AI Associate Data for AI Practice Test 10 — 15 Questions

Question 1

A data analyst is troubleshooting Einstein Article Recommendations that are not showing up on the site. Which TWO checks should be performed first? (Choose 2)

Accepted Answer

Check that article view events are being captured in the data. Option C is correct because Einstein Article Recommendations relies on captured article view event data to train and drive its recommendation model, so if view events are not being logged, no recommendations can be generated or surfaced. Option E is correct because a recommendation model that is not published and active will not serve recommendations to the site, making this a fundamental first check. Option A is not required, since the 100-article threshold is not a documented prerequisite for Einstein Article Recommendations to function. Option B is irrelevant, as author profile permissions do not control whether recommendations appear to site visitors. Option D is incorrect because changing the recommendation frequency from daily to hourly does not fix a missing-data or inactive-model problem and is not a valid troubleshooting step.

Answer

Ensure at least 100 articles are in the knowledge base

Answer

Confirm that article authors have the correct profile permissions

Answer

Increase the recommendation frequency from daily to hourly

Question 2

A data scientist needs to feed customer interaction data into Einstein Discovery for predictive analysis. Which data format is required?

Accepted Answer

CSV. Einstein Discovery requires customer interaction data to be supplied as a flat, tabular dataset, and CSV is the supported format for importing that data into a Discovery dataset. CSV's row-and-column structure maps directly to Discovery's fields and records, which is why option A is correct. XML (B), Parquet (C), and JSON (D) are not the required format for feeding data into Einstein Discovery, even though they are common data-interchange or columnar-storage formats elsewhere.

Answer

XML

Answer

Parquet

Answer

JSON

Question 3

A company uses Salesforce Data Cloud to unify customer data from multiple sources for AI model training. After adding a new data source, model performance degrades significantly. What is the most likely cause?

Accepted Answer

Data duplication from overlapping sources. Data duplication from overlapping sources. When a new source overlaps with existing sources, Data Cloud's identity resolution can create duplicate unified profiles, inflating or skewing the training dataset and degrading model performance. Insufficient compute resources (A) would affect processing speed, not model accuracy. Data labeling errors (B) are possible but less directly tied to adding a source, and Data Cloud unification focuses on identity, not labels. A schema mismatch (C) would typically cause ingestion or mapping failures rather than subtle performance degradation.

Answer

Insufficient compute resources

Answer

Data labeling errors

Answer

Data schema mismatch

Question 4

Which data type is most commonly used for image recognition AI models?

Accepted Answer

Unstructured data. Unstructured data, because image recognition models consume raw pixel data from images, which has no predefined schema or tabular organization and is therefore classified as unstructured. Images (e.g., JPEG/PNG files) are the canonical example of unstructured data used to train convolutional neural networks and other vision models. Structured data (B) fits neatly into rows and columns, such as relational database records, and cannot represent raw image content directly. Time-series data (C) consists of timestamped measurements like sensor readings, and semi-structured data (D) follows loose formats such as JSON or XML, neither of which describes image files.

Answer

Structured data

Answer

Time-series data

Answer

Semi-structured data

Question 5

A team has limited labeled data for a Salesforce predictive model but wants to leverage a pre-trained model from a related task. Which machine learning approach should they use?

Accepted Answer

Transfer learning. Transfer learning (option A) is correct because it reuses a pre-trained model from a related task and fine-tunes it on the small labeled dataset, which is exactly the scenario of limited labeled data in Salesforce predictive modeling. This approach reduces the amount of task-specific labeled data needed and speeds up training by leveraging learned feature representations. Unsupervised learning (B) does not use labeled data and cannot directly fine-tune a pre-trained model for a supervised prediction target. Supervised learning (C) would require training from scratch on the limited labels, which is inefficient and prone to overfitting. Reinforcement learning (D) learns via reward signals from environment interaction, not from labeled examples or pre-trained task models.

