End users submit feedback using the thumbs up / thumbs down buttons on the Precursive AI component, or directly from email digest links. As a builder, you can review all feedback submitted for the agents you manage and use it to refine your Rules, Prompt Instructions, and salience configuration.

Where to Find Feedback

Feedback records are stored against the Insight Result they relate to. To review feedback:

1. Go to the App Launcher in Salesforce
2. Search for Insight Feedback
3. Use list views to filter by Zen Agent, date range, rating, or status

You can also navigate to a specific Insight Result record and view its related feedback directly from the related list.

Understanding Feedback Records

Each feedback record captures the following information:

Users can only view and modify their own feedback. As a builder, you have read access to all feedback across agents you manage.

Managing Feedback with the Status Workflow

Each feedback record has a Status field that lets you track your response to it.

Update the Status and use the Admin Notes field to document what you found and what action (if any) you took. This creates a clear audit trail and helps you track which feedback has been addressed.

Using Feedback to Improve Agent Quality

Patterns to look for

Inaccurate - the AI made a claim that was factually wrong or misrepresented the data. Review the citations on the affected insights to understand what data the AI was working with. This usually points to a Context Source returning unexpected data, or a Prompt Instruction that is too loose.

Not relevant - the insight was technically correct but not useful for the recipient. This often means the filter criteria is too broad, or the salience prompt is not filtering out low-value insights effectively.

Incomplete - the insight missed something important. Check whether the relevant data is being pulled in by your Context Sources, and whether your Prompt Instructions are specific enough to direct the AI's attention to what matters.

Formatting issue / Too verbose / Too brief - these are often quick wins. Adjust the Prompt Instructions to specify the output format more precisely, for example "Provide exactly 3 bullet points" or "Keep the summary to two sentences."

Citation - the user found a citation to be wrong or misleading. Download the context file from the Insight Result to verify what data the AI actually received, and review whether the Context Source is returning the right records.

Action - the recommended action was not appropriate. Review the Action Instructions and your agent-specific Agent Action Instructions on the linked Action, or consider changing its Execution Mode from Autonomous to Supervised so users can approve before it runs.

Actions to take

• For inaccurate or incomplete insights, download the context file from the affected Insight Result record to see exactly what data the AI received, then adjust the relevant Context Source or Prompt Instructions
• For not relevant or low-value insights, tighten the Source Record Filter Criteria or refine the Salience Prompt to raise the threshold
• For formatting feedback, revise the Prompt Instructions with more specific output guidance
• After making changes, use the Preview action on the Zen Agent to test with a representative record before re-enabling scheduled runs
• Mark the feedback Status as Applied or Dismissed once you have resolved it, and document your decision in Admin Notes

Tracking Improvement Over Time

As you iterate on your agent configuration, track the ratio of positive to negative feedback over time using list views filtered by date. You should see the proportion of positive feedback increase as you address common patterns.

If feedback on a particular agent remains consistently negative despite adjustments, it may indicate that the underlying Context Sources need to be reviewed by an admin, or that the use case requires a fundamentally different agent structure.