Courseiva

CCNA Data for AI Questions

38 of 113 questions · Page 2/2 · Data for AI · Answers revealed

Practice these questionsDomain overviewAll questions
76
MCQeasy

A company wants to use Einstein Article Recommendations to surface relevant knowledge articles to its support agents. What two data components are required to set up this feature?

A.Email-to-case logs and Knowledge Article feedback
B.Knowledge Article View event data and Case records
C.Knowledge Article categories and Case priority
D.Community user activity and Knowledge Article ratings
AnswerB

Article views show which articles were read; Cases provide context for recommendations.

Why this answer

Einstein Article Recommendations uses historical Knowledge Article View event data to understand which articles agents have found useful in the past, and Case records to provide context about the current issue. By analyzing patterns between case attributes and article views, the AI can predict and surface the most relevant articles for a given case. Without both data components, the recommendation engine cannot learn the association between case details and article usefulness.

Exam trap

Salesforce often tests the distinction between optional enhancement data (like ratings or categories) and the mandatory data sources (view events and case records) required to train the recommendation model.

How to eliminate wrong answers

Option A is wrong because Email-to-case logs are used for email-to-case routing and parsing, not for training article recommendations, and Knowledge Article feedback is a secondary signal, not a required data component. Option C is wrong because Knowledge Article categories and Case priority are metadata fields that can influence recommendations but are not the two required data components; the feature specifically needs view event data and case records. Option D is wrong because Community user activity is unrelated to agent-facing article recommendations, and Knowledge Article ratings are optional feedback, not a core requirement.

Practice this question →
77
MCQeasy

A company wants to use Einstein Reply Recommendations in Service Cloud. What data is required to train the model?

A.Knowledge articles only.
B.Email templates.
C.Case comments and chat transcripts.
D.Historical email replies and customer satisfaction ratings.
AnswerC

These capture actual agent-customer interactions used for training.

Why this answer

Einstein Reply Recommendations in Service Cloud uses historical service interactions to suggest relevant replies to agents. The model is trained on case comments and chat transcripts because these contain the natural language patterns and resolutions that agents actually use in real-time service conversations, enabling the AI to learn effective response strategies.

Exam trap

Salesforce often tests the misconception that Einstein Reply Recommendations uses Knowledge articles or email templates, when in fact it relies on unstructured service conversation data like case comments and chat transcripts to learn agent-specific reply patterns.

How to eliminate wrong answers

Option A is wrong because Knowledge articles are structured content used for knowledge-based answers, not the conversational reply patterns needed for Einstein Reply Recommendations. Option B is wrong because email templates are pre-written, static responses that lack the dynamic, contextual language variations found in actual service interactions. Option D is wrong because historical email replies and customer satisfaction ratings are not the primary training data; Einstein Reply Recommendations specifically requires case comments and chat transcripts to learn from direct agent-customer exchanges.

Practice this question →
78
MCQmedium

A telecom company uses Einstein Discovery to predict customer churn. The training dataset contains 100,000 records, but only 5% represent churned customers. The model achieves 95% accuracy on a holdout test set, but the recall for churn is only 20%. The business wants to proactively retain at-risk customers, so they need to identify as many churners as possible. What action should the data scientist take to improve churn recall?

A.Increase the regularization parameter to prevent overfitting.
B.Collect more data, especially of churned customers.
C.Oversample the minority class using SMOTE to create synthetic churn examples.
D.Undersample the majority class to match the minority class size.
AnswerC

SMOTE generates synthetic instances of the minority class, balancing the dataset and improving recall without losing information.

Why this answer

Class imbalance causes the model to favor the majority class. Oversampling the minority class (e.g., using SMOTE) balances the dataset, helping the model learn churn patterns better and improve recall.

Practice this question →
79
MCQhard

You are a Salesforce AI Specialist at a mid-sized manufacturing company. The company uses Einstein Lead Scoring to prioritize leads. The model was trained on historical lead data and has been in production for three months. Recently, the sales team reports that high-scoring leads are not converting as expected. You investigate and find that the model's data source includes leads from the past 18 months. However, six months ago, the company changed its lead qualification process: they started requiring a demo before scoring leads as 'qualified.' As a result, the definition of a converted lead changed. What is the best course of action to improve model performance?

A.Manually adjust the model's prediction threshold to account for the new process
B.Retrain the model using only leads from the last six months after the process change
C.Remove the 'Demo Scheduled' field from the model to avoid bias
D.Add more historical leads from before the process change to increase data volume
AnswerB

This ensures the model learns from data that reflects the current conversion criteria.

Why this answer

The change in lead qualification process six months ago introduced a data distribution shift (concept drift), making older leads no longer representative of the current conversion behavior. Retraining the model on only the last six months of data aligns the training set with the new definition of a 'converted lead,' allowing Einstein Lead Scoring to learn the updated patterns and improve prediction accuracy.

Exam trap

The trap here is that candidates may think adjusting the threshold (Option A) is sufficient, but they fail to recognize that a change in the definition of the target variable requires retraining on a representative dataset, not just tuning a post-processing parameter.

How to eliminate wrong answers

Option A is wrong because manually adjusting the prediction threshold does not address the underlying change in the definition of a converted lead; it only shifts the cutoff for scoring, which cannot compensate for a fundamentally different target variable. Option C is wrong because removing the 'Demo Scheduled' field does not solve the problem—the issue is the change in the conversion definition, not bias from that field; in fact, the field may now be more predictive under the new process. Option D is wrong because adding more historical leads from before the process change would exacerbate the data mismatch, as those leads follow the old qualification rules and would dilute the model's ability to learn the current conversion patterns.

Practice this question →
80
Multi-Selecthard

Which THREE are key dimensions of data quality that directly impact AI model performance?

Select 3 answers
A.Consistency.
B.Timeliness.
C.Accuracy.
D.Data volume.
E.Completeness.
AnswersA, C, E

Inconsistent data (e.g., different formats) confuses models and degrades accuracy.

