Troubleshooting Scenario Questions

Retrain the model to improve its accuracy for that lead.

Why wrong: Retraining for one lead is inefficient; insights are more direct.

Override the score manually to reflect their judgment.

Why wrong: Manual overrides reduce model trust.

Disable Einstein Lead Scoring for that lead record.

Why wrong: Disabling doesn't investigate; it removes scoring.

Group Alpha

Why wrong: Accuracy is high at 0.95.

Group Gamma

Why wrong: Accuracy is 0.93, similar to others.

Group Delta

Why wrong: Accuracy is 0.91, not unusual.

Privacy

Why wrong: Privacy concerns data protection, not bias or fairness in model outcomes.

Accountability

Why wrong: Accountability addresses who is responsible for AI decisions, not the bias in predictions.

Transparency

Why wrong: Transparency involves explaining how decisions are made, not the fairness of outcomes.

Privacy

Why wrong: Privacy relates to data handling, not the ability to explain recommendations.

Reliability

Why wrong: Reliability concerns consistent performance, not the ability to explain decisions.

Human oversight

Why wrong: Human oversight involves human review, not the AI's own explainability.

Route all complex requests directly to human agents without bot interaction.

Why wrong: This undermines the bot's purpose.

Increase the confidence threshold for intent matching to avoid misclassification.

Why wrong: Higher threshold may cause the bot to not understand more queries.

Reduce the number of dialogue options to simplify the bot's logic.

Why wrong: This limits the bot's capability.

The data source is not refreshed daily

Why wrong: Data refresh frequency does not prevent model from generating predictions; it affects recency.

The object is a custom object

Why wrong: Custom objects are fully supported by Prediction Builder.

The object has fewer than 1000 records

Why wrong: Prediction Builder requires at least 100 rows, not 1000.

Increase the confidence threshold to 90%

Why wrong: Higher threshold reduces false positives but may cause more unrecognized intents.

Use a hierarchical intent structure

Why wrong: Hierarchy helps organize intents but does not directly improve recognition accuracy.

Reduce the number of intents to two

Why wrong: Fewer intents may not cover customer needs, leading to poor service.

Ensure at least 100 articles are in the knowledge base

Why wrong: No strict minimum; even fewer articles can work if views exist.

Confirm that article authors have the correct profile permissions

Why wrong: Author permissions do not affect recommendation display.

Increase the recommendation frequency from daily to hourly

Why wrong: Frequency is not usually the cause of missing recommendations.

Retrain the Opportunity Scoring model

Why wrong: Retraining does not address missing historical data.

Verify that users have the 'View Einstein Scores' permission

Why wrong: Permissions affect visibility, not score generation.

Wait 48 hours for the model to update

Why wrong: Waiting does not resolve data insufficiency.

The connected email service is not authenticated for the organization.

Why wrong: Authentication error would be different.

The contact with ID 003B0000007v4ZW does not have a valid email address.

Why wrong: Email exists but domain not configured.

The user 'admin@mycompany.com' has exceeded their email capture quota.

Why wrong: Quota warning is not the cause.

The number of fields used as input features.

Why wrong: Feature count is less likely the primary cause.

The API version used in the integration.

Why wrong: API version affects connectivity, not confidence.

The permission set for the Einstein AI feature.

Why wrong: Permissions affect access, not output.

Lead_Score_Model because its accuracy is lower than expected.

Why wrong: 92% is high; no issue indicated.

Campaign_Response because it has the oldest training date.

Why wrong: It is deployed and has good accuracy.

Opportunity_Forecast because it is still training.

Why wrong: Training is a normal state; no investigation needed unless stuck.

The model was not retrained after the last data load.

Why wrong: Retraining might be needed but wrong sentiment despite high confidence suggests labeling issues.

The training data contains predominantly neutral examples.

Why wrong: Imbalanced data can cause low confidence, not high confidence wrong predictions.

The field mapping for the sentiment field is incorrect.

Why wrong: Mapping issue would likely cause errors, not confident wrong predictions.

Review the Exchange Online mailbox audit logs to see if emails are being accessed.

Why wrong: This might be useful later but is not the first step.

Check the email synchronization frequency in the Einstein Activity Capture settings.

Why wrong: Sync frequency is not likely the cause of the 60% rate; it's usually real-time.

Ensure that all users have the 'Email Integration' permission enabled in their profiles.

Why wrong: This permission is not required for Activity Capture.

The data transform includes a filter that removes records with null Amount.

Why wrong: A filter would remove nulls, not cause a failure; the error suggests nulls are being processed.

The source object has a validation rule.

Why wrong: Validation rules apply to UI record creation/update, not data flows.

The data flow schedule is incorrect.

Why wrong: The error is data-related, not schedule-related.