AI Associate Data for AI Practice Test 12 — 15 Questions

Question 1

A large enterprise uses Data Cloud to power an Einstein model for lead scoring. The model's feature pipeline includes dozens of fields from multiple data streams. Performance has degraded, and the team suspects slow feature retrieval. What is the most efficient way to speed up feature computation in Data Cloud?

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

Accepted Answer

Use Calculated Insights to pre-compute and cache common features. The correct option is C: Use Calculated Insights to pre-compute and cache common features. Calculated Insights in Data Cloud are designed to compute and persist aggregated or derived metrics (features) ahead of time, so the Einstein model's feature pipeline can retrieve them quickly instead of recalculating them on every scoring request, directly addressing the slow feature retrieval. Option A (increasing data stream parallelism) affects ingestion throughput, not feature computation latency. Option B (external caching in the application layer) adds complexity and does not leverage Data Cloud's native feature computation. Option D (moving all data to a single data lake object) does not optimize feature computation and may even worsen performance by consolidating unrelated streams.

Answer

Increase the parallelism of the data streams

Answer

Implement external caching in the application layer

Answer

Move all data to a single data lake object

Question 2

A financial institution must ensure that customer data used for AI models does not expose personally identifiable information (PII) to unauthorized users. Which Data Cloud feature should be applied to the data model?

Accepted Answer

Apply data masking and field-level security on sensitive fields. Data Cloud supports data masking and field-level security, which allow sensitive PII fields to be obfuscated or restricted while still permitting the data to be used in AI models by authorized users. This directly addresses the requirement to prevent unauthorized exposure of PII without destroying the underlying data model. Option A is too destructive and would remove data needed for AI processing, while Option B only aggregates data and does not protect field-level PII exposure. Option D is insufficient because object-level permissions do not restrict access to individual sensitive fields within an object.

Answer

Delete PII fields from the data model

Answer

Use Calculated Insights to aggregate sensitive data only

Answer

Rely on user permissions to restrict access to the entire object

Question 3

A data architect notices that a Data Stream from an external ERP system is failing intermittently with schema mismatch errors. The ERP team says the schema changes occasionally. What is the most effective long-term solution?

Accepted Answer

Implement Data Stream schema validation and flexible mapping. Implementing schema validation and flexible mapping allows the Data Stream to accommodate unexpected schema changes from the ERP system gracefully, addressing the root cause. Option C is wrong because increasing retries does not fix schema mismatches. Option B is wrong because changing the primary key does not address schema changes. Option D is wrong because manual notifications are not scalable or reliable.

Answer

Use a different primary key for the data model

Answer

Increase the number of retries for the Data Stream

Answer

Ask the ERP team to manually notify of changes

Question 4

A news outlet wants to build an AI model that predicts article popularity using real-time social media mentions. Which data source type should they use to ingest tweets?

Accepted Answer

Data Stream with Ingestion API connector. The correct option is D, Data Stream with Ingestion API connector, because ingesting real-time social media mentions such as tweets requires a streaming ingestion mechanism that can continuously receive and process high-velocity data as it arrives. An Ingestion API connector is designed for exactly this purpose, enabling near real-time data flow into the platform for use by an AI model predicting article popularity. The other options do not fit: A, Calculated Insight, refers to derived metrics rather than raw ingestion; B, Data Transform, is used to reshape or process data after ingestion; and C, Data Lake Object, is a storage abstraction, not a real-time ingestion source.

Answer

Calculated Insight

Answer

Data Transform

Answer

Data Lake Object

Question 5

A manufacturer wants to improve demand forecasting by enriching its CRM orders with external demographic data. The external data is available via a SOAP API. How should the data architect implement this?

Accepted Answer

Configure a Data Action to call the API on a schedule or trigger. Option A is correct because a Data Action is the Pega mechanism designed to invoke external services such as a SOAP API, either on a schedule or triggered by an event, to enrich CRM order data with demographic information. Data Actions can map the SOAP response fields directly into the case or data model for use in demand forecasting. Option B is incorrect because a Data Transform only manipulates data already available in the application and cannot call an external API. Option C is incorrect because a Data Stream is used for batch or real-time ingestion of data into a repository, not for on-demand SOAP service calls. Option D is incorrect because a Calculated Insight computes derived values from existing data and does not invoke external APIs.