Why this answer

Consistency is a key dimension of data quality because AI models rely on stable patterns in the data. If the same entity is represented differently across records (e.g., 'NY' vs 'New York'), the model may learn incorrect correlations, leading to degraded prediction accuracy and unreliable outputs.

Exam trap

Salesforce often tests the distinction between data quality dimensions and data quantity metrics, so candidates mistakenly select 'data volume' thinking more data always improves AI performance, when in fact the exam focuses on accuracy, completeness, and consistency as the three critical quality dimensions.

Practice this question →
81
MCQhard

A data architect is designing a data model for Einstein Discovery. The data includes categorical variables with high cardinality (e.g., postal codes). What is the best practice to handle such features?

A.Encode them as one-hot vectors.
B.Exclude them from the model.
C.Use the raw values without transformation.
D.Group them into higher-level categories (e.g., region).
AnswerD

Reduces cardinality while preserving signal.

Why this answer

Grouping high-cardinality categories into broader categories reduces overfitting and improves model stability.

Practice this question →
82
Multi-Selecthard

Which TWO considerations are critical when planning data labeling for a computer vision project in a regulated industry?

Select 2 answers
A.Data storage location for label files
B.Mitigating labeler bias to ensure fairness
C.Compliance with data privacy regulations (e.g., GDPR)
D.Labeling timeline and budget constraints
E.Choosing between bounding boxes and segmentation masks
AnswersB, C

Bias can affect model fairness and regulatory requirements.

Why this answer

Labeler bias can introduce systematic errors into the training data, leading to models that perform unfairly or inaccurately across different demographic groups. In regulated industries, such bias can violate anti-discrimination laws and regulatory standards, making its mitigation a critical planning consideration.

Exam trap

Salesforce often tests the distinction between operational details (like storage location or annotation type) and critical regulatory or ethical considerations, leading candidates to choose technically valid but non-critical options like A or E.

Practice this question →
83
MCQmedium

A data engineer needs to create a feature that represents the average purchase amount per customer over the last 30 days. The transactional data is timestamped. Which feature engineering technique is most appropriate?

A.Sum of all purchase amounts per customer
B.Rolling average of purchase amounts over a 30-day window
C.Count of purchases per customer
D.Minimum purchase amount per customer
AnswerB

Rolling average matches the requirement.

Why this answer

A rolling average over a 30-day window directly computes the average purchase amount per customer for only the most recent 30 days of transactions, which matches the requirement of a time-sensitive feature. This technique uses a sliding window function (e.g., AVG() with a ROWS or RANGE frame in SQL, or rolling().mean() in pandas) that respects the timestamp order, ensuring only relevant data contributes to the feature.

Exam trap

Salesforce often tests the distinction between simple aggregation (like sum or count) and time-windowed aggregation, trapping candidates who overlook the 'over the last 30 days' temporal constraint and choose a static aggregate instead.

How to eliminate wrong answers

Option A is wrong because summing all purchase amounts per customer ignores the 30-day time constraint and would include historical data outside the window, producing a feature that does not reflect recent behavior. Option C is wrong because counting purchases per customer measures frequency, not the average amount spent, and also lacks the time window restriction. Option D is wrong because the minimum purchase amount per customer is a different aggregate (minimum) that does not capture the central tendency of spending and similarly ignores the 30-day window.

Practice this question →
84
MCQeasy

A company is preparing data for Einstein Article Recommendation. Which data source is most appropriate for training the model?

A.Historical article view and click data.
B.Org metadata.
C.System debug logs.
D.User profile data only.
AnswerA

This captures user preferences directly.

Why this answer

Einstein Article Recommendation uses supervised machine learning to predict which articles users are likely to find relevant. The model must be trained on historical user engagement signals—specifically article view and click data—to learn patterns of relevance. Without this behavioral data, the model cannot establish a correlation between user actions and article content.

Exam trap

Salesforce often tests the misconception that static data like user profiles or org metadata can substitute for behavioral training data, but the model fundamentally requires historical interaction signals to learn relevance.

How to eliminate wrong answers

Option B is wrong because org metadata (e.g., company name, industry) provides only static contextual information and lacks the user-article interaction signals required for training a recommendation model. Option C is wrong because system debug logs contain low-level technical events (e.g., errors, stack traces) that are irrelevant to user content preferences and would introduce noise rather than meaningful training features. Option D is wrong because user profile data alone (e.g., role, department) does not capture which articles users actually viewed or clicked, so the model cannot learn relevance from user behavior.

Practice this question →
85
MCQeasy

A company plans to use Einstein Discovery to analyze sales data. Which data preparation step is essential for time-series forecasting?

A.Remove all outliers in sales amounts
B.Ensure date fields are properly formatted and contain sufficient historical range
C.Remove duplicate records
D.Scale all numeric fields to a 0-1 range
AnswerB

Einstein Discovery relies on date fields for trend detection.

Why this answer

For time-series forecasting in Einstein Discovery, the date field must be properly formatted (e.g., as a date or datetime data type) and contain a sufficient historical range to identify patterns like seasonality and trends. Without adequate historical data, the model cannot learn temporal dependencies, making this step essential.

Exam trap

Salesforce often tests the misconception that data normalization (scaling) is always required for AI models, but for tree-based algorithms like those in Einstein Discovery, scaling is irrelevant, and the trap is that candidates pick Option D thinking it is a universal preprocessing step.

How to eliminate wrong answers

Option A is wrong because removing all outliers in sales amounts can discard legitimate extreme values that represent real-world events (e.g., holiday spikes, promotions), which are critical for accurate time-series forecasting; Einstein Discovery handles outliers through model tuning rather than blanket removal. Option C is wrong because while removing duplicate records is a general data cleaning best practice, it is not specifically essential for time-series forecasting; duplicates in date-indexed data are typically handled by aggregation or deduplication, but this step is not a prerequisite for the forecasting algorithm. Option D is wrong because scaling numeric fields to a 0-1 range is unnecessary for time-series forecasting in Einstein Discovery, as tree-based models (like Gradient Boosted Trees) used internally are invariant to monotonic transformations and do not require normalization.

