Prompt Monitoring Guardrails define the policies used to detect violations across monitored AI systems. They are the foundation of Prompt Monitoring, determining which prompts, responses, and conversations are flagged for review, and giving organizations a structured way to apply both standard protections and business-specific monitoring rules.
This page supports more than basic rule toggling. It allows organizations to apply default guardrails, create custom guardrails, configure sensitivity labels, and save those settings into reusable guardrail packages that can be applied across multiple AI systems. At the AI system level, those package settings can then be refined with targeted overrides where needed.
Purpose
Prompt Monitoring Guardrails help organizations:
- apply a baseline set of prompt monitoring policies across monitored AI systems
- detect common risks using default guardrails
- create custom guardrails for business-specific, regulatory, or operational requirements
- configure sensitivity labels for context-aware detection scenarios
- save policy settings into reusable guardrail packages
- apply consistent monitoring configurations across multiple AI systems
- support AI-system-specific overrides without changing the broader company baseline
This page is the policy backbone of Prompt Monitoring. It defines what gets detected before those detections appear in the monitoring workflow for review and incident escalation.
Overview of Page Sections
Guardrail Package
Each monitored AI system can be assigned a Guardrail Package. This acts as the baseline policy set for that system. If no package is assigned, the system uses the global defaults instead. The AI-system-level guardrails view also makes it clear when a user is overriding either the assigned package or the tenant-wide defaults for that system only.
Packages make it easier to standardize monitoring across AI systems that share the same use case, risk profile, or governance requirements.
Default guardrails
Default guardrails provide a starting point for prompt monitoring and can be used across a wide range of business cases. These are the standard policies your organization can rely on for broad coverage and consistent baseline monitoring.
Custom guardrails
Custom guardrails allow organizations to extend beyond the default set and create policies tailored to specific AI systems, business processes, or governance requirements. These are useful where standard guardrails are not enough, or where monitoring needs to reflect internal policy, legal obligations, or use-case-specific concerns. The AI-system guardrails view includes actions to add, edit, and remove custom guardrails.
Sensitivity labels
Sensitivity labels are configured from the same guardrails area, but they serve a distinct role within prompt monitoring. They support more context-aware detection scenarios and complement the broader guardrail framework rather than replacing it. Sensitivity labels can also be included as part of a reusable package configuration.
Save and reuse configurations
Guardrail settings can be saved as packages so the same mix of default guardrails, custom guardrails, and sensitivity labels can be reused across multiple AI systems. At the AI-system level, users can also save a current configuration as a package, update an existing assigned package, or save their changes as a new package. This supports both standardization and controlled variation across the monitored environment.
Guardrail packages
Guardrail packages are intended to make policy management scalable.
Instead of recreating the same monitoring configuration for every AI system, organizations can bundle their selected settings into a package and apply that package wherever the same controls are needed. This helps keep prompt monitoring consistent across similar AI systems while reducing manual setup effort. The package workflow supports creating packages, editing them, and seeing how many AI systems currently use each one.
Where a package is assigned to an AI system, users can still make system-level changes. Those changes are treated as overrides unless the user explicitly updates the package itself. The UI also makes clear when updating a package will affect other AI systems that use it.
Prompt Policies page
The broader, organization-level view for this feature sits under the Guardrails page on the Prompt Policies tab. This is the macro management view for prompt monitoring policies across the tenant. From there, administrators can:
- manage the company’s global default guardrails
- create and maintain guardrail packages
- see which packages exist across the organization
- view which AI systems are assigned to which packages
- manage custom guardrails and sensitivity labels at the broader company level
The Guardrails page explicitly presents Prompt Policies as its own tab, and the policy management area includes sections for both Guardrail Packages and AI System Assignments.
How this supports Prompt Monitoring
The relationship between the two pages is straightforward:
- Prompt Monitoring Guardrails defines what should be detected
- Prompt Monitoring shows where those detections occurred and how they should be reviewed
Without guardrails, Prompt Monitoring has no structured basis for deciding which interactions should be surfaced as alerts or violations.
Notes
- The page title should stay focused on the monitoring function, which is why Prompt Monitoring Guardrails is a stronger label than naming the page around custom guardrails or sensitivity labels alone.
- Sensitivity labels and custom guardrails are features within the broader guardrails model, not the primary frame for the page.
- Guardrail packages are the best way to apply the same settings across multiple AI systems with minimal rework.
- AI-system-level changes do not automatically change the package or the tenant-wide defaults unless a user explicitly updates that shared configuration.