Knowledge + Practice

CCNA AI Fundamentals Questions

75 of 163 questions · Page 1/3 · AI Fundamentals · Answers revealed

Practice these questions Domain overview All questions

1

Multi-Selecthard

A financial services firm wants to use AI to detect fraudulent transactions. They have a dataset with 1% fraudulent and 99% legitimate transactions. Which THREE actions should they take to address class imbalance? (Select THREE.)

Select 3 answers

A.Use a larger dataset even if it is more imbalanced

B.Use accuracy as the primary evaluation metric

C.Use F1 score as the evaluation metric

D.Assign higher class weights to the minority class during training

E.Oversample the minority class (fraud) or undersample the majority class

AnswersC, D, E

F1 score balances precision and recall and is robust to imbalance.

Why this answer

Class imbalance can be addressed by resampling (oversampling minority, undersampling majority), using class weights, or using metrics like F1 that consider precision and recall.

Practice this question →

2

MCQeasy

A nonprofit uses Einstein Vision to classify images of disaster areas. What is the primary benefit of using AI for this task?

A.It requires less training data than manual methods.

B.It eliminates all classification errors.

C.It reduces manual effort and speeds up damage assessment.

D.It can only classify images of specific disaster types.

AnswerC

Automation increases efficiency.

Why this answer

Einstein Vision automates the classification of disaster images, significantly reducing the manual effort required for damage assessment. By processing large volumes of images rapidly, it accelerates the time to insight, enabling faster response and resource allocation. This aligns with the core benefit of AI: augmenting human effort with speed and scale.

Exam trap

Salesforce often tests the misconception that AI eliminates errors entirely, when in reality AI systems have accuracy limitations and require human oversight for critical decisions.

How to eliminate wrong answers

Option A is wrong because AI models like Einstein Vision typically require large, labeled training datasets to achieve accuracy, not less data than manual methods. Option B is wrong because no AI system eliminates all classification errors; models have inherent accuracy limitations and can misclassify edge cases. Option D is wrong because Einstein Vision is not limited to specific disaster types; it can be trained to classify a wide range of disaster scenarios and image categories.

Practice this question →

3

MCQhard

During an AI ethics review, a stakeholder asks how Salesforce ensures that Einstein models do not discriminate based on protected attributes. Which mechanism addresses this concern?

A.Remove all protected attribute fields from the training dataset

B.Rely on model accuracy metrics to ensure equal treatment

C.Use the bias detection and fairness reporting built into Einstein Prediction Builder

D.Allow admins to manually override model predictions for certain groups

AnswerC

Salesforce provides tools to detect and report bias, enabling proactive fairness assessment.

Why this answer

Salesforce Einstein Prediction Builder includes built-in bias detection and fairness reporting that automatically checks models for discrimination based on protected attributes. This feature analyzes model predictions against demographic groups and generates fairness metrics, allowing stakeholders to identify and mitigate bias directly within the platform.

Exam trap

Salesforce often tests the misconception that removing protected attributes from training data is sufficient to prevent bias, but the trap here is that proxy discrimination through correlated features remains undetected, making bias detection tools like Einstein’s fairness reporting the correct answer.

How to eliminate wrong answers

Option A is wrong because simply removing protected attribute fields from the training dataset does not prevent proxy discrimination—other correlated features (e.g., zip code, income) can still encode bias against protected groups, and this approach can actually hide bias rather than eliminate it. Option B is wrong because model accuracy metrics (e.g., overall accuracy, F1 score) do not measure fairness across subgroups; a model can be highly accurate overall yet systematically discriminate against a protected class (e.g., high false positive rate for one demographic). Option D is wrong because allowing admins to manually override model predictions for certain groups introduces subjective, inconsistent, and potentially biased decision-making, and it does not provide systematic, auditable fairness checks—it is a reactive workaround, not a proactive mechanism.

Practice this question →

4

MCQhard

Refer to the exhibit. A Salesforce admin is troubleshooting email capture failures. Based on the log, which configuration step is most likely missing?

A.The connected email service is not authenticated for the organization.

B.The email domain 'acme.org' needs to be added to the Einstein Activity Capture allowed domains.

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

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

AnswerB

Domain not configured causes failures.

Why this answer

The log indicates that emails from 'acme.org' are being captured but not linked to contacts, which is a classic symptom of a missing domain in the Einstein Activity Capture allowed domains list. Without adding the domain, the system cannot associate emails from that domain with the correct contact records, even if the email addresses are valid. Option B directly addresses this configuration gap.

Exam trap

Salesforce often tests the distinction between email capture (which requires domain whitelisting) and email authentication (which is about SMTP or OAuth), leading candidates to mistakenly choose authentication issues when the real problem is domain configuration.

How to eliminate wrong answers

Option A is wrong because the log does not show any authentication failure; the email service is successfully connecting and processing emails. Option C is wrong because the log does not indicate that the contact's email address is invalid; the capture failure is due to domain association, not an invalid address. Option D is wrong because there is no mention of quota limits in the log; the error is related to domain configuration, not user-level capture quotas.

Practice this question →

5

Multi-Selectmedium

Which TWO of the following are common causes of model drift in Einstein Discovery?

Select 2 answers

A.Improved data quality after cleaning

B.Seasonal patterns that affect the target variable

C.Increased model complexity

D.Changes in customer behavior over time

E.Reduced size of the training dataset

AnswersB, D

Seasonality can introduce cyclic changes that the model may not capture if not retrained.

Why this answer

Seasonal patterns (Option B) cause model drift because the relationship between input features and the target variable changes predictably over time, such as higher sales during holidays. Einstein Discovery models trained on historical data may fail to generalize if the seasonal cycle is not captured or if the model is not retrained to account for these recurring shifts, leading to degraded prediction accuracy.

Exam trap

Salesforce often tests the distinction between factors that degrade model performance (like poor data quality or overfitting) versus the specific external or temporal changes that cause model drift, leading candidates to mistakenly select options like increased complexity or reduced dataset size.

Practice this question →

6

MCQeasy

A sales rep noticed that the Einstein Lead Scoring prediction bar shows 'No score available' for many leads. The admin confirmed that Einstein Lead Scoring is enabled and the permission set is assigned. What is the most likely cause?

A.The org does not have enough historical data to train the scoring model.

B.The leads have not yet been assigned to a user.

C.The leads were created less than 30 days ago.

D.The lead scoring model is still training.

AnswerA

A minimum number of converted leads is needed for the model to generate scores.

Why this answer

Einstein Lead Scoring requires a minimum amount of historical lead data (typically at least 2,000 converted and 2,000 unconverted leads) to train its predictive model. If the org lacks sufficient historical data, the model cannot generate scores, resulting in 'No score available' for leads. This is the most likely cause because the admin confirmed the feature and permissions are correctly enabled.

Exam trap

Salesforce often tests the misconception that 'No score available' is caused by the model still training or by recent lead creation, when in fact it points to insufficient historical data for model training.

How to eliminate wrong answers

Option B is wrong because lead assignment to a user is not a prerequisite for Einstein Lead Scoring; the model scores leads based on field values and historical patterns, not ownership. Option C is wrong because there is no 30-day age requirement for leads to receive a score; scoring applies to all leads once the model is trained, regardless of creation date. Option D is wrong because if the model were still training, the prediction bar would typically show a 'Training in progress' message, not 'No score available'; the latter indicates the model lacks sufficient data to train at all.

Practice this question →

7

MCQhard

A model trained on CRM data predicts customer lifetime value (CLV) with high accuracy, but when deployed, predictions are significantly off for new customer segments. What is the most likely cause?

A.Feature engineering was insufficient for the original segments

B.The model is overfitting to the training data

C.The model architecture is too simple

D.The training data was not representative of the new segments

AnswerD

Unrepresentative training data leads to poor generalization for unseen segments.

Why this answer

If the training data was not representative of the new segments, the model will not generalize, leading to poor out-of-sample performance.

Practice this question →

8

MCQmedium

A service team trains an Einstein Bot on historical chat transcripts. After deployment, the bot frequently fails to understand customer intents. Which action is most likely to improve performance?

A.Add more diverse training phrases per intent

B.Increase the confidence threshold to 90%

C.Use a hierarchical intent structure

D.Reduce the number of intents to two

AnswerA

Diverse examples improve natural language understanding and reduce failure to recognize intents.