Practice this question →
86
MCQmedium

During the data preparation phase for an AI model, a data engineer discovers that the 'AnnualRevenue' field contains some negative values. What is the best course of action?

A.Delete all records with negative revenue
B.Replace negative values with the mean of positive values
C.Keep negative values as they might represent returns or refunds
D.Investigate the data source to correct the negative values
AnswerD

Correcting at source ensures data integrity.

Why this answer

Negative revenue values typically indicate data entry errors, system bugs, or incorrect data transformations. The best practice in data preparation is to investigate the source system to understand why negative values were generated and correct them at the origin, ensuring data integrity before any imputation or deletion. Simply deleting or imputing without root-cause analysis can introduce bias or mask underlying data quality issues.

Exam trap

Salesforce often tests the misconception that imputation (e.g., mean replacement) is a safe default for handling invalid data, when in fact the correct first step is always to trace and fix the root cause at the data source.

How to eliminate wrong answers

Option A is wrong because deleting records with negative revenue can introduce selection bias and reduce the dataset size, potentially discarding valid data if negative values represent legitimate business events like refunds. Option B is wrong because replacing negative values with the mean of positive values artificially inflates the central tendency and distorts the distribution, which can degrade model performance, especially for regression tasks. Option C is wrong because keeping negative values as-is without investigation assumes they are valid, but in most financial datasets, revenue is non-negative by definition, and unverified negative values will mislead the model during training.

Practice this question →
87
MCQeasy

An admin is setting up Einstein Article Recommendations. Which type of data is essential for the model to learn which articles are relevant?

A.Article publication dates
B.Article view events from users
C.User job titles
D.Article author names
AnswerB

View events are the primary input for collaborative filtering.

Why this answer

Einstein Article Recommendations uses a collaborative filtering model that learns article relevance from user interaction signals, specifically article view events. The model analyzes patterns of which articles users view together to identify related content, making view events the essential training data for generating recommendations.

Exam trap

Salesforce often tests the distinction between essential training data (user behavior signals like view events) and optional metadata (like publication dates or author names), leading candidates to mistakenly choose metadata that seems relevant but is not required for the collaborative filtering model to learn article relevance.

How to eliminate wrong answers

Option A is wrong because article publication dates are metadata that influence recency but are not used as primary training signals for collaborative filtering; the model learns relevance from user behavior, not timestamps. Option C is wrong because user job titles are demographic attributes that could be used for personalization but are not essential for the core recommendation model, which relies on interaction data like views. Option D is wrong because article author names are content metadata that do not provide the behavioral signals needed for the model to learn which articles are relevant to users.

Practice this question →
88
MCQeasy

A company wants to use Einstein Prediction Builder to predict customer churn. Which data preparation step is essential before building the model?

A.Ensure the data is in a Salesforce connected data source like Data Cloud.
B.Define the prediction objective and the target date field.
C.Create a formula field to calculate the churn probability.
D.Create a new custom object to store the prediction results.
AnswerB

The prediction objective (e.g., churn) is required to train the model.

Why this answer

Einstein Prediction Builder requires you to define the prediction objective (e.g., 'Will this customer churn?') and specify the target date field that marks the event. This step is essential as it tells the model what to predict and over what time window, enabling the automated feature engineering and model training process.

Exam trap

Salesforce often tests the misconception that data must come from Data Cloud or that you need to pre-create storage objects, when in fact the core prerequisite is simply defining the prediction objective and target date field.

How to eliminate wrong answers

Option A is wrong because while Data Cloud is a supported data source, it is not mandatory; Einstein Prediction Builder can also use standard or custom objects directly in Salesforce. Option C is wrong because formula fields cannot be used to calculate churn probability; the model generates probability scores automatically after training, and you do not pre-compute them. Option D is wrong because prediction results are stored automatically in a standard Salesforce object (PredictionResult) or can be written to a field on the record; you do not need to create a custom object for storage.

Practice this question →
89
MCQhard

A bank uses Einstein Discovery to generate insights about loan approval decisions. After deployment, they notice the model denies loans to a higher percentage of applicants from a certain postal code. Which action should be taken to ensure responsible AI?

A.Ignore the discrepancy because postal code is not a protected attribute
B.Retrain the model using only recent loan data
C.Audit model outcomes for fairness across demographic groups and retrain if needed
D.Remove the postal code field from the model
AnswerC

Bias audit and mitigation is a standard responsible AI practice.

Why this answer

Responsible AI requires auditing model outcomes for fairness across demographic groups, even when the disparity correlates with a non-protected attribute like postal code. In Einstein Discovery, postal code can act as a proxy for protected attributes such as race or socioeconomic status, and ignoring this could lead to discriminatory lending practices. Auditing allows the team to detect and mitigate bias, and retraining with fairness constraints ensures the model aligns with ethical AI principles.

Exam trap

Salesforce often tests the misconception that removing a sensitive feature (like postal code) automatically eliminates bias, when in reality proxy features and correlated variables can still cause unfair outcomes.

How to eliminate wrong answers

Option A is wrong because ignoring the discrepancy is irresponsible; postal code can be a proxy for protected attributes (e.g., race or income), and model fairness must be evaluated even if the field itself is not protected. Option B is wrong because retraining on only recent loan data does not address the root cause of bias; it may even amplify existing disparities if recent data still reflects historical biases or sampling issues. Option D is wrong because simply removing the postal code field does not guarantee fairness; other correlated features (e.g., income, credit history) can still encode the same bias, and the model may still discriminate indirectly through proxy variables.

Practice this question →
90
MCQmedium

A company uses Salesforce Data Platform to store customer data. They want to use this data to train an AI model for lead scoring, but they are concerned about data quality. Which step should they take first to ensure the data is suitable for AI?

