AI Associate Practice Test 26 — 15 Questions

Question 1

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

Accepted Answer

Relevant features selected for the prediction task. Option A is correct because feature relevance directly determines whether the model learns signals that actually predict the CRM outcome (e.g., churn, lead scoring); irrelevant or noisy features degrade accuracy regardless of algorithm choice. Option C is correct because representative training data ensures the model's learned distribution matches the real customer population it will score, preventing bias and poor generalization on live CRM records. Option E is correct because clean and complete data removes missing values, duplicates, and inconsistent entries that would otherwise corrupt training and skew predictions. Option B is not correct because a large number of parameters increases capacity and risks overfitting; it does not by itself improve accuracy. Option D is not correct because inference speed is a performance/latency concern, not an accuracy factor.

Answer

Large number of model parameters

Answer

Higher inference speed

Question 2

While building a prediction model in Einstein Studio, the system warns about "high cardinality" for a categorical field. What should the admin do?

Accepted Answer

Use frequency encoding or binning to reduce cardinality. The correct option is B: use frequency encoding or binning to reduce cardinality. High cardinality means a categorical field has too many distinct values, which can cause overfitting and poor model performance, so frequency encoding (replacing categories with their occurrence counts) or binning (grouping rare categories together) reduces the number of distinct levels while preserving useful signal. Option A is wrong because converting a categorical field to numeric type does not reduce the number of distinct values and can mislead the model into treating category codes as ordered quantities. Option C is overly aggressive since the field may still carry predictive value, and option D is unrelated because adding more trees increases model complexity rather than addressing the cardinality issue.

Answer

Convert the field to a numeric type

Answer

Remove the field from the model

Answer

Increase the model complexity by adding more trees

Question 3

An admin is training a new Einstein Prediction Builder model to predict whether a support case will be escalated (binary). They have selected the prediction field 'Escalated__c' and the data set of all cases from the past year. Which step is essential to ensure the model can distinguish between escalated and non-escalated cases?

Accepted Answer

Ensure the 'Escalated__c' field has both 'True' and 'False' values in the training data.. Einstein Prediction Builder requires the target prediction field to contain both positive and negative examples (e.g., 'True' and 'False') in the training data. Without both values, the model cannot learn the decision boundary between escalated and non-escalated cases, making binary classification impossible.

Answer

Set the prediction window to the next 30 days.

Answer

Select at least 20 features from the case object.

Answer

Ensure the data set contains at least 500 records.

Question 4

A data scientist is building a churn prediction model. What THREE factors are most critical for model success?

Accepted Answer

Relevant features correlated with churn. Option B is correct because feature relevance drives predictive power: features that genuinely correlate with churn (e.g., tenure, usage decline, support tickets) give the model signal, whereas irrelevant features add noise and risk overfitting. Option C is correct because a representative dataset covering all customer segments ensures the model generalizes across the population and avoids sampling bias that would make predictions unreliable for underrepresented groups. Option D is correct because supervised churn prediction depends on accurate, consistent historical labels; mislabeled or dirty data teaches the model wrong patterns regardless of algorithm choice. Option A is not critical—more parameters increase capacity and overfitting risk without guaranteeing better performance, and simpler models often suffice. Option E is not inherently critical—dataset size helps only when data is relevant, representative, and well-labeled; large but biased or noisy data can degrade results.

Answer

High number of model parameters

Answer

Large dataset size

Question 5

A multinational corporation uses Einstein Discovery to predict employee performance. An audit reveals potential bias against employees in certain countries. Which THREE actions should they take to address ethical concerns? (Choose three.)

Accepted Answer

Investigate the training data for biased labels or sampling bias across countries. Option A is correct because bias in Einstein Discovery typically originates in the training data, so auditing for biased labels or sampling bias across countries is the essential first step to identify and remediate the root cause. Option B is correct because engaging local ethics representatives from affected regions provides the cultural and legal context needed to interpret fairness findings and design appropriate mitigations. Option E is correct because implementing a fairness metric with defined acceptable thresholds allows the organization to quantitatively measure and monitor performance parity across country groups, ensuring ongoing accountability. Option C is not appropriate because discontinuing the model worldwide is a disproportionate, blanket response that ignores the possibility of remediation and would discard business value. Option D is not appropriate because restricting use to countries showing no bias does not fix the underlying bias, may create inequitable treatment, and fails to address the ethical concern systematically.

Answer

Discontinue use of the AI model worldwide immediately

Answer

Use the model only in countries where it shows no bias

Question 6

Which TWO approaches are recommended for mitigating bias in AI models?

Accepted Answer

Re-weighting training samples. Re-weighting training samples (C) is a recommended bias-mitigation technique because it assigns higher weights to underrepresented or historically disadvantaged groups during training, so the learned loss better reflects balanced group outcomes without discarding data. Adversarial debiasing (E) is also recommended: it trains a predictor alongside an adversary that tries to infer the protected attribute from the predictor's output, penalizing the predictor when the adversary succeeds and thereby reducing disparate impact. By contrast, increasing model depth (A) only raises capacity and can amplify bias, and adding regularization (D) controls overfitting but does not specifically target group fairness. Removing all sensitive attributes (B) is not sufficient because proxy variables (e.g., ZIP code or name) can still encode protected information and reintroduce bias.

Answer

Increasing model depth

Answer

Removing all sensitive attributes

Answer

Adding regularization

Question 7

A developer is implementing retrieval augmented generation (RAG) for a customer service bot. Which component is essential for supplying real-time data to the prompt?