Why this answer

Adding more diverse training phrases per intent directly addresses the root cause of the bot's failure to understand customer intents: insufficient coverage of the varied ways customers express the same goal. Einstein Bot uses natural language understanding (NLU) models that rely on example phrases to learn intent patterns; increasing the diversity of these phrases improves the model's ability to generalize to unseen utterances, reducing misclassification.

Exam trap

Salesforce often tests the misconception that increasing the confidence threshold or reducing intents will improve accuracy, when in fact those actions only mask poor training data or limit the model's scope, rather than fixing the underlying NLU training deficiency.

How to eliminate wrong answers

Option B is wrong because increasing the confidence threshold to 90% would make the bot more conservative, causing it to reject more utterances as 'unknown' rather than improving its understanding of intents; it does not address the lack of training data diversity. Option C is wrong because a hierarchical intent structure organizes intents into parent-child relationships but does not fix the fundamental issue of insufficient or non-diverse training phrases; it can even complicate classification if base intents are poorly trained. Option D is wrong because reducing the number of intents to two would oversimplify the model, likely forcing many distinct customer intents into a single bucket, which increases confusion and degrades performance rather than improving understanding.

Practice this question →

9

MCQeasy

A company wants to use Einstein Activity Capture to log emails and events automatically. Which two considerations should the admin evaluate before enabling this feature?

A.The feature is only available for Sales Cloud.

B.All lead and contact fields must be visible to users.

C.Users must link their email client (Gmail or Outlook).

D.Users must install a browser plugin.

E.Users must have a Salesforce license.

AnswerC, E

Linking is required for activity capture.

Why this answer

Einstein Activity Capture requires users to link their email client (Gmail or Outlook) to Salesforce via OAuth 2.0 authentication. This linkage allows the feature to automatically log emails and events from the connected email and calendar systems without manual user intervention.

Exam trap

The trap here is that candidates often confuse Einstein Activity Capture with Einstein Activity Insights or assume it requires a browser plugin like the Outlook Salesforce add-in, when in fact it uses a server-side OAuth connection.

How to eliminate wrong answers

Option A is wrong because Einstein Activity Capture is available for both Sales Cloud and Service Cloud, not exclusively Sales Cloud. Option B is wrong because while certain fields must be accessible for mapping, not all lead and contact fields need to be visible to users; only the fields used for email-to-Contact/Lead matching (like Email) are required. Option D is wrong because Einstein Activity Capture uses server-side integration via OAuth and does not require users to install any browser plugin; the connection is established through Salesforce's backend.

Practice this question →

10

Multi-Selecthard

A company is developing a sentiment analysis model for customer reviews. The team wants to ensure the model is fair and does not exhibit bias. Which TWO actions are MOST effective? (Choose two.)

Select 2 answers

A.Use a more complex model to capture nuanced patterns

B.Audit the training data for imbalanced representation across demographic groups

C.Use only historical data as-is without modification

D.Ensure the training data includes reviews from a diverse range of customers

E.Remove all negative reviews from the training data to avoid negative bias

AnswersB, D

Identifying and correcting imbalances helps reduce bias.

Why this answer

Auditing training data for bias and ensuring diverse representation helps mitigate bias. Using a complex model doesn't address bias; removing all negative examples would bias the model. Using historical data as-is may perpetuate existing biases.

Practice this question →

11

MCQhard

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?

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

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

C.Confirm that the Salesforce connected app is authorized in Exchange Online.

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

AnswerC

Authorization issues can cause partial logging; checking this first is efficient.

Why this answer

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.

Exam trap

Salesforce often tests the distinction between user-level permissions (like permission sets) and service-level authorization (like connected app consent), leading candidates to focus on user settings when the real issue is a missing OAuth authorization between the two systems.

How to eliminate wrong answers

Option A is wrong because Exchange Online mailbox audit logs track user actions like read or delete, not the system-level access by Einstein Activity Capture; investigating those logs would not reveal why emails are not being captured. Option B is wrong because the email synchronization frequency in Einstein Activity Capture settings controls how often sync occurs, but if the connection is not authorized, no sync will happen regardless of frequency. Option D is wrong because the 'Email Integration' permission is a legacy setting for older email integrations, not required for Einstein Activity Capture, which relies on the correct permission set and connected app authorization.

Practice this question →

12

MCQmedium

A Salesforce admin wants to use Einstein GPT to generate personalized email content for a marketing campaign. To ensure the AI does not produce responses that include sensitive customer data or violate company policies, which Salesforce feature should the admin configure?

A.Prompt Builder

B.Data Cloud

C.Einstein Studio

D.Einstein Trust Layer

AnswerD

Einstein Trust Layer provides data masking, toxicity detection, and adherence to privacy policies for AI-generated content.

Why this answer

Einstein Trust Layer is the correct feature because it acts as a governance and security layer between Salesforce and the large language model (LLM). It automatically masks sensitive customer data (e.g., personally identifiable information) before the prompt is sent to the LLM and then unmasks the response, ensuring the AI never sees or exposes sensitive information. This directly addresses the admin's need to prevent responses containing sensitive data or violating company policies.

Exam trap

The trap here is that candidates often confuse Prompt Builder (which controls the prompt content) with the Trust Layer (which controls data security), assuming that defining strict prompts alone is sufficient to prevent sensitive data leakage, when in fact the Trust Layer's automated masking is required for true data protection.

How to eliminate wrong answers

Option A is wrong because Prompt Builder is a tool for creating and managing prompt templates that define the structure and context of AI-generated content, but it does not include built-in data masking or policy enforcement to prevent sensitive data leakage. Option B is wrong because Data Cloud is a customer data platform that unifies data from various sources for analytics and segmentation, not a feature for governing AI-generated outputs or masking sensitive data during LLM interactions. Option C is wrong because Einstein Studio is a low-code environment for building custom AI models and pipelines, but it does not provide the automatic data masking and safety controls that the Einstein Trust Layer offers for generative AI responses.

Practice this question →

13

Multi-Selectmedium

A retail company wants to use AI to predict next month's sales for each product category. They have five years of monthly sales data. Which THREE factors are most critical for the accuracy of the predictive model? (Select THREE.)

Select 3 answers

A.Choosing the most popular AI algorithm on the market

B.Including seasonality and trend components from historical data

C.Using a deep learning model instead of a simpler model

D.Ensuring the training data is clean and free of missing values

E.Incorporating external factors like holiday calendars and promotions

AnswersB, D, E

Critical for capturing recurring patterns.

Why this answer

Historical patterns (seasonality), external factors (holidays), and data quality are key for time series forecasting. Model complexity and algorithm brand are less critical.

Practice this question →

14

MCQhard

Refer to the exhibit. A Salesforce CLI output shows the status of Einstein models in the org. Which model should the administrator investigate first?

A.Lead_Score_Model because its accuracy is lower than expected.

B.Campaign_Response because it has the oldest training date.

C.Opportunity_Forecast because it is still training.

D.Case_Escalation because it has an Error status.

AnswerD

Error models require troubleshooting.

Why this answer

An Einstein model with an 'Error' status indicates a critical failure that prevents the model from generating predictions or scoring records. This requires immediate investigation to restore functionality, as the model is non-operational and may impact business processes relying on its output.

Exam trap

Salesforce often tests the distinction between a model's operational status (e.g., Error) and its performance metrics (e.g., accuracy), leading candidates to mistakenly prioritize accuracy concerns over a non-functional model.

How to eliminate wrong answers

Option A is wrong because 'accuracy lower than expected' is a performance metric, not an immediate operational issue; Einstein models can have varying accuracy based on data quality and configuration, and this alone does not warrant priority over a non-functional model. Option B is wrong because the oldest training date does not inherently indicate a problem; models can be retrained on demand, and age alone is not a sign of failure or urgency. Option C is wrong because a model that is 'still training' is in a normal state; Einstein models require training time, and this status is expected during the learning phase, not an error condition.

Practice this question →

15

MCQeasy

Refer to the exhibit. A Salesforce admin evaluates an Einstein Prediction Builder model for customer churn. What should be the admin's primary concern based on the exhibit?

A.The model accuracy is too low for production use.

B.The model has detected data drift, indicating the training data may no longer represent current patterns.

C.The AUC is low, so the model is not better than random.

D.The model uses too few features to be reliable.

AnswerB

Data drift makes predictions unreliable.