A.Profile the data to identify missing values, outliers, and inconsistencies
B.Immediately normalize all numerical features
C.Create a labeled dataset using historical lead outcomes
D.Set up a data pipeline to stream data in real-time
AnswerA

Profiling is the first step to assess data quality.

Why this answer

Profiling the data is the essential first step because it systematically identifies missing values, outliers, and inconsistencies that degrade model performance. Without this baseline assessment, any subsequent normalization or labeling would be applied to flawed data, leading to unreliable lead scoring predictions. Salesforce Data Platform supports profiling via tools like Einstein Analytics or Data Prep, which scan fields for nulls, range violations, and format errors.

Exam trap

Salesforce often tests the misconception that data preparation begins with feature engineering (like normalization) or pipeline setup, rather than with foundational data quality assessment through profiling.

How to eliminate wrong answers

Option B is wrong because normalizing numerical features is a preprocessing step that should only occur after data quality issues (like missing values or outliers) have been identified and resolved; applying normalization prematurely can amplify the impact of corrupt data. Option C is wrong because creating a labeled dataset is a critical step for supervised learning, but it assumes the raw data is already clean and consistent, which is not the case when data quality is a concern. Option D is wrong because setting up a real-time data pipeline addresses data velocity and freshness, not data quality; streaming dirty data into the pipeline would only propagate errors faster.

Practice this question →
91
MCQhard

An organization uses Salesforce Data Cloud to unify customer data from multiple sources. They want to ensure that data lineage is tracked for AI models. Which practice supports data lineage?

A.Use data partitioning to improve query performance.
B.Implement role-based access control on datasets.
C.Maintain metadata that records source, transformations, and dependencies.
D.Regularly run data profiling to check completeness.
AnswerC

Metadata enables lineage tracking.

Why this answer

Maintaining metadata that records source, transformations, and dependencies is the correct practice because data lineage for AI models requires a complete audit trail of where data originated, how it was transformed, and its dependencies. In Salesforce Data Cloud, this metadata is captured through the Data Catalog and Data Lineage feature, which tracks the flow of data from source objects through calculated insights and segments to AI model inputs, ensuring transparency and reproducibility.

Exam trap

Salesforce often tests the distinction between data management practices that improve performance or security versus those that specifically support auditability and traceability, leading candidates to confuse data partitioning or access control with lineage tracking.

How to eliminate wrong answers

Option A is wrong because data partitioning improves query performance by dividing data into smaller segments, but it does not track the origin, transformation steps, or dependencies of data, which are essential for lineage. Option B is wrong because role-based access control (RBAC) governs who can view or modify datasets, but it provides no record of data provenance or transformation history. Option D is wrong because data profiling checks completeness, accuracy, and consistency of data, but it does not capture the sequence of transformations or source-to-target mappings required for lineage.

Practice this question →
92
MCQhard

A team is labeling text data for a sentiment analysis model. To ensure consistency and quality, which practice should they prioritize?

A.Use a single expert labeler for all data.
B.Use majority voting among multiple labelers.
C.Label all data by a single expert labeler.
D.Allow each labeler to interpret guidelines freely.
AnswerB

Majority voting aggregates judgments, improving accuracy and consistency.

Why this answer

Majority voting among multiple labelers reduces individual bias and errors, improving label consistency and quality for training data. This approach is standard in supervised learning for sentiment analysis because it aggregates diverse judgments, leading to more reliable ground truth labels.

Exam trap

Salesforce often tests the misconception that a single expert labeler guarantees higher quality, when in fact multiple labelers with majority voting reduce bias and improve reliability for training data.

How to eliminate wrong answers

Option A is wrong because using a single expert labeler introduces individual bias and lacks error checking, which can degrade model performance due to inconsistent or subjective labels. Option C is wrong because labeling all data by a single expert labeler is identical to Option A and suffers from the same lack of consensus and quality assurance. Option D is wrong because allowing each labeler to interpret guidelines freely leads to high inter-labeler variability, undermining consistency and making the dataset unreliable for training a robust model.

Practice this question →
93
MCQmedium

Refer to the exhibit. What effect does this masking policy have on the data used for training an Einstein model?

A.Only SSN is masked.
B.SSN and CreditCard fields are encrypted.
C.SSN and CreditCard fields are completely removed from training data.
D.SSN and CreditCard fields are partially masked, showing only the last four characters.
AnswerD

Explicitly defined by showLastFour and maskingType partial.

Why this answer

The masking policy in Einstein applies a partial mask to sensitive fields like SSN and CreditCard, showing only the last four characters while obscuring the rest. This ensures that the data used for training retains its structural utility for model learning without exposing full sensitive values, which is why option D is correct.

Exam trap

Salesforce often tests the distinction between masking, encryption, and removal, where candidates mistakenly think masking is equivalent to encryption or complete deletion, but masking specifically preserves partial data for model training while hiding sensitive details.

How to eliminate wrong answers

Option A is wrong because the masking policy applies to both SSN and CreditCard fields, not just SSN, as indicated by the exhibit showing both fields being masked. Option B is wrong because masking is not encryption; encryption transforms data into a ciphertext that requires a key to reverse, whereas masking irreversibly obscures parts of the data for privacy. Option C is wrong because the policy does not completely remove the fields; it partially masks them, leaving the last four characters visible for training purposes.

Practice this question →
94
Multi-Selectmedium

Which TWO data preparation steps are critical for ensuring high-quality training data?

Select 2 answers
A.Increasing dataset size by adding noise.
B.Removing duplicate records.
C.Normalizing all features.
D.Handling missing values appropriately.
E.Using only labeled data.
AnswersB, D

Duplicates can overrepresent certain patterns and skew model training.

Why this answer

Duplicate records in a dataset can cause the model to overfit to repeated patterns, biasing the learned distribution and reducing generalization. Removing duplicates ensures each data point contributes equally to training, which is essential for robust model performance.

Exam trap

Salesforce often tests the distinction between data preparation steps that ensure data quality (like removing duplicates and handling missing values) versus optional preprocessing or augmentation techniques, leading candidates to mistakenly select normalization or noise addition as critical steps.

