AI Associate Hard Questions — 25 Difficult Questions

Question 1

A data scientist is building a churn prediction model using Einstein Discovery. They want to ensure the model does not rely on sensitive attributes like race or gender, even if those are correlated with other features. Which technique is MOST aligned with Salesforce's data minimisation principle?

Choose an answer to begin — your selection is scored in the full session.

Accepted Answer

Remove sensitive attributes from the training data and avoid using proxies that strongly correlate with them. Data minimisation means using only the data necessary for the task. Excluding sensitive attributes from the feature set is the most direct way to prevent them from being used, even if they are correlated with other features. Correlation does not imply causation, and if those features are not essential, they should be removed.

Answer

Include all features and rely on the AI to ignore biased ones

Answer

Apply a fairness constraint after training to adjust predictions

Answer

Use differential privacy to add noise to the training data

Question 2

A Salesforce admin notices that Einstein Account Scoring is not generating scores for all accounts. Some accounts have no score even though they meet the data requirements. What is the most likely cause?

Accepted Answer

The org has fewer than 100 account records. Einstein Account Scoring requires a minimum of 100 account records to generate meaningful scores. If the org has fewer than 100 accounts, the model cannot establish a baseline for scoring, resulting in no scores for any account. This is a hard threshold enforced by the Einstein AI engine, not a configurable setting.

Answer

Einstein features are not enabled in the org

Answer

Users do not have the 'View Scores' permission

Answer

The accounts lack custom fields for scoring

Question 3

Which THREE capabilities are provided by Einstein GPT in Sales and Service?

Accepted Answer

Summarizing case conversations from service interactions.. Einstein GPT leverages generative AI to automatically summarize case conversations from service interactions, enabling agents to quickly grasp key details without reading entire transcripts. This capability is built on Salesforce's proprietary AI models that process natural language from service records to produce concise, actionable summaries.

Answer

Generating report charts and dashboards.

Answer

Creating workflow rules based on user behavior.

Question 4

A financial services company uses Salesforce AI to detect fraudulent transactions. The dataset has 1 million legitimate transactions and only 1,000 fraudulent ones. The model trained with default parameters achieves 99.9% accuracy but identifies no fraud (precision and recall of 0). The data scientist wants to maximize fraud detection (recall) while minimizing false positives. Which approach is most effective?

Accepted Answer

Use SMOTE to generate synthetic fraud samples to balance the dataset.. The dataset is extremely imbalanced (1,000 fraud vs. 1,000,000 legitimate), causing the model to achieve high accuracy by simply predicting the majority class. SMOTE (Synthetic Minority Over-sampling Technique) generates synthetic examples of the minority class (fraud) to balance the class distribution, forcing the model to learn patterns that distinguish fraud. This directly improves recall for the minority class while still allowing the model to learn from the majority class, thus minimizing false positives. Other methods like class weighting can help, but SMOTE is specifically designed to address imbalance by creating new, plausible minority instances, making it highly effective for this scenario.

Answer

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

Answer

Train multiple models on different random subsets and average predictions.

Answer

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

Question 5

A financial services company is deploying Einstein AI and must comply with regulations requiring explainable decisions. Which Einstein capability allows them to understand why an AI model made a specific prediction?

Accepted Answer

Einstein Trust Layer with model explainability features.. The Einstein Trust Layer includes model explainability features that provide insights into why a specific prediction was made, such as highlighting the key input features that influenced the outcome. This directly addresses regulatory requirements for explainable AI decisions by offering transparency into the model's reasoning process.

Answer

Data Mask to protect sensitive information.

Answer

Salesforce Shield with encryption and monitoring.

Answer

Field Audit Trail to track changes to data.

Question 6

A company deploys an AI recommender system that personalizes content. The system is trained on user click data. After deployment, the company notices that the system increasingly recommends sensationalist content, leading to user polarization. Which principle is being violated?