Accepted Answer

A vector database for embedding and retrieval. The correct option is D, a vector database for embedding and retrieval, because RAG requires an external retrieval mechanism that stores document embeddings and performs similarity search to fetch relevant, up-to-date context that is injected into the prompt at inference time. This retrieval step is what supplies real-time or frequently changing data to the model, rather than relying solely on the model's pretrained parameters. A fine-tuned LLM (A) bakes knowledge into weights and cannot dynamically fetch live data, a prompt template in Prompt Builder (B) only structures the input and does not retrieve information, and a static knowledge base (C) is not updated in real time and lacks the embedding-based semantic retrieval needed for RAG.

Answer

A fine-tuned large language model

Answer

A prompt template in Prompt Builder

Answer

A static knowledge base

Question 8

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

Accepted Answer

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

Answer

Sentiment analysis

Answer

Entity extraction

Answer

Intent detection

Question 9

A data analyst wants to create an automatic statistical analysis of sales data that includes waterfall charts and improvement suggestions. Which feature provides these capabilities?

Accepted Answer

Einstein Discovery. Einstein Discovery automatically analyzes datasets, surfaces statistically significant drivers of outcomes, and generates waterfall charts showing how each factor contributes to a prediction, along with plain-language improvement suggestions. It is purpose-built for automated statistical storytelling and what-if analysis. Einstein GPT is a generative text tool, Prediction Builder builds binary prediction scores, and Einstein Analytics (CRM Analytics) is a BI/dashboard platform without the automated narrative and suggestion engine.

Answer

Einstein GPT

Answer

Einstein Prediction Builder

Answer

Einstein Analytics

Question 10

What is the purpose of grounding in the Einstein Trust Layer?

Accepted Answer

To connect the AI to relevant CRM records so responses are accurate and context-aware. Grounding in the Einstein Trust Layer connects the AI model to relevant CRM records (e.g., Accounts, Opportunities, Cases) so that generated responses are based on accurate, up-to-date customer data rather than the model's general training data. This ensures context-aware and factual outputs while maintaining data privacy by not exposing raw records to the model.

Answer

To limit the AI's responses to pre-approved templates

Answer

To encrypt all data sent to the AI model

Answer

To ensure the AI only uses data from Salesforce, not third-party sources

Question 11

Which Einstein feature analyzes call recordings to identify keywords, talk-time metrics, and suggested next steps?

Accepted Answer

Einstein Conversation Insights. Einstein Conversation Insights is the correct feature because it is specifically designed to analyze call recordings and transcripts using natural language processing (NLP) to extract keywords, measure talk-time metrics (e.g., speaker ratio, silence duration), and generate suggested next steps. It integrates with telephony systems to process audio data and provide actionable insights directly within Salesforce, unlike other Einstein features that focus on different data sources or tasks.

Answer

Einstein Discovery

Answer

Einstein Activity Capture

Answer

Einstein Email Insights

Question 12

Which THREE factors should be considered when selecting features for a predictive model in Salesforce?

Accepted Answer

Correlation between features to avoid multicollinearity. Option B is correct because highly correlated features introduce multicollinearity, which destabilizes model coefficients and reduces the reliability of feature-importance scores in Salesforce Einstein Discovery, so correlation between features must be evaluated during selection. Option C is correct because a feature must have a meaningful relationship with the target variable; irrelevant features add noise, inflate dimensionality, and degrade predictive accuracy. Option E is correct because Salesforce predictive models are consumed by business users, and features that are interpretable and actionable make the resulting insights and recommendations easier to trust and operationalize. Option A is not among the marked answers because data volume alone does not determine whether a feature should be selected; a small but highly relevant feature can outperform a large but noisy one. Option D is not among the marked answers because privacy compliance is a governance and data-handling requirement that applies to the overall project, not a feature-selection criterion for model quality.

Answer

Volume of data available for each feature

Answer

Compliance with data privacy regulations

Question 13

A company wants to build a chatbot for their customer portal that can handle returns and refunds. They want the bot to understand phrases like 'I want to return my order' or 'refund request'. What must they configure in Einstein Bots to recognize these variations?

Accepted Answer

Define an intent named 'Return_Refund' and add training phrases like those examples. Einstein Bots uses Natural Language Understanding (NLU) to map user utterances to intents. To recognize variations like 'I want to return my order' or 'refund request', you must define an intent (e.g., 'Return_Refund') and provide multiple training phrases that represent how users might express that intent. The bot then classifies incoming messages to the correct intent based on these examples, enabling it to handle returns and refunds appropriately.

Answer

Enable the Einstein Bots API for external integrations

Answer

Create a dialogue flow that asks clarifying questions

Answer

Set up bot analytics to monitor user utterances

Question 14

Refer to the exhibit. The error log from an AI recommendation system indicates that it cannot explain a decision. Which ethical concern does this directly raise?

Accepted Answer

Lack of transparency and explainability. Lack of transparency and explainability, because the scenario states that the AI recommendation system cannot explain a decision, which directly describes an explainability failure where the system's reasoning is opaque to users or auditors. Explainability concerns the ability to understand and justify how an AI model arrived at a given output, so an error log showing an inability to explain a decision is a textbook transparency issue. The other options do not fit: bias in recommendations (A) concerns unfair or skewed outputs, safety risks (B) concern potential harm from system behavior, and lack of human accountability (D) concerns unclear responsibility for outcomes, none of which are directly raised by an inability to explain a decision.

Answer

Bias in recommendations

Answer

Safety risks

Answer

Lack of human accountability

Question 15

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