Hard Difficulty Questions

Randomly select 2,000 tickets and label them manually.

Why wrong: Random selection doesn't target uncertainty.

Use a pre-trained model to label all tickets automatically.

Why wrong: Automated labeling may introduce noise.

Have subject matter experts label all 10,000 tickets.

Why wrong: Labeling all is costly and not active learning.

One-hot encode the region field using 50+ dummy variables.

Why wrong: Too many dummy variables can lead to overfitting and poor performance on small regions.

Remove the region field to avoid bias.

Why wrong: Region is potentially valuable; removing it loses information.

Use region as a numeric rank based on past conversion rates.

Why wrong: Ranking introduces ordering that may not exist; better to keep as categorical with grouping.

Increase data retention period for lead records

Why wrong: Does not address data incompleteness or staleness.

Rebuild the Einstein Lead Scoring model using different fields

Why wrong: May help but root cause is data quality, not field selection.

Increase the model training frequency to weekly

Why wrong: More frequent training on dirty data worsens issues.

Run a lead scoring batch job manually

Why wrong: Scoring runs automatically once enabled.

Create a custom report type for Lead Score

Why wrong: Standard report types already include the field.

Assign the 'View Lead Score' permission to users

Why wrong: Field-level security may be needed, but not a specific permission.

Update account records via process builder

Why wrong: No impact on resolution time.

Configure a flow to send churn alerts

Why wrong: Alerts don't reduce churn.

Create a custom report on churn

Why wrong: Only visualizes existing data, not prescriptive.

Generating report charts and dashboards.

Why wrong: Reports are created with Einstein Analytics, not GPT.

Creating workflow rules based on user behavior.

Why wrong: Workflow rules are declarative automation, not GPT.

Create a dialogue that triggers on the exact phrase 'I want to return a purchase'

Why wrong: Exact phrase matching fails for variations; intents handle natural language variations.

Create an entity called 'Return Purchase' and map it to a dialogue

Why wrong: Entities are for details like product name; 'return purchase' is an intent.

Use an intent called 'Customer Service' and a custom entity for return

Why wrong: Using a broad intent with a custom entity is less accurate; a dedicated intent is better.

Increase the weight of the majority class in the loss function.

Why wrong: Weighting the majority class would worsen recall; the model already overweights the majority.

Train multiple models on different random subsets and average predictions.

Why wrong: Bagging may not address imbalance; the ensemble will still be biased toward the majority.

Use a simpler model to avoid overfitting on the majority class.

Why wrong: Simplicity does not fix imbalance; the model may still ignore the minority class.

Accuracy

Why wrong: The system may be accurate in predicting clicks.

Privacy

Why wrong: Privacy may not be violated if data is properly handled.

Transparency

Why wrong: Transparency alone does not prevent harm.

Data Mask to protect sensitive information.

Why wrong: Data Mask protects privacy, not explainability.

Salesforce Shield with encryption and monitoring.

Why wrong: Shield provides security, not explanation of AI decisions.

Field Audit Trail to track changes to data.

Why wrong: Audit trail tracks data changes, not model reasoning.

Object Detection

Why wrong: Object Detection is for images, not text.

Image Classification

Why wrong: Image Classification is for images, not text.

The model is overfitted to English

Why wrong: Overfitting would not cause language-specific failure; the model simply hasn't seen Spanish.

The inference threshold is too high

Why wrong: Threshold affects confidence, not language understanding.

The model uses the wrong algorithm

Why wrong: Algorithm choice is less likely the cause than data imbalance.

Prompt Builder

Why wrong: Used for prompt templates, not for building autonomous agents.

Einstein Copilot

Why wrong: Copilot is a conversational AI assistant, not part of Agentforce agent building.

Use a single expert labeler for all data.

Why wrong: Single labeler may have personal bias and doesn't allow for cross-validation.

Label all data by a single expert labeler.

Why wrong: Same as B; no redundancy.

Allow each labeler to interpret guidelines freely.

Why wrong: Lack of standardized guidelines leads to inconsistent labels.

Profitability.

Why wrong: Profitability is a business goal, not an ethical pillar.

Scalability.

Why wrong: Scalability is a technical requirement.

Prediction scores

Why wrong: Prediction scores are not a bot component.

Training phrases

Why wrong: Training phrases are part of intent setup, not a separate component.

Use the AI model as the sole decision-maker.

Why wrong: Sole reliance on AI can be unfair.

Use all available data including protected attributes.

Why wrong: Using protected attributes may cause bias.

Ignore adverse impact if the model is accurate.

Why wrong: Accuracy does not justify adverse impact.

The trigger is executed before the lead is saved

Why wrong: It's after insert, so lead is saved.

The trigger is querying a related object without an index

Why wrong: No query is shown; the trigger uses API call.

The Score__c field is not writeable

Why wrong: The error is not about field writeability.

Timeliness.

Why wrong: Timeliness matters for some use cases, but it is not a core dimension like the others.

Data volume.

Why wrong: Volume refers to quantity, not quality.