# AI guardrail routing and fallback actions

Control what INSIDE does when an AI safety check blocks input or output.

*Release 3.3 · admin, operator, visitor*

**Settings › AI › Accounts** · **Settings › AI › Connections**

## What it is

**Routing Rules** decide what happens when an AI safety check blocks a request or answer. INSIDE can stop work, allow it to continue, trigger a workflow, apply censorship handling or ask the AI to revise its answer. The available action depends on whether the check concerns incoming or outgoing content.

## Who uses it and why

- Administrators agree safety handling with Powerfront.
- Operators need to understand why AI stopped, revised an answer or handed a conversation to a workflow.
- Visitors may receive a revised answer or follow the configured workflow instead of receiving the blocked answer.

## Where to find it

Powerfront manages **Routing Rules** at **Settings › AI › Connections** and safety configuration at **Settings › AI › Accounts**.

## How it works

1. A configured guardrail checks AI input, output or a message being delivered.
2. INSIDE resolves applicable routing rules. Rules on the guardrail connection take precedence over account rules.
3. It looks for a specific policy and context match, then a reason-pattern match, then the catchall. The first matching rule within each group wins.
4. It applies the chosen action.

| Action | Result |
|---|---|
| **Block** | Stops the blocked work or output. |
| **Continue (allow)** | Allows processing to continue despite that block decision. |
| **Trigger Workflow** | Invokes the selected workflow. |
| **Pushback (auto-retry)** | Requests a revised outgoing answer within the supported retry handling. |
| **Censorship** | Applies censorship handling to incoming content. |

For pushback, a failed retry uses the configured **If pushback retry fails** action. That action can block, allow continuation or trigger a workflow.

## Configure it

Ask Powerfront to configure your guardrail and agree the rules for incoming and outgoing content. Specify which policy or context should match and what should happen next.

For **Trigger Workflow**, Powerfront must select an enabled workflow. For **Pushback (auto-retry)**, agree the action to take if the revised answer still cannot be accepted.

The default catchall action is **Block**. Missing or unrecognised actions also resolve to blocking.

### Settings

| Setting | Where | Applies to | Default | Who can change it | What it does |
|---|---|---|---|---|---|
| Routing Rules | Settings › AI › Connections | website | Block for the catchall route | Powerfront, on request | Determines how a guardrail decision is handled, including blocking, allowing, workflow handover, censorship or a revised answer. |
| Policy | Settings › AI › Connections | website | Any | Powerfront, on request | Matches the safety policy reported by the guardrail. |
| Context | Settings › AI › Connections | website | Any | Powerfront, on request | Narrows the rule to a context within the selected policy. |
| Direction | Settings › AI › Connections | website | Both | Powerfront, on request | Selects Inbound, Outbound or Both for the rule. |
| Action | Settings › AI › Connections | website | Block | Powerfront, on request | Selects what INSIDE does when the rule matches. |
| If pushback retry fails | Settings › AI › Connections | website | Block | Powerfront, on request | Selects Block, Continue (allow) or Trigger Workflow when the revised answer cannot be accepted. |

## Limits and good to know

- **Pushback (auto-retry)** is an outgoing-content option; **Censorship** is an incoming-content option.
- Rules can apply to **Inbound**, **Outbound** or **Both** directions.
- Brief AI progress or holding messages are also subject to message-delivery checks.
- A safety block does not guarantee a particular visitor-facing explanation. The configured route and workflow determine the result.
- An external agent may use the pushback failure action without rerunning its work.
- A blocked spam check supplies no spam verdict. A blocked spelling check can return the original draft. Neither outcome means the text passed a safety check.

## Depends on and works with

### This feature uses

- **Requires** [AI guardrail policy connections](AIP-022.md): A configured guardrail supplies the safety block to route.
- **Configured in** [AI accounts and shared defaults](AIP-001.md): AI account settings provide safety configuration and fallback routing rules.
- **Configured in** [Shared AI service connections](AIP-013.md): Guardrail connection rules take precedence over account rules.
- **Configured in** [Virtual Assistant task setup](ENG-009.md): A selected workflow task defines what happens after a guardrail rule triggers it.
- **Works better with** [Censorship and abuse warnings](CHAT-021.md): Inbound guardrail decisions can use censorship handling.

### Used by

- **Configures** [AI Experiments Lab for response testing](AI-032.md): Guardrail experiments evaluate the routing rules configured on the selected connection.
- **Configures** [AI guardrail policy connections](AIP-022.md): Guardrail routing determines blocking, continued execution, regeneration, censorship or a workflow response.
- **Enhances** [AI spam checks for incoming conversations](AI-011.md): Configured guardrails can stop a spam check without returning a spam verdict.
- **Enhances** [AI run limits](AI-014.md): Guardrails can block or change processing before an answer is delivered.
- **Enhances** [AI agent service connections](AI-024.md): Configured safety decisions can block external agent output.

## FAQ and troubleshooting

### Why did AI stop instead of trying another answer?

The matched rule may be **Block**, or the pushback failure action may block further work.

### Can I change a safety rule myself?

Powerfront manages these settings. Ask Powerfront to change the relevant **Routing Rules**.

### Why did the conversation enter a workflow?

A matching safety rule, or a failed pushback rule, can select **Trigger Workflow**.

### Why did a specific rule take precedence over another rule?

Specific policy and context rules are considered before reason-pattern rules. Direction also affects which rules can match.