AI guardrail policy connections
Guardrails check selected AI inputs and outputs against an external content policy and apply the configured response to an intervention.
What it is
Guardrails connect selected AI flows to an external content policy. INSIDE can check text before generation and check supported output before delivery. A policy intervention then follows the configured routing rules.
Who uses it and why
Administrators agree the content policy and intervention behaviour with Powerfront. Operators may receive a conversation after an intervention, while visitors may see a revised reply or another configured response.
Where to find it
Powerfront manages Guardrails in Settings › AI › Accounts and Amazon Bedrock Guardrails in Settings › AI › Connections. Ask Powerfront to change the connection or Routing Rules for you.
How it works
Before generation, the external service checks user-message text in the prompt. If there is no user-message text, it can check the instructions instead.
For supported outgoing work, checks can include the reply, text within rich content, reply options and relevant tool results. Configured business context can also accompany the check.
When the policy intervenes, INSIDE keeps assessment information and chooses a routing action. Available actions can block, continue, trigger a workflow, ask the AI to rephrase, or use censorship handling. The visible result depends on the AI feature and its intervention configuration.
Routing Rules are evaluated from top to bottom. The first matching rule wins, with a catch-all rule last. A rule's default action is to block; a failed rephrasing attempt also defaults to blocking.
Configure it
Ask Powerfront to prepare the policy connection, enter its API Key and Guardrail ID, and confirm Region and Version. The defaults are us-east-1 and DRAFT. Ask for a specific policy version when you need checks against that version.
Powerfront can select the connection under Guardrails on the applicable AI account and configure Routing Rules. No guardrail connection is selected by default. Agree the inbound and outbound actions, then test examples that should pass and examples that should intervene.
Settings
| Setting | Where | Applies to | Default | Who can change it | What it does |
|---|---|---|---|---|---|
| Guardrails | Settings › AI › Accounts | website | None | Powerfront | Selects the optional guardrail connection used by the account's supported AI flows. |
| Region | Settings › AI › Connections | website | us-east-1 | Powerfront | Selects the Amazon Bedrock region for the guardrail connection. |
| Version | Settings › AI › Connections | website | DRAFT | Powerfront | Selects the guardrail policy version to apply. |
| Policy | Settings › AI › Connections › Routing Rules | website | Any | Powerfront | Selects the policy category that a routing rule matches; Any leaves the category unrestricted. |
| Context | Settings › AI › Connections › Routing Rules | website | Any | Powerfront | Selects the policy context that a routing rule matches; Any leaves the context unrestricted. |
| Direction | Settings › AI › Connections › Routing Rules | website | Both | Powerfront | Chooses whether a routing rule applies to inbound checks, outbound checks or both. |
| Action | Settings › AI › Connections › Routing Rules | website | Block | Powerfront | Selects the response to a matching policy intervention. |
Fields
| Field | Where | Required | What it is for |
|---|---|---|---|
| API Key | Settings › AI › Connections | Yes | Authenticates the Amazon Bedrock guardrail connection. |
| Guardrail ID | Settings › AI › Connections | Yes | Identifies the policy to apply. |
Limits and good to know
- A policy intervention and an unavailable guardrail service are different outcomes. Ordinary service failures can allow processing to continue.
- Checking selected text does not establish coverage of every attachment, image or live audio stream.
- Choosing a connection affects supported features using that account.
- Guardrail checks have their own usage information. Model token cost estimates do not establish the full cost of the external checks.
Depends on and works with
View all connections as a list
This feature uses
- AI guardrail routing and fallback actionsConfigured by
Guardrail routing determines blocking, continued execution, regeneration, censorship or a workflow response.
- Rich content in AI repliesEnhances
Output checks include supported rich-content text and reply options.
- AI accounts and shared defaultsConfigured by
The AI account selects a guardrail connection.
- Shared AI service connectionsRequires
External service configuration supplies the policy connection.
- AI execution tracing with LangfuseRelated
Use optional execution traces to investigate recorded guardrail assessments and interventions.
Used by
- AI guardrail routing and fallback actionsRequires
A configured guardrail supplies the safety block to route.
- AI Experiments Lab for response testingEnhances
A configured guardrail service enables guardrail-policy experiments.
- INSIDE AiFlow test dashboard · managed by PowerfrontEnhances
The dashboard includes guardrail testing tools.
- AI accounts and shared defaultsEnhances
An optional guardrail connection checks AI work.
- Shared AI service connectionsEnhances
Amazon Bedrock Guardrails connections provide external guardrail checks.
FAQ and troubleshooting
Why was a flagged answer still allowed?
Ask Powerfront to check the matched Routing Rules action. A rule can explicitly allow processing to continue. Also confirm that the check ran successfully.
Why did the visitor receive a different response after a policy intervention?
The configured action may request rephrasing or trigger a workflow. Review the intervention behaviour with Powerfront.
Does a service outage automatically block AI replies?
No. Ordinary guardrail service failures can allow processing to continue. Do not rely on service failure as an automatic block.
Why did a policy edit change the checks?
The connection may use DRAFT. Ask Powerfront to confirm Version and whether a specific published policy version should be selected.