Answer

Unsupervised learning

Answer

Supervised learning

Answer

Reinforcement learning

Question 6

After deploying an AI model in Salesforce, the data scientist notices high accuracy on the training set but poor accuracy on new incoming data. What is this phenomenon called?

Accepted Answer

Overfitting. Overfitting, which occurs when a model learns the training data too well, including its noise and idiosyncrasies, so it performs excellently on the training set but generalizes poorly to new, unseen data—exactly the high-training-accuracy, low-new-data-accuracy pattern described. Overfitting typically arises from an overly complex model, too many features relative to the number of training examples, or insufficient regularization. Data leakage (A) would inflate training performance via information from outside the training set, but it does not by itself describe the generalization gap as the named phenomenon. Underfitting (C) is the opposite problem, producing poor accuracy on both training and new data. Concept drift (D) refers to the statistical properties of the target changing over time, not to a model that fits training data too closely.

Answer

Data leakage

Answer

Underfitting

Answer

Concept drift

Question 7

To ensure AI model fairness and avoid biased outcomes, which practice is most critical when preparing training data?

Accepted Answer

Use balanced training data. Use balanced training data, because fairness in AI models depends primarily on the training set representing all relevant groups proportionally, so the model does not learn skewed patterns that disadvantage underrepresented populations. Balanced data reduces the risk of bias by preventing overrepresentation of one class or demographic, which would otherwise lead the model to favor that group in its predictions. Using only recent data (A) does not address representation and may even exclude historically marginalized groups, while increasing model complexity (B) can amplify bias rather than fix it. Adding more features (D) is also not a fairness remedy and can introduce proxy variables that worsen discriminatory outcomes.

Answer

Use only recent data

Answer

Increase model complexity

Answer

Use more features

Question 8

A company wants to integrate external customer behavior data into Salesforce to enhance AI predictions. Which Salesforce Data Cloud feature is specifically designed to ingest and map external data?

Accepted Answer

Data Streams. Data Streams (option B) is the Salesforce Data Cloud feature specifically designed to ingest external data from sources such as cloud storage, APIs, and third-party systems, then map that data into Data Cloud's data model for use in AI predictions. In this scenario, the company needs to bring in external customer behavior data, which is exactly what Data Streams handles through connectors and ingestion/mapping configuration. Apex triggers (A) are code-based automation on Salesforce records, not a Data Cloud ingestion feature. Einstein Studio (C) is used to build and train AI models on Data Cloud data, not to ingest external data. Flow Builder (D) is an automation tool for business processes and does not provide native external data ingestion into Data Cloud.

Answer

Apex triggers

Answer

Einstein Studio

Answer

Flow Builder

Question 9

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?

Accepted Answer

Audit the training data for bias. The first step when a model shows disparate impact is to audit the training data for bias — examining whether the data itself underrepresents or mislabels the affected group, or contains proxy variables correlated with the protected attribute. Without understanding the root cause in the data, any algorithmic fix (new model, dropping features, more data) is premature and may not address the actual source of bias.

Answer

Use a different algorithm

Answer

Remove demographic features from the model

Answer

Retrain the model with more data

Question 10

Which TWO are best practices for data labeling in AI projects? (Choose two.)

Accepted Answer

Have multiple labelers cross-check annotations. Option A is correct because having multiple labelers cross-check annotations (inter-annotator agreement) reduces individual bias and random errors, improving label quality and consistency, which is a core best practice in AI data labeling. Option E is correct because clear labeling guidelines ensure all annotators apply the same definitions and edge-case rules, which is essential for consistent, reproducible labels and directly supports model accuracy. Option B is not a best practice because labeling all available data immediately wastes resources; teams should prioritize and label data strategically, often starting with a representative subset. Option C is wrong because validation and test sets also require labeled data to evaluate model performance, not just training data. Option D is wrong because automated labeling tools alone are unreliable for many tasks and typically need human review or verification to ensure quality.

