AI Associate Practice Test 25 — 15 Questions

Question 1

To integrate external data into Salesforce for AI, which tool is recommended by Salesforce for building data pipelines?

Accepted Answer

Salesforce Data Pipelines. Salesforce Data Pipelines (option C) is the recommended tool for building data pipelines to integrate external data into Salesforce for AI, as it provides a modern, low-code solution for ingesting, transforming, and harmonizing data from external sources. It is purpose-built for data integration and preparation, aligning with Salesforce's AI and Data Cloud strategy. Salesforce Connect (option A) is for real-time external object virtualization, not for building pipelines. Data Export Service (option B) is for extracting data out of Salesforce, and Apex Data Loader (option D) is a bulk data loading utility, neither of which is designed for building external data pipelines.

Answer

Salesforce Connect

Answer

Data Export Service

Answer

Apex Data Loader

Question 2

A company wants to use Einstein Next Best Action but needs to ensure data privacy. What is the required step for anonymizing customer data in Data Pipelines?

Accepted Answer

Use the Data Mask transformation in Data Pipelines. The correct answer is D: Use the Data Mask transformation in Data Pipelines. Data Pipelines provides a built-in Data Mask transformation that lets you obfuscate or anonymize sensitive customer fields (such as names, emails, or IDs) as data flows through the pipeline, which is the required step to protect privacy before feeding data into Einstein Next Best Action. Option A is wrong because a sandbox with scrambled data is a separate Salesforce environment feature, not a Data Pipelines anonymization step. Option B is wrong because Einstein does not automatically anonymize source data on its own. Option C is wrong because Permission Sets control user access to fields, but they do not anonymize or mask the underlying data.

Answer

Create a sandbox with scrambled data

Answer

Rely on Einstein's built-in anonymization

Answer

Use Permission Sets to hide fields from users

Question 3

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

Accepted Answer

Narrow AI is designed for a specific task, while general AI can understand and reason across a wide range of tasks. Narrow AI (also called weak AI) is designed to perform a specific task or a limited set of tasks, such as image recognition or language translation. General AI (strong AI) would possess human-like general intelligence, able to understand, learn, and reason across a wide range of tasks. This distinction is the core definition tested.

Answer

Narrow AI is rule-based, while general AI uses machine learning

Answer

Narrow AI requires labeled data, while general AI can learn without data

Answer

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

Question 4

Refer to the exhibit. An organization implements this AI fairness policy for their Einstein Prediction Builder model. What is the most significant ethical gap in this policy?

Accepted Answer

It does not specify a process for reviewing and remediating models that fail the fairness threshold. The policy states a fairness threshold and a monitoring cadence but never defines who reviews a failing model, what remediation steps follow, or how a model is retrained, replaced, or retired, which is the critical ethical gap for an enforceable AI fairness policy. A fairness policy is only meaningful if it closes the loop from detection to corrective action, including accountability, escalation, and documentation of outcomes. Options A and B are not the most significant gap here: the number of protected attributes and the choice of demographic parity are design considerations that may be appropriate depending on the exhibit's context, but they are not the missing governance mechanism. Option D is also not the key issue, since monitoring frequency alone does not ensure that failures are acted upon.

Answer

It does not include enough protected attributes

Answer

The demographic parity metric is inappropriate for this use case

Answer

The monitoring schedule is too frequent

Question 5

Refer to the exhibit. A data analyst receives an error when trying to use this model configuration for Einstein AI predictions. Which issue is most likely causing the error?

Accepted Answer

The feature field "Usage__c" does not exist in the data source.. The correct answer is A: the feature field "Usage__c" does not exist in the data source. In Einstein AI model configuration, every field referenced as a feature or target must exist in the underlying dataset; if "Usage__c" is missing, the model cannot build the feature vector and prediction fails with a field-not-found error. Option B is incorrect because a 0.8 split ratio is a standard, allowed value for classification models. Option C is incorrect because the prediction window is not required to be longer than the training window in that way, and this would not produce the described field error. Option D is incorrect because Einstein AI can handle text target fields for classification, so a text target alone would not cause this error.

Answer

The split ratio of 0.8 is not allowed for classification.

Answer

The prediction window is shorter than the training window.

Answer

The target field "Churn__c" is a text field instead of an integer.

Question 6

Which TWO actions are required to prepare data for an Einstein Discovery model?

Accepted Answer

Ensure the data is stored in a Salesforce object or a connected data source.. Option C is correct because Einstein Discovery requires the source data to reside in a Salesforce object or be accessible through a connected data source (such as a CRM Analytics dataset or external connection) before a model can be created. Option E is correct because defining the outcome (dependent) field is mandatory — the model must know which variable it is predicting, whether numeric or categorical, to build the story and generate predictions. Options A, B, and D are not required: Einstein Discovery automatically handles missing values rather than requiring their removal, it can automatically select predictors (manual selection of exactly 10 is not a requirement), and it internally splits data for training and validation rather than requiring a separate user-created dataset.

Answer

Remove all records with missing values in any field.

Answer

Select exactly 10 predictor fields manually.

Answer

Create a separate dataset for training and validation.

Question 7

An admin wants to create a prompt template that generates a personalized email to a lead. Which tool should they use to create and manage prompt templates for Einstein GPT?

Accepted Answer

Prompt Builder. Prompt Builder is the correct tool because it is specifically designed within the Einstein GPT platform to create, version, and manage prompt templates for generative AI use cases, such as generating personalized emails to leads. Unlike other options, Prompt Builder provides a no-code interface to define the prompt structure, variables (like lead name or company), and output format, directly integrating with Salesforce data.