Why this answer

The exhibit shows a data drift alert from Einstein Prediction Builder, which indicates that the statistical properties of the input data have changed over time. This is the admin's primary concern because a model trained on outdated patterns will produce unreliable predictions, even if its accuracy or AUC were initially high. Data drift directly undermines the model's validity in production.

Exam trap

Salesforce often tests the distinction between model performance metrics (accuracy, AUC) and model health indicators (data drift), leading candidates to focus on missing or irrelevant metrics instead of the explicit alert shown.

How to eliminate wrong answers

Option A is wrong because the exhibit does not display an accuracy metric; the alert shown is specifically for data drift, not low accuracy. Option C is wrong because the exhibit does not show an AUC value, and a low AUC would indicate poor discriminative power, but the primary issue here is data drift, not AUC. Option D is wrong because the number of features is not indicated in the exhibit, and data drift can occur regardless of feature count; the concern is about the distribution of existing features, not their quantity.

Practice this question →

16

MCQeasy

A marketing team wants to recommend products to customers based on their past purchases and browsing behavior. Which type of AI is most appropriate?

A.Generative AI

B.Reinforcement learning

C.Predictive AI

D.Unsupervised learning

AnswerC

Correct: predictive AI uses historical patterns to forecast future behavior, e.g., product recommendations.

Why this answer

Product recommendation is a predictive AI task that uses historical data to predict which products a customer is likely to purchase.

Practice this question →

17

MCQmedium

A marketing director wants to use Einstein Engagement Scoring to prioritize leads. She has enabled Einstein and assigned the permission set to users. However, the Engagement Score field is not visible on any lead record. The admin checked the field-level security and it is visible to all profiles. What should the admin do next?

A.Add the Engagement Score field to the lead page layout.

B.Verify that there are at least 500 leads with activity in the last 30 days.

C.Run the 'Calculate Einstein Engagement Scores' scheduled job.

D.Wait 24 hours for the model to train.

AnswerA

Field visibility requires being on the page layout.

Why this answer

The Engagement Score field is a standard field that must be added to the lead page layout to be visible on the record. Even though field-level security grants access, the field will not appear on the record detail page unless it is explicitly placed on the page layout. This is a common layout-level visibility requirement in Salesforce.

Exam trap

The trap here is that candidates confuse field-level security with page layout visibility, assuming that enabling FLS automatically makes the field appear on the record, when in fact both settings must be configured independently.

How to eliminate wrong answers

Option B is wrong because the 500-leads-with-activity threshold is a prerequisite for the Einstein Engagement Scoring model to train, not a cause for the field not being visible on the record. Option C is wrong because the 'Calculate Einstein Engagement Scores' scheduled job is used to trigger scoring calculations, but the field must already be on the layout to display the results; running the job does not make the field appear. Option D is wrong because waiting 24 hours addresses model training time, not the layout visibility issue; the field will remain hidden regardless of training completion if it is not on the layout.

Practice this question →

18

MCQeasy

Which statement best describes 'inference' in the context of machine learning?

A.The process of training a model on labeled data

B.The process of collecting and preparing data

C.The process of using a trained model to make predictions on new data

D.The process of evaluating a model's accuracy

AnswerC

Inference is applying the model to new, unseen data.

Why this answer

Inference is the process of using a trained model to make predictions on new data. Training is the learning phase, and evaluation is assessing performance. Data collection is separate.

Practice this question →

19

Multi-Selectmedium

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?

Select 2 answers

A.F1 score

B.R-squared

C.Mean squared error

D.Recall

E.Precision

AnswersD, E

Measures coverage of actual positives.

Why this answer

Precision and recall are key classification metrics. Accuracy can be misleading if classes are imbalanced. F1 is composite; but the question asks for two metrics, precision and recall are fundamental.

Practice this question →

20

MCQmedium

A company wants to use AI to automatically extract key information (e.g., invoice number, date, total amount) from scanned PDF invoices. Which AI capability should they use?

A.Generative AI for text generation

B.Predictive analytics

C.Computer vision with optical character recognition (OCR)

D.Sentiment analysis

AnswerC

Correct: OCR converts scanned images to text, then NLP extracts fields.

Why this answer

Document scanning combined with OCR and NLP entity extraction can parse structured fields from documents. This falls under computer vision (OCR) and NLP.

Practice this question →

21

MCQeasy

A company wants to use Einstein Discovery to analyze sales data and automatically uncover key drivers of deal closure. What must the admin provide to create a story?

A.At least one numeric field to predict

B.A date field for time series analysis

C.A foreign key to relate objects

D.A text field for sentiment analysis

AnswerA

Discovery predicts numeric values or binary outcomes; a numeric target is required.

Why this answer

Einstein Discovery requires at least one numeric field as the prediction target (e.g., deal amount, probability score) to train its regression or classification model. Without a numeric field to predict, the story cannot define what outcome the AI should analyze or uncover key drivers for.

Exam trap

Salesforce often tests the misconception that Einstein Discovery requires a date field for time series or a foreign key for relational data, when in fact the core requirement is a numeric field to define the prediction target.

How to eliminate wrong answers

Option B is wrong because a date field is optional for time series analysis but not mandatory; Einstein Discovery can create stories without any temporal component. Option C is wrong because a foreign key is not required; Einstein Discovery works on a single object or dataset and does not need relational joins to build a story. Option D is wrong because a text field for sentiment analysis is not a prerequisite; Einstein Discovery focuses on structured numeric and categorical fields, not unstructured text analysis.

Practice this question →

22

MCQhard

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?

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

B.The model was not activated for all record types.

C.The model's confidence threshold is too high, causing no predictions.

D.The data prep steps included all required fields.

AnswerC

A high threshold means even correct predictions may be suppressed.

Why this answer

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.

Exam trap

Salesforce often tests the misconception that 'No Prediction' is caused by missing page layout fields or record type activation, but the actual cause is the confidence threshold filtering out all predictions.

How to eliminate wrong answers

Option A is wrong because if the prediction field were not on the page layout, the field would not appear at all on the case record, rather than showing 'No Prediction'. Option B is wrong because activation for record types is not a setting in Einstein Prediction Builder; models are applied globally or by object, not per record type. Option D is wrong because data prep steps including all required fields are necessary for model training, but if the model is already built and published, missing fields would have caused an error during training, not a 'No Prediction' result on existing records.

Practice this question →

23

MCQeasy

A data analyst is preparing data for a machine learning model. They notice that many records have missing values for the 'industry' field. What is the best first step?

A.Delete all records with missing values

B.Replace missing values with a placeholder or impute them

C.Ignore the missing values and train the model as is

D.Use a model that automatically handles missing data

AnswerB

Correct. Imputation or placeholder ensures data completeness.

Why this answer

Data quality is critical. Missing values must be handled (imputed or removed) to avoid garbage-in-garbage-out.

Practice this question →

24

MCQmedium

A retail company implements an AI chatbot to recommend products. After launch, they notice the chatbot frequently suggests expensive items to budget-conscious customers. Which AI bias is most likely occurring?

A.Confirmation bias

B.Anchoring bias

C.Sample bias (biased training data)

D.Overconfidence bias

AnswerC

If the training data overrepresents high-spending customers, the model may learn to recommend expensive products to all users.

Why this answer

The chatbot's frequent suggestion of expensive items to budget-conscious customers indicates that the training data was biased toward high-cost products, leading the model to learn and replicate that preference. This is a classic case of sample bias (biased training data), where the dataset does not accurately represent the target user population, causing systematic errors in the model's recommendations.

Exam trap

Salesforce often tests the distinction between human cognitive biases (like anchoring or confirmation bias) and data-driven biases (like sample bias), so the trap here is that candidates confuse a human reasoning flaw with a machine learning training data issue.

How to eliminate wrong answers

Option A is wrong because confirmation bias refers to a human tendency to favor information that confirms preexisting beliefs, not a data-driven AI model's output skew. Option B is wrong because anchoring bias is a cognitive heuristic where humans rely too heavily on the first piece of information encountered, not a bias originating from training data distribution. Option D is wrong because overconfidence bias relates to a model or human assigning excessive certainty to predictions, not to a systematic skew in recommendation outputs due to imbalanced training data.

Practice this question →

25

Multi-Selectmedium

A company wants to use AI to analyze customer feedback from surveys and social media. Which TWO capabilities are most relevant?

