AI Associate AI Fundamentals Practice Test 2 — 15 Questions

Question 1

Refer to the exhibit. A Salesforce admin configured the Einstein Trust Layer policy shown. What is the effect of this policy on AI model usage?

Accepted Answer

AI models receive masked data for those fields, but feature importance insights are still available.. The Einstein Trust Layer policy configured to mask specific fields ensures that sensitive data is replaced with masked values before being sent to the AI model. This preserves data privacy while still allowing the model to generate predictions and insights. Feature importance insights remain available because they are computed from the masked data, not the original values.

Answer

All fields in the org will be masked to protect customer privacy.

Answer

AI models will not be able to use the configured fields, and model insights are disabled.

Answer

AI models can still use the fields but feature importance insights are blocked.

Question 2

A financial services company is deploying Einstein Prediction Builder to predict customer churn. The data includes both numerical and categorical fields. Which step is essential to ensure the model is not biased against protected attributes like race or gender?

Accepted Answer

Exclude any protected attributes from the training data and ensure the model does not use correlated proxies.. Excluding protected attributes like race or gender from the training data and ensuring the model does not use correlated proxies is essential to prevent bias in Einstein Prediction Builder. This approach directly removes the risk of the model learning discriminatory patterns based on these attributes, as the platform relies on the data provided and does not automatically enforce fairness constraints. Including such attributes or relying on built-in fairness would not guarantee unbiased predictions because the model could still infer protected characteristics from correlated features.

Answer

Include race and gender as predictors to allow the model to adjust for them.

Answer

Rely on the model's built-in fairness constraints.

Answer

Use a deep learning algorithm to automatically handle bias correction.

Question 3

A global retail company with 50,000+ users has deployed Einstein Activity Capture across Sales and Service Clouds. After two weeks, the VP of Sales reports that only 60% of emails sent from Outlook are being logged in Salesforce. Users have installed the Einstein Activity Capture plugin and have the correct permission set. The admin has verified that the email logging settings are enabled for all users. The company uses Exchange Online. What should the admin investigate first?

Accepted Answer

Confirm that the Salesforce connected app is authorized in Exchange Online.. Einstein Activity Capture for Exchange Online requires an OAuth 2.0 connected app authorization in Exchange Online to access mailbox data. Even if the plugin is installed and permissions are set, without the connected app being authorized, the service cannot retrieve email metadata, causing a significant logging gap. This is the first thing to verify since it is a common misconfiguration during initial setup.

Answer

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

Answer

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

Answer

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

Question 4

A company uses Einstein Prediction Builder to create a custom model that predicts whether a support case will be escalated. The model is built and published, but when the admin looks at the case record, the prediction field shows 'No Prediction' for all cases. The prediction is set to run on case creation and update. What should the admin check?

Accepted Answer

The model's confidence threshold is too high, causing no predictions.. When Einstein Prediction Builder shows 'No Prediction' for all cases, a common cause is that the model's confidence threshold is set too high. The prediction field only displays a value when the model's confidence in its prediction exceeds that threshold; if no case meets the threshold, all predictions are suppressed. The admin should lower the confidence threshold in the model settings to allow predictions to appear.

Answer

The prediction field is not added to the case page layout.

Answer

The model was not activated for all record types.

Answer

The data prep steps included all required fields.

Question 5

A sales manager wants to use Einstein Opportunity Scoring to prioritize deals. Which two requirements must be met?

Accepted Answer

There must be at least 50 closed-won opportunities in the last 365 days.. Option A is correct because Einstein Opportunity Scoring requires a sufficient volume of historical data to train its model, specifically at least 50 closed-won opportunities within the last 365 days. Option E is correct because Einstein features must first be enabled in Setup (Einstein > Einstein Features / Settings) before Opportunity Scoring can be turned on and used. Option B is not required for using scoring; the 'Manage Einstein' permission is for administering Einstein features, not for a sales manager to consume scores. Option C is not a prerequisite; field history tracking is not required to generate Opportunity Scores. Option D is incorrect because the requirement is about closed-won historical deals (at least 50), not 500 open opportunities.