Practice this question →
95
MCQeasy

A company plans to train an AI model using data from Salesforce CRM and an external marketing automation platform. What is the first step to unify these data sources in Data Cloud?

A.Define a Data Model that maps fields from both sources to a unified customer object
B.Create two separate Data Streams to bring data in
C.Build a Calculated Insight to merge the data
D.Set up a Data Transformation to blend the sources
AnswerA

Unifies the schema before ingestion.

Why this answer

Creating a data model that maps fields from both sources to a common object ensures consistency. Option B is wrong because data streams come after the model. Option C is wrong because Calculated Insights are for aggregations.

Option D is wrong because data transformations are applied later.

Practice this question →
96
MCQhard

A system administrator receives an error when running a Data Cloud data transform: 'Row-level security settings are preventing access to the source data.' The admin has appropriate permissions. What is the most likely cause?

A.The target object has field-level security.
B.The data stream is scheduled during maintenance.
C.The data transform is set to run as the admin's default user.
D.The source object has sharing rules that restrict access for the data transform's running user.
AnswerD

Row-level security is about sharing; the running user may not see all rows.

Why this answer

Data Cloud data transforms run under a specific running user context, and row-level security (RLS) settings on the source object can restrict that user's access to rows, even if the admin has broad permissions. The error indicates that the running user lacks visibility to certain source data rows due to sharing rules or RLS policies, which is a common cause of this specific error message.

Exam trap

Salesforce often tests the distinction between row-level security and field-level security, and candidates mistakenly choose field-level security (Option A) because they confuse the two concepts, not realizing the error message explicitly points to row-level restrictions.

How to eliminate wrong answers

Option A is wrong because field-level security (FLS) controls access to fields, not rows, and the error explicitly mentions 'row-level security,' not field-level. Option B is wrong because scheduled maintenance would typically cause a different error (e.g., 'service unavailable' or timeout), not a row-level security access denial. Option C is wrong because running as the admin's default user would inherit the admin's permissions, which should have access; the error states the admin has appropriate permissions, so the issue is with the running user context, not the default user setting.

Practice this question →
97
MCQhard

An organization is preparing data for Einstein Next Best Action. They have multiple action types (discounts, product suggestions, content). Which data model approach best ensures accurate recommendations?

A.Train a separate model per customer segment and then merge.
B.Create a separate model for each action type and combine results manually.
C.Build an ensemble of models and average their outputs.
D.Use a single model that includes all action types in the training data.
AnswerD

A unified model captures interactions between actions, leading to better optimization of the next best action.

Why this answer

Einstein Next Best Action is designed to learn from all action types simultaneously within a single model. By including all action types (discounts, product suggestions, content) in the training data, the model can capture cross-action patterns and relative effectiveness, leading to more accurate and contextually relevant recommendations. A unified model avoids fragmentation and ensures consistent scoring across actions.

Exam trap

Salesforce often tests the misconception that separate models per action or segment improve accuracy, when in fact Einstein Next Best Action requires a single unified model to learn cross-action patterns and deliver coherent recommendations.

How to eliminate wrong answers

Option A is wrong because training a separate model per customer segment and merging results introduces fragmentation and ignores cross-segment patterns, reducing the model's ability to generalize and leading to inconsistent recommendations. Option B is wrong because creating a separate model for each action type and manually combining results loses the interdependencies between actions, such as when a discount makes a product suggestion more effective, and adds unnecessary complexity without leveraging Einstein's built-in multi-action support. Option C is wrong because building an ensemble of models and averaging outputs does not align with Einstein Next Best Action's architecture, which expects a single unified model to optimize across all actions; averaging can dilute the signal from specific action types and reduce recommendation precision.

Practice this question →
98
MCQeasy

A company wants to use Einstein Article Recommendations to suggest knowledge articles to support agents. What is a prerequisite for this feature?

A.Articles must be of a specific type, such as FAQ.
B.The org must be enabled for Einstein features.
C.A case must be open for the recommendation to appear.
D.Knowledge articles must be created and published.
AnswerD

Articles must exist to be recommended.

Why this answer

Einstein Article Recommendations requires that knowledge articles are created and published in the Salesforce Knowledge base. The feature uses natural language processing (NLP) to match the context of a case or conversation with published articles, so unpublished or draft articles cannot be recommended. Without published articles, the AI model has no content to analyze or suggest.

Exam trap

Salesforce often tests the distinction between general Einstein enablement and feature-specific prerequisites, so candidates mistakenly select Option B thinking that enabling Einstein is the only requirement, when in fact published articles are the critical prerequisite for Article Recommendations to function.

How to eliminate wrong answers

Option A is wrong because Einstein Article Recommendations does not require articles to be of a specific type like FAQ; it works with any standard or custom article type defined in Salesforce Knowledge. Option B is wrong because while Einstein features generally require the org to be enabled for Einstein, this is a platform-level prerequisite for all Einstein services, not a specific prerequisite for Article Recommendations—the question asks for a prerequisite specific to this feature. Option C is wrong because a case does not need to be open for recommendations to appear; Einstein can also suggest articles in other contexts such as Chat, Email-to-Case, or even in the Knowledge tab without an open case.

Practice this question →
99
MCQeasy

Refer to the exhibit. A data transformation configuration is shown. Which of the following describes the outcome of applying this transformation?

A.The transformation is invalid because one-hot encoding cannot be combined with scaling.
B.Only 'color' is transformed; 'price' and 'weight' are unchanged.
C.'color' is one-hot encoded into multiple binary columns; 'price' and 'weight' are standardized to have mean 0 and variance 1.
D.'color' is scaled to [0,1] and 'price', 'weight' are one-hot encoded.
AnswerC

Correct interpretation of the config.

Why this answer

The transformation configuration applies a one-hot encoder to the 'color' categorical column, creating multiple binary columns, and applies a standard scaler to the 'price' and 'weight' numerical columns, centering them to mean 0 and scaling to unit variance. This is a common preprocessing pipeline that handles mixed data types appropriately.