Select 2 answers

A.Speech recognition

B.Sentiment analysis

C.Image recognition

D.Text classification

E.Predictive lead scoring

AnswersB, D

Sentiment analysis measures emotional tone, key for understanding customer feelings.

Why this answer

Sentiment analysis detects overall sentiment (positive/negative), and text classification can categorize feedback by topic. Entity extraction might be useful but is less central for broad feedback analysis.

Practice this question →

26

MCQeasy

A marketing team wants to use Einstein Engagement Scoring to prioritize leads. What is the primary input for this AI feature?

A.Lead interaction history with emails and web activity.

B.Historical conversion data from closed opportunities.

C.Lead demographic information like industry and company size.

D.Social media posts and mentions of the company.

AnswerA

Engagement is measured by interactions.

Why this answer

Einstein Engagement Scoring analyzes lead interactions (email opens, clicks, web visits) to calculate engagement scores. Option A is correct. Option B is wrong because demographic data is not the primary input.

Option C is wrong because historical conversion data is used for predictive scoring, not engagement. Option D is wrong because social media data is not a direct input.

Practice this question →

27

MCQeasy

A sales manager wants to implement Einstein Automated Contacts to automatically create contacts from email interactions. The admin enables the feature and assigns the permission set. However, no contacts are being created automatically. What is the most likely reason?

A.The users have not logged any emails to Salesforce.

B.The admin did not set up the contact creation criteria in the Einstein Automated Contacts settings.

C.The feature requires at least 500 emails to be processed before it starts creating contacts.

D.The users need to install a plugin for their email client.

AnswerB

Rules must be configured; default is no creation.

Why this answer

Einstein Automated Contacts requires the admin to explicitly define the criteria that trigger automatic contact creation from email interactions. Simply enabling the feature and assigning the permission set does not configure the rules for when and how contacts should be created. Without these criteria, the system has no instructions to act upon, so no contacts are generated.

Exam trap

Salesforce often tests the misconception that enabling a feature and assigning permissions is sufficient for it to work, when in reality, configuration of specific rules or criteria is required to trigger the intended behavior.

How to eliminate wrong answers

Option A is wrong because the feature processes emails that are already logged to Salesforce; if users have not logged any emails, there would be no data to process, but the question states no contacts are being created, implying the feature is not triggering even when emails are present. Option C is wrong because Einstein Automated Contacts does not require a minimum number of emails to be processed before it starts creating contacts; it works on individual email interactions once criteria are set. Option D is wrong because Einstein Automated Contacts is a native Salesforce feature that works with the standard email integration (e.g., Salesforce for Outlook or Gmail) and does not require a separate plugin installation.

Practice this question →

28

MCQmedium

A data scientist is training a model to predict churn. The model achieves 99% accuracy on training data but only 60% on test data. Which issue is most likely occurring?

A.Concept drift

B.Overfitting

C.Data leakage

D.Underfitting

AnswerB

Correct. Large gap between training and test performance indicates overfitting.

Why this answer

Overfitting: the model learns training data patterns too well, including noise, failing to generalize to new data.

Practice this question →

29

Multi-Selectmedium

A sales team uses an AI model to prioritize leads. The model's predictions are not improving despite adding more data. Which THREE factors could explain this? (Choose three.)

Select 3 answers

A.The new data is noisy or of low quality

B.The model is underfitting and lacks capacity

C.The model is already overfitting to the training data

D.The features used are not predictive of lead conversion

E.The team has not added enough features to the model

AnswersA, B, D

Garbage in, garbage out; low-quality data hinders improvement.

Why this answer

Poor data quality, irrelevant features, and model underfitting can all cause lack of improvement. Adding more data helps only if it's high-quality; more features can help if relevant, but not necessarily. Overfitting would show good training performance.

Practice this question →

30

MCQeasy

Which type of machine learning is used to predict customer churn based on historical labeled data?

A.Reinforcement learning

B.Unsupervised learning

C.Supervised learning

D.Self-supervised learning

AnswerC

Supervised learning trains on labeled examples to predict outcomes, making it ideal for churn prediction.

Why this answer

Supervised learning uses labeled data (historical churn outcomes) to train a model to predict future churn.

Practice this question →

31

MCQeasy

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?

A.Disable and re-enable Einstein Lead Scoring.

B.Enable field history tracking on the Lead object and retrain the model.

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

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

AnswerB

Field history tracking provides the necessary historical data for scoring accuracy.

Why this answer

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.

Practice this question →

32

MCQhard

A data scientist trains a lead scoring model that achieves 99% accuracy on training data but only 65% accuracy on a held-out test set. What is the most likely issue?

A.Underfitting due to insufficient model complexity

B.Overfitting due to model complexity or insufficient regularization

C.Label noise in the training data

D.Data leakage from the test set into training

AnswerB

Overfitting explains the large gap between high training accuracy and low test accuracy.

Why this answer

Overfitting occurs when the model memorizes training data noise instead of learning generalizable patterns, leading to poor test performance.

Practice this question →

33

Multi-Selectmedium

Which TWO actions are recommended when preparing data for an Einstein Prediction Builder model?

Select 2 answers

A.Ensure the data set contains at least 500 records with the outcome field populated.

B.Include all available fields on the object, even if unrelated.

C.Use external data sources and upload CSV files without any preprocessing.

D.Select fields that are logically related to the prediction outcome.

E.Include data from the last two days only for the most current trends.

AnswersA, D

Minimum sample size is required for model training.

Why this answer

Einstein Prediction Builder requires a minimum of 500 records with the outcome field populated to ensure statistical significance and reliable model training. Fewer records can lead to overfitting or insufficient pattern recognition, making the model less accurate.

Exam trap

Salesforce often tests the misconception that more data (all fields) or recent data only is always better, but the trap here is that Einstein Prediction Builder requires a minimum record threshold and logically relevant features to avoid noise and ensure model validity.

Practice this question →

34

Multi-Selecthard

A company uses AI to generate personalized email content for marketing campaigns. They notice the AI occasionally produces factually incorrect statements. Which THREE actions should they take to mitigate this?

Select 3 answers

A.Implement retrieval-augmented generation (RAG) to ground outputs in verified data

B.Reduce the model's temperature parameter to lower randomness

C.Include a human review step before emails are sent

D.Train the model on more unlabeled emails from the internet

E.Remove all safety guardrails to allow more creative content

AnswersA, B, C

RAG provides factual context, reducing hallucinations.

Why this answer

Grounding the model with retrieval, adding a human review step, and adjusting temperature can reduce hallucinations. Training on more data alone may not fix the issue.

Practice this question →

35

Multi-Selectmedium

A CRM administrator is planning to implement predictive AI for lead scoring. Which TWO actions should be taken to ensure data quality?

Select 2 answers

A.Ensure lead source values are standardized (e.g., 'Web' not 'website' or 'web form')

B.Remove duplicate leads and leads older than 5 years with no activity

C.Use as many fields as possible to increase model complexity

D.Manually label all historical leads again to remove bias

E.Use a smaller dataset to train faster

AnswersA, B

Consistent categorical values prevent the model from learning spurious patterns.

Why this answer

Removing outdated or duplicate records and ensuring consistent formatting across fields are key data quality steps. Standardizing lead source values also ensures consistent labels.

Practice this question →

36

MCQmedium

A sales operations manager wants to use AI to predict which leads are most likely to convert. The CRM has historical data on past leads, including whether they were won or lost, along with demographic and behavioral attributes. Which machine learning type should be used?

A.Generative AI

B.Supervised learning

C.Unsupervised learning

D.Reinforcement learning

AnswerB

Supervised learning trains on labeled data (outcomes) to predict new outcomes, ideal for lead scoring.

Why this answer

Supervised learning uses labeled historical data (won/lost outcomes) to predict future outcomes, making it the best fit for lead scoring.

Practice this question →

37

MCQhard

A company uses Einstein Bots to handle customer service inquiries. The bot often fails to understand complex requests, leading to escalations. Which improvement strategy is most effective?

A.Train the bot with additional intents and example phrases for complex scenarios.

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

C.Increase the confidence threshold for intent matching to avoid misclassification.

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

AnswerA

More training data improves NLU accuracy.

Why this answer

Training the bot with additional intents and example phrases directly addresses the root cause of the bot's failure: insufficient training data for complex scenarios. By expanding the training corpus, the natural language understanding (NLU) model can better recognize and classify nuanced user inputs, reducing misclassifications and unnecessary escalations.