Answer

The user must have the 'Manage Einstein' permission.

Answer

Field history tracking must be enabled for Opportunity.

Answer

There must be at least 500 open opportunities.

Question 6

Refer to the exhibit. A Salesforce admin runs the Einstein model list command and sees the output. Which model is currently available for use in predictive scoring?

Accepted Answer

Lead Score Model. Lead Score Model, because the Einstein model list output shows this model as the one currently available for predictive scoring. In Einstein/CRM Analytics predictive scoring, only models that appear as active or available in the model list can be used to generate scores, so the listed Lead Score Model is the usable option. The other options do not fit: A is too broad because not every model is necessarily available, while C and D are specific model names that are not the one shown as available for scoring in the exhibit.

Answer

All models are available for scoring.

Answer

Case Deflection

Answer

Opportunity Win Rate

Question 7

A company has been using Einstein Lead Scoring for six months. Recently, the lead score confidence has dropped from 85% to 60%. The admin reviews the model and finds that many leads have missing data in custom fields used by the model. The admin also notices that field history tracking is not enabled on the Lead object. The lead volume is adequate with over 10,000 leads. What should the admin do to improve the model's confidence?

Accepted Answer

Enable field history tracking on the Lead object and retrain the model.. Field history tracking is required for Einstein Lead Scoring to capture changes over time, and retraining after enabling it will incorporate historical data. Option A is incorrect because manually updating all leads is impractical and doesn't address the root cause. Option C is incorrect because the lead volume is already adequate. Option D is incorrect because disabling and re-enabling will reset the model but not fix the missing field history.

Answer

Disable and re-enable Einstein Lead Scoring.

Answer

Increase the lead volume to at least 50,000 records.

Answer

Manually update all leads with missing data to have complete records.

Question 8

A customer support center wants to automatically route incoming cases to the appropriate department based on the issue description. Which NLP task is most relevant?

Accepted Answer

Text classification. Text classification assigns an incoming text to one or more predefined categories — here, routing a support case to the correct department means classifying the issue description into a department label. Intent detection is a specialized form of classification used in conversational systems, but for routing free-text case descriptions to departments, text classification is the broader and more accurate task.

Answer

Intent detection

Answer

Entity extraction

Answer

Sentiment analysis

Question 9

A retailer wants to use computer vision to automatically identify products from images uploaded by customers for a return process. Which computer vision task is required?

Accepted Answer

Product recognition. Product recognition is the computer vision task specifically designed to identify and classify products from images, often using a combination of object detection and classification tailored to a product catalogue. In a returns process, the system must recognise which product the customer is returning from an uploaded photo, which requires matching the image to a known product SKU. This goes beyond generic object detection or image classification because it involves fine-grained recognition against a product database.

Answer

Object detection

Answer

Image classification

Answer

Text extraction (OCR)

Question 10

A company is building a sentiment analysis model for customer reviews. They want to measure its performance. Which TWO metrics are most appropriate for evaluating a classification model?

Accepted Answer

Recall. Recall (D) is correct because it measures the proportion of actual positive cases (e.g., positive or negative reviews) that the model correctly identified, which is essential for understanding how well the sentiment classifier captures all relevant instances. Precision (E) is correct because it measures the proportion of predicted positive cases that are truly positive, which is critical when false positives have a cost, such as mislabeling a negative review as positive. Together, precision and recall provide a balanced view of a classification model's performance, especially for imbalanced sentiment datasets. F1 score (A) is also a classification metric, but it is not marked as correct here; it combines precision and recall into a single value. R-squared (B) and mean squared error (C) are regression metrics, not appropriate for evaluating a classification model like sentiment analysis.

Answer

F1 score

Answer

R-squared

Answer

Mean squared error

Question 11

A company wants to generate personalized product recommendations for each customer based on their purchase history and browsing behavior. Which approach is MOST appropriate?

