Review Queue

The Review Queue displays messages that have been held by the Safety Gateway for human review before they can be sent. This page allows you to approve, deny, or edit held messages.

Accessing the Review Queue

1. Navigate to Build > Governance > Safety Gateway
2. Click on the Review Queue tab

The queue is also accessible from the Safety Settings page via the navigation tabs.

Dry-Run Mode Indicator

When the Safety Gateway is in Dry Run mode, the Review Queue displays a prominent banner indicating that all decisions are simulated. In dry-run mode:

• A blue banner appears at the top of the queue: "Gateway in Dry Run Mode - Decisions are simulated"
• Each queue item shows its simulated decision badge (what would have happened if enforcement were active)
• Items are not actually held - they were allowed through but logged for observation
• This helps you tune thresholds before enabling full enforcement

Queue Overview

Status Tabs

The queue is organized into three tabs:

Tab	Description
Pending	Messages awaiting review (default view)
Approved	Messages that were approved and sent
Denied	Messages that were blocked

Queue Statistics

At the top of the page, you'll see:

• Pending Count - Number of items awaiting review
• Oldest Pending - How long the oldest item has been waiting
• Avg. Wait Time - Average time items spend in queue
• Today's Reviews - Number of items reviewed today

Pending Items

Item Information

Each pending item displays:

Field	Description
Timestamp	When the message was held
Agent	Which agent generated the message
Mailbox	Source mailbox for the original email
Recipients	Who the message would be sent to
Classification	Internal, External, or Mixed recipients
Danger Score	0.0 - 1.0 risk score (color-coded)
Flags	Risk indicators detected by the gateway

Danger Score Colors

• Green (0.0 - 0.3): Low risk
• Yellow (0.3 - 0.6): Medium risk
• Orange (0.6 - 0.8): High risk
• Red (0.8 - 1.0): Critical risk

Flags

Common flags include:

• pii_detected - PII found in message
• external_recipient - Message going outside organization
• sensitive_content - LLM detected potentially sensitive content
• unusual_request - Message seems out of character for the agent
• high_recipient_count - Many recipients (potential mass mailing)
• new_external_domain - First time contacting this domain

Reviewing an Item

View Details

Click on any item to open the detail panel showing:

1. Original Email - The email that triggered the agent
2. Generated Response - The message that was held
3. Recipients - Full recipient list with classifications
4. Risk Analysis - Detailed breakdown of why it was flagged
5. PII Detected - List of PII types found (if any)
6. LLM Analysis - Full reasoning from the safety analysis

Actions

For each pending item, you can:

Approve

Send the message as-is. Use when:

• False positive detection
• Appropriate business communication
• Risk is acceptable

Approve with Edits

Modify the message before sending. Use when:

• Message is mostly fine but needs adjustment
• PII should be removed or redacted
• Tone needs modification

Editing flow:

1. Click "Edit & Approve"
2. Modify the message content
3. Review changes
4. Click "Approve"

The edited version is re-scanned by the Safety Gateway before sending. If the edited content still exceeds the safety threshold:

• The edit modal displays the new danger score and analysis
• Highlighted issues show specific problematic content with severity levels (high, medium, low)
• Suggestions are provided for how to revise the content
• You can continue editing to address the issues
• A Force Send option is available for authorized users to override the re-scan

The edited version is sent; the original is preserved in the audit log.

Deny

Block the message from being sent. Use when:

• Clear policy violation
• Inappropriate content
• Sensitive data exposure risk

When denying, you can optionally:

• Add reviewer notes explaining the denial
• Send feedback to improve the agent

Request More Info

Flag the item for additional investigation. Use when:

• Need context from another team member
• Unclear business justification
• Requires escalation

Bulk Actions

For efficiency with multiple similar items:

Select Multiple

• Click checkboxes next to items
• Or use "Select All" for current page

Bulk Approve

Approve all selected items at once. Useful for:

• Clearing false positives
• Processing routine communications

Bulk Deny

Deny all selected items at once. Use carefully.

Filtering and Sorting

Filters

Filter	Options
Agent	Filter by specific agent
Classification	Internal, External, Mixed
Min Danger Score	Show only items above threshold
Flags	Filter by specific flags
Date Range	Limit to specific time period

Sorting

Sort by:

• Timestamp (oldest/newest first)
• Danger Score (highest/lowest first)
• Urgency (urgent/high/normal)

Time-Sensitive Items

Some held items may be marked as time-sensitive:

• Displayed with a clock icon
• Shows deadline (if any)
• Sorted to top by default
• May trigger additional notifications

Best Practices

Review Workflow

1. Start with highest risk - Sort by danger score descending
2. Check time-sensitive - Handle urgent items first
3. Batch similar items - Use bulk actions for clear cases
4. Add notes - Document reasoning for training and audit
5. Review daily - Avoid accumulating backlog

Decision Guidelines

Scenario	Recommended Action
False positive, routine email	Approve
Minor PII (phone, email) to internal	Approve
Credit card / SSN detected	Deny or Edit
External + sensitive content	Review carefully
High recipient count	Verify intent

Quality Assurance

• Periodically review approved items in audit log
• Look for patterns that could improve thresholds
• Adjust settings based on common false positives
• Document new edge cases for team training

Troubleshooting

Queue Growing Too Fast

• Review threshold settings (may be too conservative)
• Enable "Skip LLM for Internal" if appropriate
• Add more reviewers if volume is high
• Check for specific agents generating excessive holds

Items Timing Out

• Set up queue alerts at lower threshold
• Assign dedicated review times
• Consider higher thresholds for low-risk categories
• Enable emergency override for critical situations

Can't See All Items

• Check filter settings (may be hiding items)
• Verify permissions (need safety:queue:read)
• Clear date range filter

Related Topics

• Safety Gateway Overview - Understand how the gateway works
• Safety Settings - Configure thresholds
• Audit Log - View historical decisions