Exam trap

Salesforce often tests the misconception that increasing confidence thresholds or simplifying logic improves accuracy, when in fact these actions reduce the bot's ability to handle complex inputs, leading to more escalations.

How to eliminate wrong answers

Option B is wrong because routing all complex requests to human agents without bot interaction bypasses the bot entirely, failing to improve its capability and defeating the purpose of using an AI bot to handle escalations. Option C is wrong because increasing the confidence threshold for intent matching would cause the bot to reject more queries, leading to even more false negatives and escalations, not fewer. Option D is wrong because reducing dialogue options simplifies the bot's logic but does not improve its understanding of complex requests; it may actually increase escalations by limiting the bot's ability to handle varied inputs.

Practice this question →

38

MCQmedium

A nonprofit organization uses Salesforce to manage donor relationships. They have implemented Einstein Prediction Builder to predict which donors are likely to upgrade their donation level in the next 90 days. The model was built using a custom object "Donation" with fields like Amount, Frequency, and Campaign. After deployment, the predictions seem random and do not correlate with donor engagement. The admin suspects the model is not trained on enough records. The organization has 500 donors with at least two donations each. What should the admin do to improve the model?

A.Increase the prediction window from 90 to 180 days to capture more upgrade events.

B.Use a different field as the prediction outcome, such as 'donation amount increase'.

C.Ensure that at least 500 records exist where the donor actually upgraded, and retrain the model.

D.Add more fields to the model, such as donor age and geographic location.

AnswerC

Sufficient positive examples are needed.

Why this answer

Einstein Prediction Builder requires a minimum number of positive outcome records (upgrade events) to train a reliable model. With only 500 donors and likely far fewer upgrades, the model lacks sufficient signal. Ensuring at least 500 actual upgrade records provides the necessary positive examples for the algorithm to learn meaningful patterns, reducing randomness in predictions.

Exam trap

Salesforce often tests the misconception that adding more data fields or changing the prediction window can compensate for a lack of positive training records, when in fact the core requirement is a sufficient number of outcome examples for the model to learn from.

How to eliminate wrong answers

Option A is wrong because increasing the prediction window does not address the root cause of insufficient positive training records; it may dilute the signal by including more non-upgrade events. Option B is wrong because changing the prediction outcome field does not solve the data scarcity issue; the model still needs enough historical upgrade events to learn from. Option D is wrong because adding more fields without sufficient positive records will not improve model accuracy and may introduce noise, as the algorithm still lacks enough examples to identify meaningful correlations.

Practice this question →

39

MCQhard

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?

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

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

C.Limit the data used for training to only essential fields.

D.Use Einstein Trust Layer features to mask personally identifiable information (PII) in the model.

AnswerD

Trust Layer masks PII so the model does not see raw sensitive data.

Why this answer

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.

Exam trap

Salesforce often tests the distinction between procedural compliance steps (like obtaining consent) and technical enforcement mechanisms (like the Einstein Trust Layer), leading candidates to choose Option B because it sounds correct in a general privacy context, but the question specifically asks about deploying the model, where a platform-native feature is the correct answer.

How to eliminate wrong answers

Option A is wrong because storing sensitive data in an external data lake and connecting via APIs does not inherently address privacy compliance; it merely shifts the storage location and still requires proper governance, encryption, and consent mechanisms to meet regulations. Option B is wrong because while obtaining explicit consent is a fundamental privacy practice, it is a procedural step, not a technical feature of the Einstein AI platform, and the question asks for a practice to follow when deploying the model, implying a built-in technical solution. Option C is wrong because limiting data to essential fields reduces exposure but does not guarantee compliance; sensitive fields may still be included, and without masking or anonymization, the model could inadvertently memorize and leak PII, violating privacy laws.

Practice this question →

40

MCQhard

A company uses generative AI to create personalized email content for each customer. They notice that occasionally the AI produces content that is factually incorrect. What is this phenomenon called?

A.Hallucination

B.Overfitting

C.Concept drift

D.Bias

AnswerA

Correct. Hallucination is the generation of false information by an AI.

Why this answer

Hallucination occurs when generative models produce plausible-sounding but incorrect information.

Practice this question →

41

MCQeasy

A nonprofit organization wants to use Einstein Bots to handle inquiries on their website. They are concerned that the bot may give incorrect or insensitive responses. Which feature should they prioritize to maintain trustworthy AI?

A.Use a larger training dataset from generic internet sources.

B.Allow the bot to generate responses only from pre-authored articles.

C.Implement a fallback to human agent for uncertain queries.

D.Disable sentiment analysis to avoid misinterpretation.

AnswerC

Fallback improves trust by human handling.

Why this answer

Implementing a fallback to a human agent for uncertain queries directly addresses the risk of incorrect or insensitive responses by ensuring that when the bot cannot confidently answer, the conversation is escalated to a human. This aligns with the principle of maintaining trustworthy AI, as it prevents the bot from generating potentially harmful or inaccurate responses and provides a safety net for complex or sensitive inquiries.

Exam trap

The trap here is that candidates may confuse 'trustworthy AI' with 'restricting the bot's knowledge' (Option B) or 'disabling features' (Option D), rather than recognizing that a fallback mechanism is the standard industry practice for handling uncertainty and maintaining safety in conversational AI systems.

How to eliminate wrong answers

Option A is wrong because using a larger training dataset from generic internet sources does not guarantee accuracy or sensitivity; it may introduce noise, bias, or irrelevant information, and does not address the specific concern about incorrect or insensitive responses. Option B is wrong because allowing the bot to generate responses only from pre-authored articles restricts the bot's ability to handle dynamic or nuanced inquiries, and does not provide a mechanism for handling queries that fall outside the pre-authored content, potentially leading to unhelpful or inappropriate responses. Option D is wrong because disabling sentiment analysis removes the bot's ability to detect and appropriately respond to user emotions or tone, which could actually increase the risk of insensitive responses rather than reduce it.

Practice this question →

42

MCQeasy

A marketing manager wants to predict which customers are most likely to respond to a new email campaign. Which type of machine learning is most appropriate?

A.Generative AI

B.Supervised learning

C.Unsupervised learning

D.Reinforcement learning

AnswerB

Correct. Uses labeled data (responded/did not respond) to train a model.

Why this answer

Supervised learning uses labeled historical data (e.g., past response data) to predict a specific outcome (respond vs. not respond).

Practice this question →

43

MCQeasy

A company wants to use Einstein Prediction Builder to predict customer churn. They have a dataset with 10,000 records and 50 features. What is the primary consideration for model accuracy?

A.The dataset size is too small for reliable predictions.

B.All features must be numerical and normalized.

C.The dataset must be balanced between churned and non-churned customers.

D.The model needs to be retrained daily.

AnswerC

Balancing prevents bias towards majority class.

Why this answer

Einstein Prediction Builder uses automated machine learning (AutoML) to train models, and class imbalance is a critical factor that directly impacts model accuracy. If the dataset is highly skewed (e.g., 95% non-churned, 5% churned), the model may achieve high accuracy by simply predicting the majority class, but it will fail to identify actual churners. Einstein Prediction Builder includes built-in handling for imbalanced data, but the user must ensure the dataset is reasonably balanced or use techniques like oversampling to improve predictive performance.

Exam trap

Salesforce often tests the misconception that dataset size is the primary driver of accuracy, but the trap here is that class balance is more critical than raw record count for classification models in Einstein Prediction Builder.

How to eliminate wrong answers

Option A is wrong because 10,000 records is generally sufficient for a binary classification task like churn prediction, especially with 50 features; Einstein Prediction Builder can work with datasets as small as a few thousand records. Option B is wrong because Einstein Prediction Builder automatically handles feature encoding and normalization; it accepts categorical, numerical, and text features without requiring manual preprocessing. Option D is wrong because retraining frequency depends on business needs and data drift, not on model accuracy; Einstein Prediction Builder supports scheduled retraining but does not require daily retraining as a primary consideration for accuracy.

Practice this question →

44

MCQmedium

A customer service team wants to automatically route incoming emails to the appropriate department based on content. Which NLP capability is essential for this task?

A.Sentiment analysis

B.Text classification

C.Entity extraction

D.Intent detection

AnswerB

Text classification assigns a category to the email, enabling routing.

Why this answer

Text classification categorizes text into predefined classes, enabling automated routing.