Accepted Answer

Beneficence. The recommender system's shift toward sensationalist content, which polarizes users, violates the principle of beneficence because it causes harm (user polarization) rather than promoting well-being. Beneficence requires AI systems to act in the best interests of users and society, not to optimize for engagement metrics at the expense of ethical outcomes.

Answer

Accuracy

Answer

Privacy

Answer

Transparency

Question 7

A retail company deploys an AI system that adjusts prices dynamically based on customer browsing history. The system charges higher prices to returning customers. This practice is known as:

Accepted Answer

Unfair price discrimination. Unfair price discrimination, because the system charges different prices to customers based on their personal browsing history, specifically penalizing returning customers with higher prices, which is a discriminatory pricing practice. This goes beyond legitimate personalization since it exploits customer loyalty data to charge more, raising ethical and fairness concerns. Option A, lack of transparency, is a related but distinct issue—the problem here is the discriminatory pricing itself, not merely that it is undisclosed. Option C, personalization, refers to tailoring offerings or experiences to users, but it does not inherently involve charging different prices to different customers. Option D, dynamic pricing, involves adjusting prices based on factors like demand, time, or inventory, not on individual customer identity or history.

Answer

Lack of transparency

Answer

Personalization

Answer

Dynamic pricing

Question 8

A Salesforce admin is configuring Einstein Conversation Insights. They want to automatically capture next steps from sales call recordings. Which component of Conversation Insights provides this functionality?

Accepted Answer

Next step capture (powered by Einstein NLP). Einstein Conversation Insights uses Einstein Natural Language Processing (NLP) to automatically analyze sales call transcripts and extract actionable next steps, such as follow-up tasks or commitments. This feature is specifically designed to identify and capture these items without manual effort, leveraging AI to parse conversational context.

Answer

Call recording transcription

Answer

Keyword tracking

Answer

Talk-time metrics

Question 9

A mid-sized company uses Salesforce for sales and service. They have implemented Einstein Prediction Builder on a custom object 'Support_Ticket__c' to predict whether a ticket will be escalated (field: 'Escalated__c' Boolean). The model was trained with 10,000 records and 15 fields including 'Subject', 'Description_Summary__c', 'Priority__c', 'Hours_to_Resolution__c', and others. After deployment, the model's precision for escalated tickets is only 30%, while recall is 80%. The business finds too many false positives. The admin notices that the 'Priority__c' field has many missing values (60% null) and that the field 'Is_Critical__c' (a formula field) was included though it flags tickets as critical only rarely. The data spans 12 months but the last 3 months have a significantly higher escalation rate due to a product bug that has since been fixed. Which course of action will most likely improve the model's precision without harming recall?

Accepted Answer

Filter training data to exclude tickets from the last 3 months and impute missing 'Priority__c' values with the most common priority.. It addresses both the data drift and data quality issues that degrade precision. Excluding the last 3 months removes the biased escalation pattern caused by a fixed product bug, ensuring the model learns from stable historical patterns. Imputing missing 'Priority__c' values with the most common priority reduces noise from nulls without discarding the field entirely, which helps maintain recall by preserving predictive signal.

Answer

Roll back the model to the version trained 6 months ago when escalation rates were lower.

Answer

Exclude the 'Priority__c' field from the model and retrain.

Answer

Remove the 'Is_Critical__c' field and increase training data to 50,000 records.

Question 10

Which THREE factors should be considered when evaluating the quality of a dataset for an AI model?

Accepted Answer

Presence of outliers that may skew the model.. Option B is correct because outliers can disproportionately influence model parameters (especially in distance-based or linear models), skewing predictions and degrading generalization, so their presence must be assessed during dataset quality evaluation. Option D is correct because a high percentage of missing values in key fields reduces the effective training signal, can introduce bias if missingness is non-random, and may require imputation or exclusion decisions that directly affect model reliability. Option E is correct because duplicate records inflate the apparent size of the dataset, cause overrepresentation of certain patterns, and can lead to data leakage between training and validation splits, undermining honest performance estimates. Option A is not a quality factor per se—record count reflects dataset size/quantity rather than quality, and a large dataset can still be noisy, biased, or incomplete. Option C is also not a quality metric—the number of distinct labels describes the target's cardinality or class structure (relevant to task type and class balance), not the intrinsic quality of the data itself.