Answer

Label all available data immediately

Answer

Label only training data

Answer

Use automated labeling tools exclusively

Question 11

Which THREE are key considerations for data privacy when using AI models that process customer data? (Choose three.)

Accepted Answer

Limit data access to authorized personnel only. Option B is correct because limiting data access to authorized personnel only enforces least-privilege access controls, reducing the risk of unauthorized exposure of customer data processed by AI models. Option C is correct because obtaining user consent for data usage satisfies core privacy principles such as transparency and lawful processing, ensuring customers agree to how their data is used by the AI system. Option E is correct because anonymizing personally identifiable information (PII) removes or irreversibly alters identifiers, so even if data is exposed or reused, individuals cannot be re-identified. Option A is not appropriate because storing data indefinitely for future use violates data minimization and retention limits, increasing privacy and compliance risk. Option D, while a valuable security control, is not marked correct here and addresses protection of data rather than the privacy considerations of consent, access limitation, and anonymization emphasized in this scenario.

Answer

Store data indefinitely for future use

Answer

Encrypt data in transit and at rest

Question 12

Which TWO are common data quality issues that negatively impact AI model performance? (Choose two.)

Accepted Answer

Missing values. Missing values (B) are a classic data quality defect: nulls or gaps in features cause models to either drop records, fail to train, or learn biased patterns, and they must be handled via imputation, deletion, or masking before training. Outliers (C) are also a data quality issue because extreme or erroneous values distort statistical distributions, skew learned weights, and degrade generalization, requiring detection through methods like z-scores, IQR, or isolation forests. Multicollinearity (A) is a modeling/statistical concern about correlated predictors rather than a data quality defect, and it primarily affects interpretability of linear models. Data volume (D) refers to dataset size, which is a quantity/scaling matter, not a quality flaw. High dimensionality (E) is a structural property of the feature space (the curse of dimensionality) rather than a data quality problem per se.

Answer

Multicollinearity

Answer

Data volume

Answer

High dimensionality

Question 13

What is the primary purpose of this policy?

Accepted Answer

Data governance for AI. Data governance for AI, because a policy in this context defines the rules, roles, and controls for how data is collected, managed, and used in AI systems, ensuring compliance, privacy, and ethical use. Data governance for AI specifically addresses accountability, data quality, bias mitigation, and regulatory alignment across the AI lifecycle. The other options do not fit: data integration focuses on combining data from multiple sources, data transformation on converting data formats or structures, and data backup on copying data for recovery—none of which describe the primary purpose of a policy governing AI data use.

Answer

Data integration

Answer

Data transformation

Answer

Data backup

Question 14

A Salesforce admin is preparing a dataset for Einstein Prediction Builder. The dataset contains a field "Income" with many missing values. The admin wants to minimize bias in the model. What is the best practice?

Accepted Answer

Review the pattern of missingness and document reasons, then decide on imputation. The best practice for handling missing data in Einstein Prediction Builder is to first understand why values are missing—reviewing the pattern of missingness and documenting the reasons—before choosing a treatment such as imputation or exclusion, since the missingness mechanism determines whether a method introduces bias. Deleting all rows with missing Income (A) can introduce selection bias and shrink the training set, especially if missingness correlates with the target. Filling with the mean (C) ignores the missingness pattern and distorts the distribution and relationships with other fields, while replacing with 0 (D) injects a false value that can severely bias the model if 0 is not a meaningful Income value.

Answer

Delete all rows where Income is missing

Answer

Fill missing values with the mean of Income

Answer

Replace missing values with 0

Question 15

When training an Einstein Discovery model, which data type is not supported as a predictor field?

Accepted Answer

Multi-select picklist. Multi-select picklists are not supported as predictors in Einstein Discovery. Numeric, picklist, and date fields are supported.

Answer

Numeric

Answer

Picklist

Answer

Date