Practice this question →

45

MCQhard

A data scientist notices that an Einstein Discovery model predicts a low probability of conversion for all leads in a new campaign, even though the campaign targets high-value accounts. Which initial diagnostic step should be taken?

A.Retrain the model with the new campaign data included

B.Check the Einstein model recipe for incorrect filters

C.Compare the feature distributions of the training and campaign data

D.Increase the prediction confidence threshold

AnswerC

Distribution mismatch often explains low predictions; if features differ, the model may not apply.

Why this answer

The most likely cause of a model predicting low conversion for all leads in a new campaign is a shift in feature distributions between the training data and the campaign data (covariate shift). Checking these distributions is the standard initial diagnostic step to identify if the model is encountering data it was not trained on, which would invalidate its predictions. This aligns with best practices for model monitoring and data validation in Einstein Discovery.

Exam trap

Salesforce often tests the misconception that retraining is the immediate fix for poor model performance, but the trap here is that candidates overlook the fundamental diagnostic step of checking for data drift before taking any corrective action.

How to eliminate wrong answers

Option A is wrong because retraining the model with new campaign data is a premature action; the root cause (data drift) must be diagnosed first, and retraining without investigation could mask the issue or introduce bias. Option B is wrong because checking the Einstein model recipe for incorrect filters addresses configuration errors, but the scenario describes a systematic prediction pattern across all leads, which is more indicative of data distribution shift than a filter misconfiguration. Option D is wrong because increasing the prediction confidence threshold does not fix the underlying cause of low probabilities; it only changes the cutoff for classification, leaving the flawed predictions unchanged.

Practice this question →

46

MCQhard

A data scientist is evaluating a custom Einstein model for a lead scoring use case. The model's precision is 0.9, recall is 0.5. What is the most important improvement priority?

A.Increase recall to reduce false negatives

B.Increase precision to reduce false positives

C.Optimize for an F1 score of 0.7

D.Improve overall accuracy above 80%

AnswerA

Recall is low (0.5), meaning half of actual leads are missed. This should be improved.

Why this answer

With a precision of 0.9 and recall of 0.5, the model is highly selective but misses many actual leads (high false negatives). In lead scoring, false negatives mean lost sales opportunities, which is typically more costly than false positives. Therefore, increasing recall to capture more true positives is the most important improvement priority.

Exam trap

Salesforce often tests the trade-off between precision and recall in imbalanced classification scenarios, where candidates mistakenly focus on improving precision or accuracy without recognizing that low recall (high false negatives) is the critical business problem in lead scoring.

How to eliminate wrong answers

Option B is wrong because increasing precision would further reduce false positives, but the model already has high precision (0.9); the bigger issue is the low recall (0.5) causing many missed leads. Option C is wrong because optimizing for an F1 score of 0.7 is a metric goal, not a direct improvement priority; the F1 score is a harmonic mean of precision and recall, and simply targeting a number does not address the underlying imbalance. Option D is wrong because overall accuracy can be misleading in imbalanced datasets; a model could achieve high accuracy by always predicting the majority class, but that would not improve lead capture for the minority class (actual leads).

Practice this question →

47

MCQmedium

A sales operations team wants to automatically categorize incoming support cases into predefined categories (e.g., Billing, Technical, General). The team has thousands of historical cases with correct category labels. Which AI approach should they use?

A.Topic modeling (unsupervised)

B.Text classification

C.Sentiment analysis

D.Named entity recognition

AnswerB

Correct: text classification maps text to predefined categories using labeled training data.

Why this answer

Text classification is a supervised NLP task that assigns predefined labels to text. Historical labeled cases provide the training data for a classifier.

Practice this question →

48

MCQmedium

A company wants to generate personalized marketing email content for each customer, including product recommendations and tailored copy. Which AI approach is BEST?

A.Use an unsupervised learning model to cluster customers and send generic emails per cluster

B.Use predictive AI to score customers and generative AI to create email copy based on the scores

C.Use only predictive AI to send the same email to all customers with top products

D.Use only generative AI to create emails without any customer data

AnswerB

Predictive AI identifies what to recommend, and generative AI writes the personalized copy.

Why this answer

Generative AI can create original content, and predictive AI can recommend products. The combination of both can generate personalized copy with dynamic product suggestions. Using only predictive AI yields recommendations but not content; generative AI alone can create text but lacks recommendation logic.

Practice this question →

49

MCQmedium

A data scientist is evaluating the performance of an Einstein Discovery model. They observe that the model has high accuracy but low precision for a specific prediction class. What does this indicate?

A.The model is overfitted to the training data.

B.The model correctly predicts most instances but has many false positives for that class.

C.The model correctly predicts most instances but has many false negatives for that class.

D.The model rarely predicts that class, leading to high accuracy.

AnswerB

Low precision indicates many false positives.

Why this answer

High accuracy with low precision for a specific class indicates that while the model correctly classifies the majority of instances overall, it produces a high number of false positives for that class. Precision measures the proportion of positive identifications that were actually correct, so low precision means many of the predicted positive cases are false alarms. In Einstein Discovery, this trade-off is critical when optimizing for business outcomes where false positives are costly.

Exam trap

Salesforce often tests the distinction between precision and recall, trapping candidates who confuse false positives (low precision) with false negatives (low recall) when interpreting accuracy metrics.

How to eliminate wrong answers

Option A is wrong because overfitting typically leads to high accuracy on training data but poor generalization to unseen data, not specifically to low precision for a single class. Option C is wrong because many false negatives would reduce recall, not precision; the scenario describes low precision, which is about false positives. Option D is wrong because if the model rarely predicts that class, precision could be high (few predictions, mostly correct) and accuracy could be high due to class imbalance, but low precision indicates frequent incorrect predictions for that class.

Practice this question →

50

MCQmedium

A retailer wants to recommend products to customers based on their purchase history and browsing behavior. Which AI approach is most suitable?

A.Supervised learning to predict purchase probability for each product

B.Reinforcement learning with real-time rewards

C.Natural language generation to create product descriptions

D.Computer vision to analyze product images

AnswerA

Supervised learning can be trained on historical purchases to predict which products a customer is likely to buy.

Why this answer

Product recommendations are typically handled by supervised or unsupervised learning, but the most common approach in CRM is collaborative filtering or similar supervised/unsupervised models. However, given options, supervised learning on historical interactions is typical.

Practice this question →

51

Multi-Selecthard

A company wants to use AI to automatically extract key information (e.g., invoice number, date, total amount) from scanned invoices. Which THREE technologies should be combined?

Select 3 answers

A.Generative AI for text generation

B.Optical Character Recognition (OCR)

C.Sentiment analysis

D.Natural Language Processing (NLP) – entity extraction

E.Computer vision

AnswersB, D, E

Converts image text into machine-readable text.

Why this answer

Computer vision reads the document, OCR converts image to text, and NLP extracts entities. Generative AI is not needed.

Practice this question →

52

MCQeasy

What is the term for when an AI model produces confident but incorrect information, often in generative AI?

A.Bias

B.Overfitting

C.Hallucination

D.Underfitting

AnswerC

Hallucination describes the generation of false information presented as fact.

Why this answer

Hallucination refers to when a generative AI model fabricates information that seems plausible but is false.

Practice this question →

53

MCQmedium

A data scientist notices that a churn prediction model has high variance: small changes in training data cause large changes in predictions. Which technique is BEST to address this?

A.Apply regularization techniques

B.Reduce the amount of training data

C.Add more features to the model

D.Increase the number of training epochs

AnswerA

Regularization penalizes large coefficients, reducing model complexity and variance.

Why this answer

High variance indicates overfitting. Regularization (e.g., L1/L2) reduces model complexity and variance. Adding more features would increase variance, increasing training data can help but is not always feasible, and reducing training data would worsen the problem.

Practice this question →

54

Multi-Selecthard

A company is deploying an AI chatbot for customer service. Which THREE ethical considerations should be addressed? (Select 3)

Select 3 answers

A.Protecting customer data privacy during model training and inference

B.Using the most complex deep learning architecture available

C.Mitigating bias in training data that could lead to unfair responses

D.Ensuring the chatbot explains when a customer is speaking to AI

E.Maximizing the number of training epochs

AnswersA, C, D

Data privacy is a core ethical and legal requirement.

Why this answer

Transparency, bias mitigation, and data privacy are key ethical areas for AI systems.

