Share Insights from Einstein Generative AI Audit and Feedback Data

Use Pre-Built Dashboards and Reports

Use the pre-built Data 360 dashboards to:

• analyze the generative AI, and
• share insights such as data masking, toxicity in responses, user trends, acceptance rates, and response feedback.

Note When you turn on Einstein Generative AI Audit and Feedback for the first time, it can take a few minutes before you can access the pre-built reports and dashboards. If the Einstein generative AI data collection storage is turned on but you don't see the reports and dashboards, turn off Einstein Generative AI data collection and storage, and then turn it back on again.

To see more details about toxicity in responses and data masking, use the pre-built Einstein Trust Layer Toxic Response reports.

Get Insights with Queries

Use queries to analyze audit and feedback data. This example shows linking GenAIGtwyObjRecord to an sObject to GenAIGatewayRequest or GenAIFeedback. Suppose the sObject is Case. The query can get the number of generative AI requests used or called for the case. You can compare the time to resolve for cases with generative AI usage versus cases without generative AI usage. You can also link the GenAIFeedback table to determine whether generative AI was helpful, or just to give feedback to the generative AI for that particular case.

SELECT *
        FROM "GenAIGtwyObjRecord__dlm"
        JOIN "GenAIFeedback__dlm" ON "GenAIGtwyObjRecord__dlm"."parent__c" = "GenAIFeedback__dlm"."feedbackId__c"
        JOIN "GenAIGeneration__dlm" ON "GenAIFeedback__dlm"."generationId__c" = "GenAIGeneration__dlm"."generationId__c"
        JOIN "GenAIGatewayResponse__dlm" ON "GenAIGeneration__dlm"."generationResponseId__c" = "GenAIGatewayResponse__dlm"."generationResponseId__c"
        JOIN "GenAIGatewayRequest__dlm" ON "GenAIGatewayResponse__dlm"."generationRequestId__c" = "GenAIGatewayRequest__dlm"."gatewayRequestId__c"
        LIMIT 10

Build Data 360 Reports

To analyze audit and feedback data, build reports in Data 360 that pull relevant information from the data model objects (DMOs) in the Audit and Feedback data model. You can use calculated insights to create metrics that can trigger alerts. To learn more, see Data 360 Reports and Dashboards.

Tag Limits in the GenAIRequestTag Data Lake Object (DLO)

If you query GenAIRequestTag data in Data 360 for analytics, auditing, or debugging, Salesforce enforces a maximum number (default of 20) of individual rows written per LLM request. Tags that exceed that limit appear in a single REMAINDER_TAGS row containing a JSON array.

Reserved Planner Tags

Ten reserved planner tags are always individual rows. These tags don’t count toward the 20-tag limit.

• action_id
• action_name
• bot_id
• bot_session_id
• bot_version_id
• plan_id
• planner_id
• user_context
• user_id
• x-client-trace-id

Example REMAINDER_TAGS Entry

Here’s an example that shows the content of the tagValue__c field associated with the REMAINDER_TAGS entry in a single row.

[
        {
        "tagValue": "val_21_overflow",
        "tag": "np_21"
        },
        {
        "tagValue": "val_22_overflow",
        "tag": "np_22"
        },
        {
        "tagValue": "val_23_overflow",
        "tag": "np_23"
        },
        {
        "tagValue": "val_24_overflow",
        "tag": "np_24"
        },
        {
        "tagValue": "tag_test_org",
        "tag": "x-client-trace-id"
        }
]

Sample Data 360 SQL Query to Parse Tags, Including Remainder Tags

SELECT tag__c, tagValue__c, timestamp__c
        FROM "GenAIGatewayRequestTag__dll"
        WHERE timestamp__c >= TIMESTAMP '2026-03-27 04:00:00+00:00'
        AND tag__c = 'REMAINDER_TAGS'
        ORDER BY timestamp__c DESC
        LIMIT 10

Update Previous SQL Queries to Parse Remainder Tags

Previously created SQL queries that filter by tag__c don't use overflow tags as individual rows. Update these queries to parse the REMAINDER_TAGS JSON value. Any dashboards, reports, or Data Cloud recipes that assume the one-row-per-tag approach can return incomplete results rather than errors.