Guardrail Hits Reporting
Updated
Guardrail Hits is a dedicated reporting dataset available under the Sprinklr AI+ data source. This report provides visibility into all requests where one or more guardrails were triggered during AI request processing.
The Guardrail Hits report focuses exclusively on guardrail detections, allowing administrators, auditors, and AI governance teams to monitor safety, compliance, and policy enforcement across AI-powered experiences.
What is a Guardrail Hit?
A guardrail hit occurs when at least one configured guardrail evaluates a request and detects a policy violation, restricted content type, or other protected condition.
The Guardrail Hits report:
- Contains all requests where one or more guardrails were triggered.
- Includes both input guardrail and output guardrail detections.
- Records guardrail evaluations regardless of whether an AI model call was eventually made.
- Stores detailed information about the triggering request, detected guardrail, affected message, originating deployment, and associated user or business context.
Record Granularity
Each record represents a single request in which one or more guardrails were triggered.
Examples include:
- Denied topic detection
- Harmful content detection
- Safety policy violations
- Restricted content categories
- Custom guardrail detections
Use Cases
The Guardrail Hits report helps you:
- Monitor AI safety and compliance.
- Audit blocked or restricted requests.
- Identify frequently triggered guardrails.
- Analyze harmful or denied-topic trends.
- Track guardrail performance across deployments.
- Investigate end-user interactions that resulted in policy violations.
- Measure detection latency and operational efficiency.
Configure Guardrail Hits Reporting Widget
Follow these steps to configure AI+ Reporting:
Step 1: Create a Reporting Widget
Navigate to Service Reporting.
Click the + Create Widget button under a reporting dashboard to build a new report.
For detailed steps, refer to Create a Reporting Widget.
Step 2: Select the Data Source
In the widget setup screen, choose Sprinklr AI+ in the Data Source field from the dropdown.
Step 3: Configure Metrics and Dimensions
In the Add a Metric or Dimension window, expand Sprinklr AI+ from the left navigation pane.
Select Guardrail Hits under the Sprinklr AI+ category.
Browse or search for the required metrics and dimensions, such as Detection Count, Guardrail Type, Guardrail Name, Input/Output Message, or Latency.
Select the metrics and dimensions you want to add to the widget.
Select Done to apply your selections and configure the report.
Available Metrics
The following measurement metrics are available in the Guardrail Hits dataset.
Metric | Description |
Detection Count | Total number of guardrail detections. |
Latency | Time taken for guardrail evaluation, measured in milliseconds. |
Available Dimensions
The following dimensions are available for segmentation, filtering, and reporting.
Detection Information
Dimension | Description |
Guardrail Type | Type of guardrail hit. Multiple values can exist for a request if multiple guardrails are triggered. |
Guardrail Name | Specific guardrail that generated the detection. |
Applied On | Indicates whether the guardrail was evaluated against the user's input message or the model's output response. |
Detected Traits | Traits or categories identified by the guardrail during evaluation. |
Message Information
Dimension | Description |
Input/Output Message | The content of the message on which the guardrail was triggered. For input guardrails, this is the user message. For output guardrails, this is the generated model or agent response. |
Blocked Message Shown | The actual blocked message displayed to the user or agent after the guardrail hit. |
Request Information
Dimension | Description |
Request ID | Unique identifier of the guardrail evaluation request. |
AI+ Request ID | AI+ request identifier associated with the guardrail detection. This is primarily expected for output guardrails and may be available for input guardrails depending on implementation. |
Request Triggered Time | Timestamp when the request was submitted for guardrail evaluation. |
Request Completion Time | Timestamp when the guardrail evaluation response was received. |
Prompt and Deployment Information
Dimension | Description |
Prompt Node | Prompt node that triggered the guardrail. |
Prompt Node Name | Name of the prompt node within the deployment from which the request originated. |
Deployment | Deployment through which the request was executed. |
User and Workspace Information
Dimension | Description |
Workspace | Workspace in which the guardrail detection occurred. |
User | User who triggered the request that resulted in the guardrail hit. |
User Email | Email address of the user who triggered the request. |
Product and Feature Information
Dimension | Description |
Feature | Feature that triggered the request. The value corresponds to the feature name configured in Feature Access Management. |
Module | Product module associated with the triggering feature. |
Product Suite | Product suite associated with the corresponding module. |
Business Context Information
Dimension | Description |
Case Number | Case associated with the guardrail hit, if applicable. |
Conversation | Conversation ID associated with the guardrail hit, if applicable. |
Task | Bot task associated with the detection, when available. |
Building a Guardrail Hits Widget
You can use the Guardrail Hits dataset to create widgets that answer questions such as:
Most Triggered Guardrails
Guardrail Name | Detection Count |
National Elections | 2,450 |
Self-Harm Content | 1,820 |
Harmful Content | 1,321 |
Guardrail Detections by Workspace
Workspace | Detection Count |
Customer Support AI | 7,842 |
Agent Assist AI | 4,310 |
Knowledge Copilot | 2,117 |
Input vs Output Guardrail Distribution
Applied On | Detection Count |
Input | 8,125 |
Output | 3,604 |
Average Detection Latency
Guardrail Type | Average Latency (ms) |
Denied Topics | 115 |
Harmful Content | 132 |
Toxicity Detection | 148 |
Sample Analysis Scenarios
Audit Restricted Topics
Filter by:
- Guardrail Type
- Guardrail Name
- Deployment
- User
Use this view to identify which users, deployments, or workspaces most frequently encounter restricted content categories.
Review Blocked Responses
Analyze:
- Applied On = Output
- Input/Output Message
- Blocked Message Shown
This helps validate that output guardrails are functioning correctly and presenting the intended fallback responses.
Monitor Guardrail Performance
Track:
- Detection Count
- Latency
- Request Triggered Time
- Request Completion Time
Use these metrics to measure guardrail processing efficiency and identify latency spikes.
Deployment-Level Compliance Monitoring
Segment reporting by:
- Deployment
- Prompt Node
- Workspace
- Product Suite
This helps governance teams identify deployments with elevated guardrail activity.
Key Considerations
- A guardrail hit may occur even when no AI model request is executed.
- Both input and output guardrails are included in reporting.
- Multiple guardrails can be triggered within a single request.
- Detection Count represents guardrail detections captured by the reporting pipeline.
- AI+ Request ID availability may vary depending on whether the detection occurred before or after model execution.
- Business context dimensions such as Case Number and Conversation are populated only when available in the originating request context.
- Guardrail hits generated during Prompt Node > Test executions are not logged in the Guardrail Hits report to prevent test activity from cluttering production reporting data. When a guardrail is triggered during testing, the detection details are displayed directly in the test interface for immediate review.
The Guardrail Hits report provides a centralized view of all AI safety and policy enforcement events across Sprinklr AI+. By combining detection metrics with detailed contextual dimensions such as guardrail type, user, deployment, workspace, feature, and message content, the report enables administrators to monitor compliance, investigate violations, and improve the effectiveness of AI governance strategies.