Practice this question →

55

MCQmedium

A company uses an AI model to predict customer churn. The model's predictions are used to automatically assign discounts to high-risk customers. A customer complains about receiving a discount offer they did not request. Which ethical concern is most relevant?

A.Bias in training data

B.Model overfitting

C.Data quality

D.Lack of transparency of AI decisions to customers

AnswerD

Correct: customers should be informed that the discount was AI-generated and why.

Why this answer

Transparency about AI decisions is key. Customers should be informed when an AI system is making decisions that affect them.

Practice this question →

56

MCQmedium

A company uses an AI model to classify customer support cases into categories (billing, technical, general). The model performs well on training data but poorly on new cases. Which issue is MOST likely occurring?

A.Underfitting

B.Overfitting

C.Data leakage

D.Bias in training data

AnswerB

Overfitting leads to high accuracy on training data but low accuracy on new data because the model is too complex.

Why this answer

Overfitting means the model memorized the training data and fails to generalize to new, unseen data. Underfitting would show poor performance on both training and test data. Data leakage occurs when future information leaks into training.

Bias in data is a different issue.

Practice this question →

57

MCQeasy

Which type of AI is designed to perform only a specific task, such as playing chess or recommending products?

A.Narrow AI

B.Super AI

C.General AI

D.Strong AI

AnswerA

Correct: narrow AI is task-specific, like chess or product recommendations.

Why this answer

Narrow AI (weak AI) is specialized for one task and cannot generalize beyond its training.

Practice this question →

58

MCQeasy

A sales operations manager wants to predict which leads are most likely to convert to deals. The CRM has historical data on thousands of leads with outcomes (converted or not). Which type of machine learning should they use?

A.Unsupervised learning

B.Supervised learning

C.Deep learning

D.Reinforcement learning

AnswerB

Supervised learning trains on labeled historical data (features + outcome) to predict future outcomes like lead conversion.

Why this answer

Supervised learning uses labeled data (past leads with known outcomes) to predict future lead conversion. Unsupervised learning finds patterns without labels, reinforcement learning learns from rewards, and deep learning is a subset of supervised/unsupervised but not the most specific answer here.

Practice this question →

59

MCQeasy

A user asks Einstein GPT to generate a product description. The AI returns a response with a confidence score of 0.65. What does this score indicate?

A.There is a 65% probability that the response exactly matches the training data

B.The model is 65% confident that the response is accurate

C.The response is 65% shorter than the optimal length

D.The model has a 65% likelihood of generating the same response again

AnswerB

Confidence scores indicate the model's assessment of how likely the generated answer is correct.

Why this answer

The confidence score in AI models like Einstein GPT quantifies the model's internal certainty that its generated output is factually correct or contextually appropriate. A score of 0.65 means the model estimates a 65% probability that the response is accurate based on its training and inference algorithms, not that it matches training data or has a fixed length.

Exam trap

Salesforce often tests the misconception that a confidence score indicates a direct probability of correctness or a measure of output quality, when in reality it is a model's self-assessed certainty that can be misleading and is not a guarantee of factual accuracy.

How to eliminate wrong answers

Option A is wrong because the confidence score does not measure how closely the response matches training data; it reflects the model's probabilistic assessment of correctness, not a similarity metric. Option C is wrong because confidence scores are unrelated to output length; they are a probability value between 0 and 1, not a measure of brevity. Option D is wrong because the score does not indicate the likelihood of generating the same response again; that depends on the model's stochastic sampling parameters (e.g., temperature), not the confidence score.

Practice this question →

60

Multi-Selecthard

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

Select 3 answers

A.Brightness of the user interface

B.Model temperature setting

C.Clarity and specificity of the prompt

D.Quality of the grounding data provided

E.Length of the response generated

AnswersB, C, D

Higher temperature increases creativity but may reduce factual accuracy.

Why this answer

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.

Exam trap

Salesforce often tests the misconception that output length or interface settings affect model accuracy, when in fact only prompt clarity, grounding data quality, and model hyperparameters like temperature are the true determinants.

Practice this question →

61

Multi-Selectmedium

Which TWO actions are best practices when implementing Einstein Prediction Service?

Select 2 answers

A.Ignore correlated features to simplify the model.

B.Clean the data to handle missing values and outliers.

C.Select relevant features that are likely to influence the prediction.

D.Include all available fields in the dataset for maximum information.

E.Use the default field mapping without review.

AnswersB, C

Data cleaning improves model accuracy.

Why this answer

Data quality directly impacts the accuracy and reliability of Einstein Prediction Service models. Cleaning data to handle missing values and outliers ensures that the training data is representative and reduces the risk of skewed predictions, which is a fundamental prerequisite for any machine learning model within Salesforce's AI framework.

Exam trap

Salesforce often tests the misconception that more data always leads to better predictions, but in practice, irrelevant or noisy features degrade model performance, making feature selection and data cleaning critical.

Practice this question →

62

Multi-Selectmedium

A company is deploying Einstein Prediction Builder to predict equipment failure. Which three considerations are essential for building an accurate prediction model? (Choose 3)

Select 3 answers

A.Missing values in features should be handled appropriately.

B.The dataset should span a sufficient time period to capture patterns.

C.All input features must be numerical.

D.The model should be retrained only once after initial deployment.

E.The prediction horizon must be clearly defined.

AnswersA, B, E

Missing data can bias the model.

Why this answer

Missing values in features can introduce bias or cause errors in the predictive model. Einstein Prediction Builder automatically handles missing data through imputation, but understanding how missing values are treated is essential for model accuracy, as inappropriate handling can distort relationships between features and the target outcome.

Exam trap

Salesforce often tests the misconception that all input features must be numerical for machine learning models, but Einstein Prediction Builder natively supports non-numerical data types through automated preprocessing.

Practice this question →

63

MCQhard

A generative AI chatbot sometimes produces factually incorrect responses about a company's products. What is this phenomenon called?

A.Data leakage

B.Overfitting

C.Hallucination

D.Bias

AnswerC

Hallucination describes when an AI model produces confident but false information.

Why this answer

Hallucination is when an AI generates incorrect or nonsensical information.

Practice this question →

64

Multi-Selecteasy

A sales team is implementing Einstein Lead Scoring. Which two actions should they take to ensure the model is effective? (Choose 2)

Select 2 answers

A.Include only demographic data for scoring.

B.Set a fixed score threshold for all users.

C.Disable the model for low-volume leads.

D.Ensure the training data includes both converted and unconverted leads.

E.Regularly review and provide feedback on lead conversions.

AnswersD, E

Balanced data is necessary.

Why this answer

Einstein Lead Scoring requires training data that includes both converted and unconverted leads to build a predictive model that can distinguish between leads likely to convert and those that are not. Without unconverted leads, the model cannot learn the negative patterns, leading to biased predictions and poor accuracy.

Exam trap

Salesforce often tests the misconception that Einstein Lead Scoring can work effectively with only positive examples (converted leads), but the model fundamentally requires both positive and negative examples to learn the difference between high- and low-quality leads.

Practice this question →

65

Multi-Selectmedium

Which THREE factors are most important for ensuring the accuracy of an AI model in a CRM context? (Select 3)

Select 3 answers

A.Relevant features selected for the prediction task

B.Large number of model parameters

C.Representative training data

D.Higher inference speed

E.Clean and complete data

AnswersA, C, E

Feature selection directly affects model performance.

Why this answer

Data quality, representative training data, and relevant features directly impact model accuracy.

Practice this question →

66

MCQhard

Refer to the exhibit. An admin configures Einstein Next Best Action with the above JSON. The expected behavior is to recommend the top 5 actions for open leads with a score of at least 70. However, only 2 recommendations appear for some leads. Which is the most likely cause?

A.The filter on Lead Status is incorrectly excluding actions.

B.The scoreThreshold of 70 excludes many actions, so fewer than 5 meet the criteria.

C.The recommendation strategy is misconfigured for this object.

D.The maxRecommendations is set to 2 instead of 5.

AnswerB

Score threshold filters out low-scoring actions.

Why this answer

The scoreThreshold of 70 filters out any actions with a score below 70. If fewer than 5 actions meet this threshold, the system returns only those that qualify, resulting in fewer than 5 recommendations. The maxRecommendations setting defines the upper limit, but the actual number returned is constrained by the scoreThreshold.