Answer

Einstein Copilot

Answer

Einstein GPT Builder

Answer

Einstein Next Best Action

Question 8

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

Accepted Answer

Protecting customer data privacy during model training and inference. Option A is correct because deploying an AI chatbot that processes customer conversations involves collecting and using personal data, so privacy must be protected during both model training and inference, typically through data minimization, anonymization, and compliance with regulations such as GDPR or CCPA. Option C is correct because biased training data can cause the chatbot to produce unfair, discriminatory, or offensive responses, so bias detection and mitigation are essential ethical safeguards. Option D is correct because transparency requires disclosing to customers when they are interacting with an AI rather than a human agent, which supports informed consent and trust. Option B is not correct because using the most complex deep learning architecture is an engineering choice, not an ethical consideration, and complexity does not inherently improve fairness or privacy. Option E is not correct because maximizing training epochs is a hyperparameter tuning decision that can cause overfitting and has no direct ethical justification.

Answer

Using the most complex deep learning architecture available

Answer

Maximizing the number of training epochs

Question 9

A sales team uses Einstein Lead Scoring. They notice leads from certain industries are always low-scored. What should they do?

Accepted Answer

Review training data for bias. Low scores for specific industries often indicate bias in the training data, where historical lead data may have underrepresented or mislabeled those industries. Reviewing the training data for bias allows the team to identify and correct such imbalances, ensuring the Einstein Lead Scoring model produces fair and accurate predictions across all segments.

Answer

Retrain the model weekly

Answer

Ignore the scores

Answer

Use a different AI system

Question 10

Which TWO of the following are considered core ethical principles in AI according to Salesforce’s AI Ethics?

Accepted Answer

Accountability. According to Salesforce's AI Ethics guidelines, Accountability (A) is a core principle because organizations must take responsibility for the outcomes of their AI systems, ensuring human oversight and the ability to explain and remedy decisions. Transparency (C) is also a core principle, as AI systems should be designed so that their decision-making processes are open, understandable, and auditable by stakeholders. These two principles are explicitly part of Salesforce's ethical framework, alongside others such as fairness, reliability, and human-centered design. Popularity (B), Speed (D), and Profitability (E) are not ethical principles in this context; they are business or performance metrics that do not address the moral obligations of AI development and deployment.

Answer

Popularity

Answer

Speed

Answer

Profitability

Question 11

A company deploys an AI chatbot for customer service. After training on historical chats, the chatbot frequently gives incorrect answers to minority language queries. What is the likely cause?

Accepted Answer

Data imbalance. Data imbalance, because the historical chat data likely contains far fewer examples of minority language queries, so the model learns the majority language patterns well but has insufficient signal to answer minority language queries accurately. This skewed class/language distribution is a classic cause of poor performance on underrepresented groups. Underfitting (A) would show poor performance broadly, including on majority language queries, not specifically minority language ones. Lack of compute resources (C) affects training speed or model size, not language-specific accuracy. Overfitting (D) would typically cause poor generalization overall or memorization of training data, but the scenario points to a systematic gap tied to underrepresented language data.

Answer

Underfitting

Answer

Lack of compute resources

Answer

Overfitting

Question 12

A customer service manager wants to use Einstein Bots to handle common inquiries. They are concerned about the bot generating offensive responses. Which Einstein Trust Layer feature should they enable to minimize this risk?

Accepted Answer

Toxicity detection. Toxicity detection in the Einstein Trust Layer scans both prompts and completions for harmful, offensive, or toxic content and can block or flag such responses. Enabling it directly addresses the manager's concern about the bot generating offensive replies, without needing custom model changes.

Answer

Zero Data Retention

Answer

Grounding

Answer

PII masking

Question 13

A company deployed an AI model for lead scoring. After several months, they notice that leads from certain geographic regions consistently receive higher scores than leads from other regions with similar demographic profiles. The company wants to ensure ethical AI usage. What should they do first?

Accepted Answer

Retrain the model using a more diverse and balanced training dataset.. Retraining with more diverse data addresses potential bias at the source. Option A adjusts thresholds without fixing root cause. Option C ignores the discrepancy, which does not address the bias. Option D removes a feature that may be relevant but could still leak bias through correlated features.

Answer

Adjust the scoring thresholds for each region to equalize scores.

Answer

Ignore the discrepancy since the model overall accuracy is high.

Answer

Remove the geographic region feature from the model completely.

Question 14

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 15

When preparing data for Einstein Next Best Action, which two aspects must be considered for compliance with data privacy regulations? (Choose two.)

Accepted Answer

Consent management. Data privacy regulations such as GDPR and CCPA require that organizations track and honor the consent status of individuals before using their data to generate recommendations, so Einstein Next Best Action must respect consent flags when evaluating recommendation strategies. Option E (Data masking of personally identifiable information (PII)) is correct because PII must be protected or obfuscated before it is processed or displayed in recommendation outputs, ensuring that sensitive fields like names, emails, or phone numbers are not exposed to unauthorized users. The unmarked options do not belong: data compression (A) is a storage/performance technique with no bearing on privacy compliance, color coding fields (B) is purely a visualization choice, and indexing speed for real-time recommendations (C) is a performance consideration rather than a regulatory requirement.

Answer

Data compression

Answer

Color coding fields in the dataset

Answer

Indexing speed for real-time recommendations