Prompt Monitoring provides a detailed view of the prompts and responses captured for a monitored AI system. It allows users to review detected violations, search and filter activity, inspect conversation context, reveal the identity of the submitting user where needed, and escalate serious issues as incidents.
This page is the primary workspace for reviewing runtime interactions. It is designed to help governance, security, compliance, and operational teams assess how an AI system is being used in practice and respond when a prompt or response requires follow-up.
Purpose
Prompt Monitoring supports responsible AI oversight by helping organizations:
- review captured prompts and responses for a monitored AI system
- identify and investigate policy violations and alerts
- search, filter, and sort prompt activity to focus on what matters most
- understand prompts in the context of a full conversation, not only as individual messages
- reveal the identity of the submitting user when investigation requires it
- report incidents directly from a detected violation
- maintain an auditable record of review activity and identity access
This page is intended for day-to-day monitoring and investigation once an AI system is actively being observed.
Overview of Page Sections
Sync status and monitoring summary
At the top of the page, the sync status shows whether monitoring data is up to date and when prompts were last synchronized. Summary cards provide a quick view of the current monitoring state, including open violations, flagged prompts, cleared items, and monitored channels. A Refresh Now action allows users to retrieve the latest captured activity.
These indicators provide an immediate view of review workload and help users confirm that the page reflects the most recent prompt activity.
Search, filters, and actions
Prompt Monitoring includes filters and controls to help users narrow the list of captured interactions. Users can search prompt content, filter by time range and channel, limit the view to open items, sensitivity label matches, prompts with responses, or prompts already linked to an incident, and change the sort order.
The page also provides actions to:
- Scan All to run detection across the monitored prompts
- Review to focus on prompts with open violations
These controls make it easier to move quickly from broad monitoring to focused investigation.
Prompt list
The prompt list displays captured prompt activity for the selected AI system. Each row includes the prompt timestamp, a preview of the prompt text, channel, response status, incident status, and current violation status.
Prompts that belong to the same conversation are visually grouped together and color-coded by conversation thread. Each prompt in the sequence is labeled with its turn order, making it easier to understand how a conversation progressed over time.
This grouping is especially important where the risk or policy concern becomes clear only when multiple turns are reviewed together.
Prompt-level alerts and violations
Each prompt row shows its current violation state. Selecting the status control, such as Open (2), reveals the alerts attached to that prompt.
Prompt Monitoring distinguishes between different types of alerts by color:
- Purple indicates a conversation violation. These are conversation-level findings and will appear across the prompts in the related conversation chain.
- Yellow indicates a sensitivity label finding.
- Red indicates a standard policy violation, including default or custom guardrail detections.
This visual distinction helps users quickly understand whether the issue relates to a single prompt, an instruction-based self-report, or a broader conversation pattern.
Conversation details panel]
Selecting a prompt opens a detail panel on the right side of the page. This panel provides the additional context needed to review and act on the interaction.
The detail panel includes:
- conversation date and time
- session ID
- channel
- masked user identity
- a Reveal action to show the user who submitted the prompt
- a View full session action to open the entire conversation thread
- full prompt content
- full agent response
- detected policy violations and alert details
- the Identity Access Log, which records each time a user identity has been revealed
This panel is the main review space for understanding what occurred, what was detected, and whether additional action is required.
Reviewing and resolving violations
Prompt Monitoring is also where violation review is managed.
For each detected violation, users can:
- review the violation title and severity
- read the supporting detail explaining why it was flagged
- report the issue as an incident
- update the status using the dropdown
Available statuses include:
- Open
- Under Review
- Dismissed
- Mitigated
This status workflow is intentionally manual. Organizations may manage investigation, response, and remediation processes differently, including processes that occur outside the platform. Prompt Monitoring therefore supports review and status tracking without enforcing a fixed incident-handling model.
Revealing user identity
By default, the submitting user is masked in the conversation details panel. Where investigation requires attribution, users can reveal the identity of the prompt submitter. Each reveal action is logged in the Identity Access Log, creating an auditable record of identity access.
This supports controlled access to user information while still enabling investigation when accountability or escalation is required.
Reporting incidents
Where a violation requires escalation, users can select Report Incident directly from the violation entry. This allows serious issues to move from monitoring into incident management without leaving the review workflow.
Incident reporting is especially useful where the interaction may require formal investigation, governance review, or action outside the prompt monitoring process.
Notes
- Prompt Monitoring should be used as the primary review page for captured prompt activity and detected violations.
- Conversation grouping and turn order are important when reviewing multi-turn interactions, particularly where a policy concern depends on the full session context.
- Conversation violations appear across the prompts in a related conversation chain to reflect that the finding applies at the session level rather than to only one turn.
- User identity is masked by default and reveal actions are logged for audit purposes.
- Violation status updates are managed manually so organizations can align review activity with their own investigation and incident workflows.
Related Help Pages
- AI Monitoring Dashboard
- Guardrails and Sensitivity Labels
- Drift Analysis
- Knowledge & Tools Monitoring