Answer

Use a Data Transform to pull the external data

Answer

Set up a Data Stream to continuously ingest the external API

Answer

Create a Calculated Insight to reference the API

Question 6

Which TWO of the following are valid methods to improve data quality in Data Cloud before training an AI model?

Accepted Answer

Use a Data Transform to filter out invalid records. Option B is correct because a Data Transform in Data Cloud lets you apply SQL-based transformations to cleanse and filter records—such as removing rows with null or malformed key fields—so invalid data never reaches the model training dataset. Option E is correct because enabling data profiling on data streams surfaces metrics like completeness, uniqueness, and value distributions, which is exactly how you identify quality problems (duplicates, missing values, outliers) before training. Option A is not a quality method; increasing ingestion frequency only changes how often data lands, not its accuracy or consistency. Option C, data retention policies, governs how long data is stored for compliance and storage management, not its correctness. Option D, Calculated Insights, computes aggregate metrics across DMO relationships for analytics and activation; while it can surface anomalies, it is not a primary data-quality remediation method and is not the best answer for improving data quality before training.

Answer

Increase data stream ingestion frequency

Answer

Implement data retention policies

Answer

Use Calculated Insights to detect anomalies

Question 7

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

Accepted Answer

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

Answer

Remove all records with missing values

Answer

Use raw data directly without any transformation

Question 8

A financial services firm uses Data Cloud to enrich sales data with external credit scores via an API. They set up a Data Action to call the credit bureau API for each new lead. Over time, API costs are rising, and the action is slowing down lead processing. They only need credit scores for leads with a high probability of conversion. What is the best approach to reduce costs and improve performance?

Accepted Answer

Use a Calculated Insight to score leads based on internal data and only invoke the Data Action for leads with a high probability of conversion. A Calculated Insight can score leads using existing internal data, and the Data Action is then invoked only for leads that meet the high-probability-of-conversion threshold, directly reducing the number of external credit bureau API calls and therefore cost and processing latency. This conditional invocation targets the stated business need: credit scores are only required for high-conversion leads. Option A does not reduce API costs or improve performance because it removes automation and relies on manual work. Option B filters only on incomplete data, which is unrelated to conversion probability and may still call the API for many low-value leads. Option C changes timing to daily batch but still calls the API for every lead, so it does not reduce the number of paid API calls.

Answer

Remove the Data Action and manually verify credit scores for top leads

Answer

Apply a Data Transform to filter leads that have incomplete data before the Data Action

Answer

Schedule the Data Action to run daily in batch instead of real-time

Question 9

A non-profit organization uses Data Cloud to manage donor data from multiple sources (email campaigns, event attendance, donations). They want to use an AI model to predict future donations. The data scientist says the model needs a unified view of each donor with consistent fields. What is the first step the data architect should take in Data Cloud to enable this?

Accepted Answer

Define a Data Model that maps each source's fields to a common donor object. Define a Data Model that maps each source's fields to a common donor object. In Data Cloud, the Data Model (Data Cloud canonical model plus custom objects) is what provides the unified, consistent schema that harmonizes disparate source fields into a single donor profile, which is the prerequisite for any AI model needing consistent inputs. Data Streams (B) only ingest raw source data and do not by themselves reconcile differing field names or semantics across email, events, and donations. Data Actions (C) are used for triggering workflows or activations, not for establishing the unified schema, and training on raw data (D) skips the required harmonization step and would produce inconsistent features.

Answer

Create a Data Stream for each source (email, events, donations)

Answer

Set up Data Actions to clean data at each source

Answer

Immediately start training the AI model on raw data from the streams

Question 10

A company is preparing customer data to train a custom AI model for sentiment analysis. Which two data preparation best practices should they follow? (Choose two.)