Exam trap

Salesforce often tests the ability to distinguish which transformation applies to which column type, trapping candidates who confuse scaling with encoding or assume that different transformations cannot coexist in a single pipeline.

How to eliminate wrong answers

Option A is wrong because one-hot encoding and scaling can be combined in a single transformation pipeline; they are applied to different columns (categorical vs. numerical) and are not mutually exclusive. Option B is wrong because the transformation explicitly applies a standard scaler to 'price' and 'weight', so they are not unchanged; they are standardized. Option D is wrong because it reverses the operations: 'color' is one-hot encoded, not scaled to [0,1], and 'price' and 'weight' are standardized, not one-hot encoded.

Practice this question →
100
MCQeasy

Refer to the exhibit. What data quality issue does the exhibit reveal?

A.The Summer Sale campaign has duplicate records.
B.The Fall Clearance campaign has no response data.
C.The query syntax is incorrect.
D.The data is not normalized.
AnswerB

NonNullResponse is 0, meaning all responses are null.

Why this answer

The Fall Clearance campaign has zero non-null responses, indicating all response data is missing for that campaign.

Practice this question →
101
MCQmedium

An admin is troubleshooting Einstein Sentiment. The model returns high confidence but wrong sentiment (e.g., positive reviews labeled negative). What is the most likely issue?

A.The model was not retrained after the last data load.
B.The training data contains predominantly neutral examples.
C.The training data has incorrect labels for sentiment.
D.The field mapping for the sentiment field is incorrect.
AnswerC

Garbage in, garbage out: mislabeled training data leads to confident but incorrect classifications.

Why this answer

If the training data contains incorrect labels for sentiment, the model learns from erroneous ground truth, leading to high confidence in wrong predictions. In Einstein Sentiment, the model's accuracy depends directly on the quality and correctness of the labeled training data; mislabeled examples cause the classifier to associate features with the wrong sentiment class, resulting in confident but incorrect outputs.

Exam trap

Salesforce often tests the concept that high confidence does not imply high accuracy; candidates mistakenly assume retraining or data volume issues are the root cause, rather than recognizing that garbage-in (incorrect labels) leads to garbage-out (confident wrong predictions).

How to eliminate wrong answers

Option A is wrong because retraining after a data load does not fix incorrect labels; it would only reinforce the existing mislabeled patterns. Option B is wrong because predominantly neutral examples would bias the model toward neutral predictions, not cause high-confidence wrong sentiment (e.g., positive labeled as negative). Option D is wrong because incorrect field mapping would typically result in missing or misaligned data, not high-confidence misclassification of sentiment; the model would fail to train or predict altogether.

Practice this question →
102
MCQhard

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

A.Remove all 18 HIPAA identifiers from each record.
B.Generate synthetic data that mimics patient records.
C.Remove patient names and replace with IDs.
D.Anonymize data by aggregating into groups of 10 or more.
AnswerA

Safe Harbor method requires removal of all listed identifiers.

Why this answer

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

Exam trap

The trap here is that candidates often confuse de-identification with anonymization or pseudonymization, assuming that removing just names or aggregating data is sufficient, but Cisco tests the specific requirement of removing all 18 HIPAA identifiers under the Safe Harbor method.

How to eliminate wrong answers

Option B is wrong because generating synthetic data that mimics patient records does not satisfy the HIPAA Safe Harbor method; while synthetic data can be useful, it is not a recognized de-identification method under HIPAA Safe Harbor, which requires the removal of specific identifiers from actual data. Option C is wrong because removing patient names and replacing them with IDs alone does not meet the Safe Harbor standard, as it still leaves 17 other identifiers (e.g., dates, ZIP codes) that must be removed. Option D is wrong because anonymizing data by aggregating into groups of 10 or more is not a HIPAA Safe Harbor method; aggregation may reduce re-identification risk but does not guarantee removal of all 18 identifiers, and Safe Harbor requires explicit removal of those identifiers, not statistical aggregation.

Practice this question →
103
MCQmedium

A company wants to use customer purchase history to train a recommendation model. Which action is essential to comply with data privacy regulations?

A.Use only publicly available data.
B.Ignore regulations because data is internal.
C.Obtain explicit consent from customers.
D.Anonymize the data after training.
AnswerC

Explicit consent is legally required for processing personal data for AI training.

Why this answer

Data privacy regulations such as GDPR and CCPA require a lawful basis for processing personal data, and explicit consent is a primary lawful basis when using customer purchase history for training a recommendation model. Without obtaining explicit consent, the company would be processing personal data without a valid legal ground, violating regulations that mandate transparency and user control over their data.

Exam trap

The trap here is that candidates often assume internal data is exempt from privacy regulations or that anonymization after training retroactively fixes compliance, but regulations require a lawful basis before any processing begins.

How to eliminate wrong answers

Option A is wrong because using only publicly available data does not guarantee compliance; the data may still contain personal information subject to privacy regulations, and the model could inadvertently infer private attributes from public data. Option B is wrong because internal data is not exempt from data privacy regulations; laws like GDPR apply to any processing of personal data regardless of whether it is internal or external. Option D is wrong because anonymizing data after training does not address the requirement for a lawful basis at the time of collection and processing; the model may have already learned patterns from identifiable data, and retroactive anonymization does not cure the initial lack of consent.

Practice this question →
104
MCQhard

A large enterprise is using Einstein Lead Scoring and notices that the model score is not updating for leads created via a web-to-lead form. The leads have all required fields populated. The admin has verified that the model is active and the data source includes the Lead object. What could be causing the score to remain static?

A.The data source excludes leads created by web-to-lead
B.Web-to-lead leads are not supported by Einstein AI
C.The model is not yet activated
D.The model has not scored enough leads to start scoring new ones
AnswerD

Einstein requires a critical mass of scored records to calibrate the model before scoring new leads.

