard_rails` field in the POST `/v1/calls` [API](https://docs.bland.ai/api-v1/post/calls.md#param-guard-rails).

On an agent, four built-in guard rails cover the standard TCPA disclosures: AI disclosure, recording disclosure, self introduction, and opt out. Each is a switch, and a new agent starts with all four off. Built-in guard rails are available on every plan.

Guard rails are part of the agent's configuration, so a change reaches live calls once you publish and promote the agent. See [Environments](https://docs.bland.ai/agents/environments.md).

## Actions When Guard Rails Fire

You can attach an action to run when a guard rail is tripped:

* **End call**: Immediately terminate the conversation
* **Transfer**: Route the call to a phone number, such as a live agent
* **Move to a step**: Jump to a specific step in your agent, such as an apology or handoff step

An action is optional. With no action attached, the violation is recorded and the call continues.

## Custom Guard Rails

Custom guard rails cover industry-specific concerns or company policies. They require an enterprise plan, and each agent can have up to 5. You write them on the agent's **Guardrails** tab.

When you write a custom guard rail, make the prompt clear, specific, and well-defined. Ambiguous prompts can cause false positives, flagging calls that shouldn't be triggered.

> [!NOTE]
>
> Guard rails are monitoring. They do not guarantee regulatory compliance, and they are not a substitute for your own legal review of how you use the product.

***

Docs for agents: [llms.txt](https://docs.bland.ai/llms.txt)


This documentation is built and hosted on [Mintlify](https://mintlify.com), a developer documentation platform.