Exam trap

Salesforce often tests the interaction between scoreThreshold and maxRecommendations, where candidates mistakenly assume maxRecommendations is the sole determinant of the number of recommendations, overlooking that scoreThreshold can reduce the count below that limit.

How to eliminate wrong answers

Option A is wrong because the filter on Lead Status is not mentioned in the JSON exhibit; the issue is with the score threshold, not a status filter. Option C is wrong because the recommendation strategy is correctly configured for the object; the JSON shows valid strategy settings, and the problem is purely about scoring. Option D is wrong because maxRecommendations is set to 5 in the JSON (as stated in the expected behavior), not 2; the reduced count is due to the scoreThreshold, not a misconfiguration of maxRecommendations.

Practice this question →

67

MCQmedium

A customer service chatbot misinterprets user requests and often provides irrelevant answers. The development team wants to improve the chatbot's understanding of user intent. Which NLP component should they focus on?

A.Text generation

B.Sentiment analysis

C.Named entity recognition

D.Intent detection

AnswerD

Correct: intent detection determines the user's goal, enabling the correct response.

Why this answer

Intent detection is the NLP task that identifies what the user wants to do. Improving it directly addresses misinterpretation of requests.

Practice this question →

68

MCQhard

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?

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

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

C.Filter training data to exclude tickets from the last 3 months and impute missing 'Priority__c' values with the most common priority.

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

AnswerC

Removing anomalous period and fixing data quality improves model relevance.

Why this answer

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.

Exam trap

Salesforce often tests the misconception that simply removing a problematic field or adding more data will fix model performance, when the real issue is data drift and missing value handling that require both temporal filtering and imputation.

How to eliminate wrong answers

Option A is wrong because rolling back to a 6-month-old model ignores the fact that the recent 3-month spike was due to a temporary bug that has been fixed; the older model may not generalize to current data and could still have poor precision. Option B is wrong because simply excluding 'Priority__c' without addressing missing values or data drift may harm recall, as the field could still be predictive when populated; the core issue is the biased training period and null handling, not the field itself. Option D is wrong because removing 'Is_Critical__c' alone does not fix the data drift from the last 3 months, and increasing training data to 50,000 records without cleaning or rebalancing could amplify the bias from the bug period, potentially worsening precision.

Practice this question →

69

MCQmedium

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?

A.Object detection

B.Image classification

C.Product recognition

D.Text extraction (OCR)

AnswerC

Correct. Product recognition identifies specific products from images.

Why this answer

Product recognition is a specific computer vision task that identifies objects (products) in images.

Practice this question →

70

Multi-Selecthard

A data analyst is reviewing an Einstein Discovery story and notices that one input feature has a very high influence on the predicted outcome. Which two conclusions are justified based on this observation? (Choose 2)

Select 2 answers

A.The feature has a causal relationship with the outcome.

B.Removing the feature will significantly improve model performance.

C.The model is likely overfitted to that feature.

D.The feature could be a surrogate for other correlated features.

E.The feature is the most important predictor of the outcome.

AnswersD, E

Could represent collinear features.

Why this answer

A feature with high influence in an Einstein Discovery story may be acting as a proxy for other correlated features, meaning its apparent importance could be due to shared variance with other predictors. This is a known phenomenon in machine learning where collinearity can inflate a feature's influence score without it being uniquely causal.

Exam trap

Salesforce often tests the distinction between correlation and causation, and the trap here is assuming that high feature influence implies a direct causal link or that removing the feature will always improve the model.

Practice this question →

71

Multi-Selecteasy

Which TWO data types can be used as input for Einstein Vision?

Select 2 answers

A.Video files (MP4).

B.URLs pointing to images.

C.Image files (JPEG, PNG).

D.Text documents (PDF).

E.Audio files (WAV).

AnswersB, C

URLs are accepted as references to images.

Why this answer

Einstein Vision is designed to analyze and classify visual content, specifically images. It accepts image files (JPEG, PNG) and URLs pointing to images as input, allowing the model to process the visual data for object detection, classification, or other vision tasks. Video, text, and audio files are not supported because the service is not built for temporal or non-visual data.

Exam trap

Salesforce often tests the misconception that Einstein Vision can handle video or multimedia files because it is an 'AI' service, but the exam specifically limits input to static images and image URLs.

Practice this question →

72

MCQmedium

A company wants to build a customer service chatbot that answers questions about their internal policy documents. The documents are updated monthly, and the team cannot afford to retrain a model each time. Which approach is MOST appropriate?

A.Use a larger foundation model with a longer context window and paste all documents into each prompt

B.Train a custom model from scratch on the policy documents each month

C.Fine-tune a base LLM on the policy documents monthly

D.Use Retrieval-Augmented Generation (RAG) with the policy documents indexed in a vector store

AnswerD

RAG retrieves relevant document chunks at query time, ensuring the chatbot always answers from the latest uploaded documents without any model retraining.

Why this answer

RAG (Retrieval-Augmented Generation) allows the LLM to retrieve relevant document sections at inference time, so knowledge stays current without retraining. The other options either require expensive retraining for each update or lack document grounding.

Practice this question →

73

MCQeasy

What is the primary purpose of Einstein Studio in the Salesforce AI ecosystem?

A.Create and manage prompt templates

B.Unify customer data from multiple sources

C.Train, test, and deploy custom AI models using AutoML

D.Deploy pre-built Einstein GPT models

AnswerC

Einstein Studio provides AutoML capabilities for building custom models without deep data science expertise.

Why this answer

Einstein Studio is specifically designed to allow users to train, test, and deploy custom AI models using AutoML, without requiring deep data science expertise. It provides a no-code interface for building models on Salesforce Data Cloud data, enabling predictive scoring and recommendations directly within the Salesforce ecosystem.

Exam trap

The trap here is confusing Einstein Studio's custom AutoML model building with Einstein GPT's pre-built generative AI capabilities, leading candidates to select Option D or A when the question specifically asks about the primary purpose of Einstein Studio.

How to eliminate wrong answers

Option A is wrong because creating and managing prompt templates is the primary function of Einstein GPT Trust Layer and Prompt Builder, not Einstein Studio. Option B is wrong because unifying customer data from multiple sources is the role of Data Cloud (formerly Customer Data Platform), not Einstein Studio. Option D is wrong because deploying pre-built Einstein GPT models is handled by Einstein GPT and its pre-built actions, whereas Einstein Studio focuses on custom model creation using AutoML.

Practice this question →

74

MCQhard

A company uses an NLP model to detect intent in customer messages. The model works well for English but fails for Spanish messages. What is the most likely cause?

A.The model is overfitted to English

B.The inference threshold is too high

C.The model uses the wrong algorithm

D.The training data is predominantly English

AnswerD

Models perform poorly on languages unseen in training data.

Why this answer

If the training data lacks Spanish examples, the model cannot learn Spanish patterns.

Practice this question →

75

MCQeasy

A marketing manager uses Einstein recommendations on their website, but customers are receiving suggestions for products they already purchased. What is the most likely cause?

A.The product catalog is not updated with purchase history.

B.The model is using real-time browsing data that includes past purchases.

C.The recommendation model is not filtering out previously purchased items.

D.The recommendations are based on collaborative filtering without personalization.

AnswerC

Einstein recommendations can exclude purchased items.

Why this answer

C is correct because the most likely cause is that the recommendation model is not configured to exclude previously purchased items. Einstein Recommendations uses customer purchase history to personalize suggestions, but if the model's filtering logic does not explicitly remove items the customer has already bought, those items will continue to appear in the recommendations. This is a common oversight in model configuration rather than a data sync or algorithm type issue.

Exam trap

Salesforce often tests the distinction between data source issues (e.g., catalog not updated) versus model configuration issues (e.g., missing filters), and the trap here is assuming the problem is a data sync failure when it is actually a missing business rule in the recommendation logic.

How to eliminate wrong answers

Option A is wrong because the product catalog typically contains product metadata (e.g., name, price, category), not individual customer purchase history; purchase history is stored separately in order or transaction objects. Option B is wrong because real-time browsing data includes current session behavior, not past purchases; past purchases are historical data, not real-time. Option D is wrong because collaborative filtering inherently personalizes recommendations based on user-item interactions; the issue is not the algorithm type but the lack of a filter to exclude purchased items.

Practice this question →

Page 1 of 3 · 163 questions totalNext →

Ready to test yourself?

Try a timed practice session using only AI Fundamentals questions.

Start 20-question session