Why this answer

Einstein Lead Scoring requires a minimum number of scored leads (typically 500) before it begins scoring new leads. Until that threshold is met, the model remains in a 'training' or 'pending' state and will not update scores for any leads, including those from web-to-lead forms. The admin has confirmed the model is active and the Lead object is included, so the most likely cause is that the model has not yet processed enough leads to start scoring.

Exam trap

Salesforce often tests the concept that Einstein AI models require a minimum data threshold before they become operational, and candidates mistakenly assume that an 'active' model immediately scores all leads, ignoring the training/pending state requirement.

How to eliminate wrong answers

Option A is wrong because the data source for Einstein Lead Scoring includes all leads in the Lead object by default; there is no option to exclude leads based on creation method (e.g., web-to-lead). Option B is wrong because Einstein AI fully supports leads created via web-to-lead forms; there is no restriction on lead source. Option C is wrong because the admin has already verified that the model is active, so the model is not in an 'inactive' or 'not yet activated' state.

Practice this question →
105
MCQeasy

A company wants to use its data from Salesforce to train an Einstein AI model. However, they need to exclude records where the customer has opted out of data use. Which field should they configure in the Data Manager?

A.A checkbox field named Data Use Opt-Out
B.The Record Type of the object
C.A picklist field on the object
D.A formula field that evaluates to true
AnswerA

Salesforce Data Manager uses a checkbox field to mark records that should be excluded from AI training.

Why this answer

The Data Manager in Einstein AI uses a standard checkbox field named 'Data Use Opt-Out' to identify records that should be excluded from model training. When this checkbox is true, the system automatically filters out those records during data preparation, ensuring compliance with data privacy requirements.

Exam trap

Salesforce often tests the misconception that any field indicating consent can be used, but the Data Manager specifically requires a checkbox field named 'Data Use Opt-Out' to automatically filter records.

How to eliminate wrong answers

Option B is wrong because the Record Type field is used for business process differentiation and page layouts, not for indicating data usage consent; it has no built-in mechanism to exclude records from AI training. Option C is wrong because a picklist field, while it could theoretically store opt-out status, is not recognized by the Data Manager as a standard field for data exclusion; the system specifically looks for the checkbox field. Option D is wrong because a formula field that evaluates to true is not a standard field type that the Data Manager can interpret for opt-out filtering; the Data Manager requires a native checkbox field to trigger exclusion logic.

Practice this question →
106
MCQhard

A healthcare organization uses Salesforce to develop an AI model for patient readmission prediction. They must comply with HIPAA regulations. The dataset includes patient names, addresses, medical record numbers, and detailed clinical notes. The data scientist plans to train a supervised model using historical readmission outcomes. What is the most important data governance step before model training?

A.Use only aggregated data that does not include any patient-level details.
B.De-identify all protected health information (PHI) by removing or masking identifiers.
C.Obtain written patient consent for every record used in training.
D.Store the data in a separate, encrypted environment with access controls.
AnswerB

De-identification ensures compliance with HIPAA and protects patient privacy, allowing safe use of data for AI.

Why this answer

HIPAA mandates that protected health information (PHI) must be de-identified before it can be used for model training without patient authorization. Removing or masking identifiers such as names, addresses, and medical record numbers ensures the dataset no longer contains individually identifiable information, allowing the organization to comply with the HIPAA Privacy Rule while still using clinical notes for predictive modeling.

Exam trap

Salesforce often tests the misconception that security controls like encryption or access controls alone satisfy HIPAA compliance, when in fact de-identification is the primary requirement for using PHI in AI model training without patient consent.

How to eliminate wrong answers

Option A is wrong because using only aggregated data would remove the granular patient-level details needed to train a supervised model for readmission prediction, which requires individual outcomes to learn patterns. Option C is wrong because obtaining written patient consent for every record is impractical for large historical datasets and is not required under HIPAA if PHI is properly de-identified. Option D is wrong because while storing data in an encrypted environment with access controls is a good security practice, it does not address the core HIPAA requirement to de-identify PHI before using it for model training; encryption alone does not make data non-PHI.

Practice this question →
107
MCQeasy

When using Einstein Lead Scoring, which data source is most critical for generating accurate lead scores?

A.Lead source (e.g., Webinar, Trade Show)
B.Lead field update timestamps
C.Email open rates from marketing campaigns
D.Converted lead records with attached opportunities
AnswerD

The model learns from past conversions; opportunities show which leads actually became customers.

Why this answer

Converted lead records with attached opportunities are the most critical data source because Einstein Lead Scoring uses supervised machine learning to analyze historical patterns in leads that successfully converted into opportunities. By training on these converted records, the model learns which lead attributes and behaviors are predictive of conversion, enabling it to assign accurate scores to new leads. Without this historical conversion data, the model lacks the ground truth needed to distinguish high-quality leads from low-quality ones.

Exam trap

Salesforce often tests the misconception that any single lead attribute (like source or email engagement) is the most critical input, when in fact the model's accuracy depends entirely on having historical conversion data to learn from.

How to eliminate wrong answers

Option A is wrong because lead source is just one of many features Einstein Lead Scoring evaluates, but it is not the most critical data source; the model requires historical conversion outcomes to weight such features properly. Option B is wrong because lead field update timestamps indicate recency of activity but do not provide the conversion outcome data needed to train the predictive model. Option C is wrong because email open rates from marketing campaigns are behavioral signals that can be used as features, but they are insufficient without converted lead records to establish which behaviors actually correlate with successful conversions.

Practice this question →
108
Multi-Selecthard

A data scientist is using Einstein Discovery to analyze sales data. The model results show a high correlation between two predictor variables. Which TWO actions should the data scientist take?

Select 2 answers
A.Apply regularization.
B.Combine them into a single feature.
C.Include both to capture more information.
D.Increase the sample size.
E.Remove one of the correlated variables.
AnswersB, E

Creates a new variable that captures the combined effect.