Accepted Answer

Train a supervised learning model on historical purchase data to predict the next product a customer will buy. Training a supervised learning model on historical purchase data to predict the next product is the most appropriate approach because it directly leverages past behavior to make personalized recommendations. Supervised learning is well-suited for this task as it can learn patterns from labeled data (e.g., which products were purchased after certain browsing sequences) and generalize to new customers. This approach is standard in recommendation systems and can be implemented with algorithms like collaborative filtering or gradient boosting.

Answer

Use a generative AI model to create new product descriptions for each customer

Answer

Deploy a reinforcement learning agent that explores different recommendations in real-time

Answer

Use an unsupervised learning algorithm to cluster customers and recommend popular items in each cluster

Question 12

A team trains a model to predict customer lifetime value (CLV) using CRM data. The model's predictions are way off for new customers who have only been with the company for a month. Which factor is MOST likely contributing to this issue?

Accepted Answer

The model is overfitting to customers with long purchase histories. The model is overfitting to customers with long purchase histories because the training data is dominated by established customers whose CLV patterns are well-represented. New customers with only one month of history have very different behavioral signals, and the model has not learned to generalize to this sparse-data regime — it essentially memorized patterns from long-tenured customers. This is a classic distribution-shift/overfitting problem where the model performs poorly on underrepresented subgroups.

Answer

The model has high precision but low recall

Answer

The model is underfitting

Answer

The training data includes future transactions (data leakage)

Question 13

A company wants to use AI to automatically route customer support emails to the appropriate department (billing, technical, sales). Which THREE AI capabilities are needed? (Choose three.)

Accepted Answer

Intent detection. Intent detection (A) is required because the system must determine what the customer actually wants (e.g., a refund request versus a technical fault), which drives routing to billing, technical, or sales. Entity extraction (B) is needed to pull out key structured details such as account numbers, product names, or invoice IDs that help disambiguate and enrich the routing decision. Text classification (D) is the core capability that assigns the email to one of the predefined department categories (billing, technical, sales), making it essential for automated routing. Sentiment analysis (C) only measures emotional tone and does not determine the correct department, so it is not required. Generative AI (E) is about producing new content such as replies or summaries, not about classifying and routing incoming messages, so it is not needed here.

Answer

Sentiment analysis

Answer

Generative AI

Question 14

What is the primary difference between narrow AI and general AI?

Accepted Answer

Narrow AI is task-specific; general AI can perform any intellectual task a human can. Narrow AI (also called weak AI) is designed and trained for a specific task — such as recommending products, recognizing faces, or predicting churn — and cannot transfer its capabilities outside that domain. General AI (AGI) refers to a hypothetical system that can understand, learn, and apply intelligence across any intellectual task a human can perform, with human-like adaptability. This task-specific vs. task-agnostic distinction is the defining boundary between the two categories.

Answer

Narrow AI can perform any intellectual task; general AI is limited to one task

Answer

Narrow AI uses machine learning; general AI uses deep learning

Answer

General AI is already deployed in CRM systems; narrow AI is experimental

Question 15

A data scientist trains a model to predict customer churn. The model performs well on training data but poorly on test data. Which TWO issues are most likely?

Accepted Answer

Overfitting. A model that performs well on training data but poorly on test data is the classic signature of overfitting, where the model has memorized noise and specific patterns in the training set rather than learning generalizable relationships. Option E (Data leakage) is also correct because leakage occurs when information from the test set (or target) inadvertently influences training, producing artificially strong training performance that does not hold up on genuinely unseen test data. Option B (Insufficient training data) is not the most likely issue here, since too little data typically causes poor performance on both training and test sets rather than a large train-test gap. Option C (High bias) and Option D (Underfitting) are incorrect because both describe models that perform poorly on training data as well, which contradicts the scenario of strong training performance.

Answer

Insufficient training data

Answer

High bias

Answer

Underfitting