Accepted Answer

Ensure data is representative of all customer demographics.. Option B is correct because training data must be representative of all customer demographics to avoid sampling bias, which would cause the sentiment analysis model to perform poorly or unfairly for underrepresented groups. Option E is correct because anonymizing personally identifiable information (PII) before training satisfies data privacy and compliance requirements (e.g., GDPR, CCPA) and reduces the risk of leaking sensitive customer data into the model or its outputs. Option A is not a best practice because restricting data to only the last month introduces recency bias and may not capture the full range of sentiment expressions or seasonal patterns. Option C is wrong because dropping all records with missing values can discard valuable data and further skew the dataset; imputation or careful handling is usually preferred. Option D is wrong because labeling by a single annotator introduces individual bias and inconsistency; multiple annotators with inter-annotator agreement checks yield more reliable labels.

Answer

Use only data from the last month.

Answer

Remove all records with missing values.

Answer

Label data manually by a single annotator.

Question 11

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?

Accepted Answer

Remove all duplicate records and keep only one instance per duplicate group, resolving label conflicts by majority vote.. The model's poor AUC-ROC (0.55) is largely due to noisy labels from duplicate records with conflicting targets. Removing duplicates and resolving label conflicts via majority vote cleans the training data, reducing label noise and allowing the model to learn the true underlying pattern. This directly addresses the root cause: inconsistent labels confuse the model, leading to near-random predictions. Deduplication is a standard data quality step before training any supervised model.

Answer

Downsample duplicates to reduce their impact but keep all records.

Answer

Use the duplicates as a separate class to indicate noisy data.

Answer

Keep all duplicates because they represent multiple interactions; increase model complexity to handle them.

Question 12

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?

Accepted Answer

Oversample the minority class using SMOTE to create synthetic churn examples.. The dataset is highly imbalanced (5% churn), so the model achieves high accuracy by predicting the majority class but misses most churners (20% recall). Oversampling the minority class with SMOTE creates synthetic churn examples, giving the model more signal to learn churn patterns and improving recall for the minority class. This directly targets the class imbalance causing low recall.

Answer

Increase the regularization parameter to prevent overfitting.

Answer

Collect more data, especially of churned customers.

Answer

Undersample the majority class to match the minority class size.

Question 13

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?

Accepted Answer

De-identify all protected health information (PHI) by removing or masking identifiers.. 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.

Answer

Use only aggregated data that does not include any patient-level details.

Answer

Obtain written patient consent for every record used in training.

Answer

Store the data in a separate, encrypted environment with access controls.

Question 14

A company is building a chatbot using Einstein Bot's AI capabilities. They want to train intent recognition using historical chat transcripts. The transcripts contain many typos (e.g., 'hellp' instead of 'help') and slang (e.g., 'gonna' instead of 'going to'). The initial model performs poorly, misclassifying many intents. What data cleaning step is most important?

Accepted Answer

Normalize text by applying spell-correction and replacing slang with standard terms.. Normalizing text by applying spell-correction and replacing slang with standard terms is the most important data cleaning step. This reduces vocabulary variability and helps the intent recognition model learn consistent patterns. Typos and slang can cause the model to treat similar intents as different, leading to misclassifications. Normalization ensures that variations like 'hellp' and 'help' are treated as the same token, improving model accuracy.

Answer

Use a spell-checker only for words that appear infrequently.

Answer

Keep the raw text as is because it reflects real user behavior.

Answer

Remove all messages that contain typos or slang to clean the dataset.

Question 15

A multinational corporation uses Salesforce AI to analyze customer feedback across multiple languages. They have 10,000 English reviews, 2,000 Spanish reviews, and 500 French reviews. The sentiment model performs well on English (F1=0.85) but poorly on French (F1=0.40). The data scientist wants to improve French sentiment performance without collecting new data. What should they do?

Accepted Answer

Apply data augmentation to the French reviews using back-translation (translate to another language and back) to create more training examples.. Applying data augmentation to the French reviews using back-translation is the best approach because it generates additional training examples without collecting new data. Back-translation involves translating text to another language and back to the original, creating paraphrased versions that can improve model robustness. This directly addresses the poor French performance by increasing the effective size and diversity of the French training set.

Answer

Translate all French reviews to English and train only on English data.

Answer

Use a multilingual pre-trained model without any additional French data.

Answer

Remove French data and use only English and Spanish to avoid imbalance.