Why this answer

Removing one correlated variable or combining them reduces multicollinearity.

Practice this question →
109
MCQhard

A data scientist discovers that an AI model used for loan approval predicts high default risk disproportionately for a specific demographic group. What is the first step to address this issue?

A.Use a different algorithm
B.Audit the training data for bias
C.Remove demographic features from the model
D.Retrain the model with more data
AnswerB

Auditing helps identify and mitigate bias in data.

Why this answer

Auditing the training data for bias helps identify if the model learned biased patterns. Option A is wrong because retraining with more data may not solve the bias if the new data also contains bias. Option C is wrong because removing demographic features may not eliminate bias if other correlated features exist.

Option D is wrong because changing the algorithm does not address biased data.

Practice this question →
110
MCQeasy

A retail company has implemented a Salesforce AI lead scoring model to prioritize high-value customers. After three months, the model's AUC-ROC score is only 0.55, indicating poor performance. The data scientist reviews the training data and finds that 20% of the records are exact duplicates due to multiple data imports from different sources. The duplicates have inconsistent target labels (some labeled 'converted', others 'not converted'). What should the data scientist do to improve model performance?

A.Downsample duplicates to reduce their impact but keep all records.
B.Use the duplicates as a separate class to indicate noisy data.
C.Remove all duplicate records and keep only one instance per duplicate group, resolving label conflicts by majority vote.
D.Keep all duplicates because they represent multiple interactions; increase model complexity to handle them.
AnswerC

This cleans the data, removes noise, and provides consistent labels, likely improving model performance.

Why this answer

Duplicate records with conflicting labels confuse the model. Removing duplicates and resolving label conflicts (e.g., by majority vote) is the most effective step to clean the data and improve performance.

Practice this question →
111
MCQmedium

Refer to the exhibit. A data scientist sees this error when training an Einstein Discovery model for customer churn prediction. What is the most likely reason for the error?

A.The field count (8) exceeds the maximum of 5 allowed fields.
B.The positive examples (180) are insufficient for the number of fields (8).
C.The model name contains a version number, which is not allowed.
D.The dataset has too few records (3200) for 8 fields.
AnswerB

50 per field * 8 = 400; 180 is below the threshold.

Why this answer

Einstein Discovery requires at least 50 positive examples per predictor field. With 8 fields, at least 400 positive examples are needed. Only 180 were provided, causing the error.

Practice this question →
112
MCQmedium

A company is preparing data for Einstein Prediction Builder to forecast lead conversion. They have historical data with fields like Lead Source, Industry, Number of Employees, and Converted (boolean). Which data preparation step is most critical?

A.Mix data from all lead sources without normalization
B.Ensure data completeness by handling missing values in Lead Source
C.Use only the last 3 months of data for training
D.Remove all records with outliers in Number of Employees
AnswerB

Completeness is a key data quality dimension; missing values in a predictor reduce model reliability.

Why this answer

Handling missing values in Lead Source is critical because Einstein Prediction Builder requires complete, high-quality data to train accurate predictive models. Missing categorical fields like Lead Source can introduce bias or cause the model to ignore important patterns in lead conversion. Ensuring data completeness through imputation or removal of incomplete records is a standard data preparation step for AI/ML in Salesforce.

Exam trap

Salesforce often tests the misconception that more data or aggressive cleaning (like removing outliers or using only recent data) always improves AI model accuracy, when in fact data completeness and representative sampling are more critical for supervised learning tasks like lead conversion prediction.

How to eliminate wrong answers

Option A is wrong because mixing data from all lead sources without normalization can introduce scale differences and skew model predictions; normalization is often required for numerical features, but the key issue here is that mixing without handling categorical consistency (like Lead Source) can degrade model performance. Option C is wrong because using only the last 3 months of data may not capture seasonal trends or sufficient historical patterns, leading to overfitting or poor generalization; Einstein Prediction Builder benefits from a broader historical window (e.g., 12-24 months) to learn conversion patterns. Option D is wrong because removing all records with outliers in Number of Employees can discard valuable data points that represent legitimate business segments (e.g., large enterprises) and reduce model robustness; outlier treatment should be context-aware, not automatic removal.

Practice this question →
113
MCQhard

Refer to the exhibit. A data analyst has defined this field mapping for Einstein Prediction Builder. Which data issue would most likely arise from this mapping?

A.The 'LeadSource' field should be mapped to 'Text' instead of 'Category' to preserve verbatim values
B.The 'Amount' field should be mapped to 'Category' to discretize the values
C.The 'Id' field should be excluded as it can cause data leakage and overfitting
D.The 'CloseDate' field should be mapped to 'Text' to avoid date parsing issues
AnswerC

Unique identifiers act as keys and should not be used as predictors.

Why this answer

Including the 'Id' field in Einstein Prediction Builder can cause data leakage and overfitting. The 'Id' field is a unique identifier that has no predictive value for the target outcome, but the model could learn to memorize specific records based on it, leading to poor generalization on unseen data.

Exam trap

Salesforce often tests the concept of data leakage by including a seemingly harmless field like 'Id', tricking candidates into thinking all fields should be mapped, when in fact unique identifiers must be excluded to prevent overfitting.

How to eliminate wrong answers

Option A is wrong because 'LeadSource' is a categorical field with a limited set of values (e.g., 'Web', 'Phone'), so mapping it to 'Category' is appropriate; mapping it to 'Text' would treat each unique value as a separate token, which is not suitable for prediction models. Option B is wrong because 'Amount' is a continuous numeric field, and mapping it to 'Category' would discretize it, losing granularity and potentially reducing model accuracy; it should remain as a numeric field. Option D is wrong because 'CloseDate' is a date field, and mapping it to 'Text' would prevent proper date parsing and feature extraction (e.g., day of week, month); Einstein Prediction Builder handles date fields natively for time-based features.

Practice this question →
← PreviousPage 2 of 2 · 113 questions total

Ready to test yourself?

Try a timed practice session using only Data for AI questions.

Start 20-question session