Answer

Total number of records available for training.

Answer

Number of distinct labels in the outcome field.

Question 11

An AI Associate reviews the Lead Scoring model exhibit. What is the primary ethical concern with this model?

Accepted Answer

The model has low recall, potentially missing minority class leads.. The primary ethical concern is that the model has low recall, meaning it fails to identify a significant portion of actual positive leads (the minority class). In a lead scoring context, this can result in missed business opportunities and potential bias against certain customer segments, as the model systematically overlooks valuable leads that do not fit the majority pattern.

Answer

The model uses too many features.

Answer

The model is not explainable.

Answer

The training data is imbalanced.

Question 12

A developer is creating a prompt template for Einstein GPT to summarize customer case details. The prompt must include the case subject, description, and last 3 comments, but only when the case priority is High. Which approach best achieves this in Prompt Builder?

Accepted Answer

Use merge fields with a conditional IF statement to include fields only when priority is High. Prompt Builder supports conditional merge fields using IF statements, allowing dynamic inclusion of case subject, description, and last 3 comments only when the case priority is High. This approach ensures the prompt is concise and relevant, avoiding unnecessary tokens or manual edits.

Answer

Hardcode example cases in the prompt and rely on the model to generalize

Answer

Include all fields in the prompt and use a token limiter to truncate the response

Answer

Create a static prompt and manually edit it for each High-priority case

Question 13

A company needs to build a custom AI model that predicts whether a support case will be escalated based on account history, case description, and product category. The output should be a binary yes/no. Which tool should they use?

Accepted Answer

Einstein Prediction Builder. Einstein Prediction Builder is the correct tool because it is specifically designed for binary classification (yes/no) predictions using structured data from Salesforce objects, such as account history, case description, and product category. It allows admins to create custom predictive models without code, directly within the Salesforce platform, making it ideal for predicting case escalation.

Answer

Einstein Vision and Language Platform

Answer

Einstein Discovery

Answer

Einstein GPT

Question 14

A company wants to deploy an Einstein AI model that uses sensitive customer data. Which practice should they follow to comply with data privacy regulations?

Accepted Answer

Use Einstein Trust Layer features to mask personally identifiable information (PII) in the model.. The Einstein Trust Layer provides built-in capabilities to automatically mask or redact personally identifiable information (PII) before data is sent to the underlying AI model, ensuring compliance with data privacy regulations like GDPR and CCPA without requiring manual data handling. This feature operates at the platform level, intercepting data in transit and applying masking rules based on predefined patterns, so sensitive customer data is never exposed to the model or stored in its training logs.

Answer

Store all sensitive data in an external data lake and connect via APIs.

Answer

Obtain explicit consent from data subjects before using their data in AI models.

Answer

Limit the data used for training to only essential fields.

Question 15

A company wants to build an Einstein Bot that can handle order status inquiries and, if the customer is frustrated, hand off to a human agent. Which THREE steps are essential to implement this?

Accepted Answer

Configure a hand-off action to a human agent. Option B is correct because an Einstein Bot hands off to a live agent through a configured hand-off action (the 'Transfer to Agent' action), which routes the conversation to an Omni-Channel queue so a human agent can take over when frustration is detected. Option C is correct because the bot must define a dialog (a conversation flow built from dialogs and steps) that actually retrieves and presents the order status, which is the core task the bot is meant to handle. Option E is correct because the bot needs an intent such as 'Order Status' so its natural language model can recognize and classify the customer's request and trigger the appropriate dialog. Option A is not essential because Einstein Bots include built-in sentiment detection in the bot's dialog logic rather than requiring a separate Einstein Sentiment Analysis implementation. Option D is not essential because Einstein Bots use the platform's built-in NLP and intent model, so training a custom model with Einstein Platform Services is unnecessary for this scenario.

Answer

Use Einstein Sentiment Analysis to detect frustration

Answer

Train a custom NLP model using Einstein Platform Services

Question 16

Which THREE of the following are best practices for feature engineering in Einstein Studio?

Accepted Answer

Apply normalization to numerical features. Option B is correct because applying normalization (e.g., min-max scaling or z-score standardization) to numerical features puts them on a comparable scale, which improves the convergence and performance of many machine learning algorithms used in Einstein Studio. Option D is correct because using domain knowledge to create derived features (such as ratios, aggregates, or time-based calculations) surfaces business-relevant signal that raw columns may not express, boosting model accuracy. Option E is correct because one-hot encoding converts categorical variables into binary indicator columns, which is required since most ML algorithms in Einstein Studio cannot consume raw string categories directly. Option A is not a best practice because dropping every record with missing values discards potentially valuable data and can bias the model; imputation or missing-value indicators are preferred. Option C is not a best practice because feeding raw, untransformed data typically violates algorithm assumptions (scaling, encoding, handling skew) and degrades model quality.

Answer

Remove all records with missing values

Answer

Use raw data directly without any transformation

Question 17

Which THREE factors can affect the accuracy of an Einstein GPT response?

Accepted Answer

Model temperature setting. The model temperature setting directly controls the randomness of the output. A higher temperature (e.g., 0.9) produces more creative but potentially less accurate responses, while a lower temperature (e.g., 0.1) makes the model more deterministic and factual. This parameter is a core hyperparameter in large language models like those powering Einstein GPT, and it significantly influences response accuracy.

Answer

Brightness of the user interface

Answer

Length of the response generated

Question 18

A multi-national company wants to use Einstein Activity Capture to automatically log emails from Outlook to Salesforce. They have a requirement that emails to certain external domains (e.g., competitors) must never be logged. How should they configure this?

Accepted Answer

Configure excluded addresses in Einstein Activity Capture settings. Einstein Activity Capture provides a native configuration to exclude specific email addresses or domains from being logged. By adding competitor domains to the 'Excluded Email Addresses' list in the Activity Capture settings, the system prevents those emails from ever being captured, ensuring compliance without post-processing workarounds.

Answer

Use a Process Builder to delete the logged emails

Answer

Use Einstein Email Insights to filter out those domains after logging

Answer

Create a validation rule on the Email Message object

Question 19

A healthcare AI model uses patient data. The legal team requires that all data used for training be de-identified according to HIPAA Safe Harbor method. Which data handling process satisfies this?

Accepted Answer

Remove all 18 HIPAA identifiers from each record.. The HIPAA Safe Harbor method specifically requires the removal of all 18 identifiers listed in the HIPAA Privacy Rule from each patient record. This includes direct identifiers like names, addresses, and Social Security numbers, as well as indirect identifiers such as dates and geographic subdivisions. By removing these 18 identifiers, the data is considered de-identified and no longer subject to HIPAA restrictions, allowing it to be used for AI training.

Answer

Generate synthetic data that mimics patient records.

Answer

Remove patient names and replace with IDs.

Answer

Anonymize data by aggregating into groups of 10 or more.

Question 20

An administrator is configuring Einstein Bots and needs the bot to understand when a customer says 'I want to return a product' and route them accordingly. What must be created to map this phrase to a specific bot action?

Accepted Answer

An intent named 'ReturnProduct' and associated training phrases. Intents represent the purpose of a user input (e.g., 'ReturnProduct'), and entities capture specifics like product name. The bot uses NLP to match intents.

Answer

A new object to store return requests

Answer

A dialogue flow with condition nodes

Answer

A custom entity for 'return'