AI Associate Practice Exam 59 — 20 Questions

Question 1

A company is developing an AI system to screen job applicants. Which TWO practices are essential for ethical AI in hiring?

Accepted Answer

Auditing the model for bias against protected groups. Option B is correct because auditing the model for bias against protected groups is essential to detect and mitigate disparate impact, ensuring compliance with anti-discrimination laws and ethical fairness standards in hiring. Option E is correct because providing candidates with an explanation of decisions supports transparency and contestability, which are core ethical AI principles and often legally required under regulations like GDPR's right to explanation for automated decisions. Options A, C, and D are not essential ethical practices: using all available data including demographic details can introduce or amplify bias and violate privacy or fairness principles, relying solely on AI for final decisions removes necessary human oversight and accountability, and maximizing processing speed is a performance goal unrelated to ethical fairness or transparency.

Answer

Using all available data including demographic details

Answer

Relying solely on AI for final decisions

Answer

Maximizing processing speed

Question 2

A university uses an AI system to predict first-year student retention. The system uses factors such as high school GPA, SAT scores, and socioeconomic indicators. After two years, administrators notice that the model consistently predicts lower retention probabilities for students from low-income families, even when their academic profiles are strong. The university's mission emphasizes equity and inclusion. The admissions office is considering using the predictions to allocate support resources. The model's accuracy on historical data is 85%. What should the university do to align with ethical AI principles?

Accepted Answer

Retrain the model with a fairness constraint such as equalized odds to reduce income-based disparities.. The scenario describes a model whose predictions are systematically biased against low-income students despite strong academic profiles, so retraining with a fairness constraint such as equalized odds directly targets the income-based disparity while preserving the AI system's utility for allocating support resources. Equalized odds requires similar true positive and false positive rates across groups, which is an appropriate technical remedy when the university's mission emphasizes equity and inclusion. Option A is insufficient because simply removing socioeconomic indicators does not eliminate proxy discrimination, since correlated features like high school GPA or neighborhood data can still encode income status. Option B is overly drastic and discards a system that can be made fairer, while Option D ignores the ethical harm by prioritizing aggregate 85% accuracy over disparate impact on a protected or vulnerable group.

Answer

Remove socioeconomic indicators from the model inputs.

Answer

Abandon the AI system and rely on human advisors for resource allocation.

Answer

Use the predictions as-is because they are accurate for the majority of students.

Question 3

Which TWO are best practices for mitigating bias in AI models when using Salesforce Einstein? (Choose two.)

Accepted Answer

Use diverse and representative training data that reflects the target population.. Option A is correct because using diverse and representative training data that reflects the target population is a foundational best practice for reducing bias in Einstein models — if the data underrepresents certain groups, the model's predictions will systematically disadvantage them. Option E is correct because conducting regular bias audits with Einstein's fairness evaluation tools allows teams to detect disparate impact across protected groups and remediate issues before deployment, supporting ongoing governance. Options B, C, and D are not best practices: excluding demographic features with small sample sizes hides rather than fixes underrepresentation, prioritizing majority-group accuracy sacrifices fairness for aggregate metrics, and convenience sampling produces skewed, non-representative datasets that amplify bias.

Answer

Exclude demographic features that have small sample sizes to avoid statistical noise.

Answer

Prioritize accuracy for the majority group to maximize overall performance.

Answer

Rely on convenience sampling to quickly gather a large dataset.

Question 4

A retail company uses an AI recommendation engine to suggest products to online shoppers. The engine uses past purchase history and browsing behavior. Recently, a customer advocacy group publishes a report showing that the engine recommends higher-priced products to customers in affluent zip codes and lower-priced products to customers in lower-income areas, even when both groups have similar browsing histories. The company's revenue has increased since implementing the engine, and marketing teams are pleased. However, the company wants to maintain a reputation for fairness. Which action should the company take?

Accepted Answer

Implement fairness constraints to ensure similar recommendation distributions across demographic groups.. The correct action is D: implement fairness constraints to ensure similar recommendation distributions across demographic groups. This directly addresses the documented disparity where the engine recommends higher-priced products to affluent zip codes and lower-priced products to lower-income areas despite similar browsing histories, by enforcing comparable recommendation outcomes across groups while preserving the AI system's personalization and revenue benefits. Option A does not fit because showing only the most popular products to everyone eliminates personalization and does not guarantee fair treatment across demographic groups. Option B is overly drastic, as discontinuing the system entirely discards its business value instead of mitigating the bias. Option C is inappropriate because keeping the system solely for revenue ignores the fairness problem and the company's stated reputational goal.

Answer

Show only the most popular products to all customers regardless of browsing history.

Answer

Discontinue the AI recommendation system entirely.

Answer

Keep the current system as it maximizes revenue.

Question 5

A customer service department uses an AI chatbot to handle common inquiries. Recently, customers have reported that the chatbot sometimes responds with offensive or inappropriate language. The company wants to uphold ethical standards. Which approach is the best practice?

Accepted Answer

Implement human review of all chatbot responses that are flagged as potentially offensive.. Implementing human review of flagged chatbot responses combines automated detection with human judgment, which is the most reliable way to catch context-dependent offensive language and uphold ethical standards while still allowing the chatbot to operate at scale. This human-in-the-loop approach addresses the root problem of inappropriate outputs rather than merely masking symptoms. Option A is insufficient because simple word filters miss context, sarcasm, and implicitly offensive phrasing, and can also over-block legitimate content. Option B would eliminate offensive responses but destroys the chatbot's usefulness by removing its ability to handle varied inquiries. Option D is flawed because no pre-trained model from any vendor can guarantee ethical behavior; model outputs still require monitoring and governance.

Answer

Implement a filter to automatically block any offensive words in the chatbot's responses.

Answer

Limit the chatbot to only respond from a fixed set of predefined answers.

Answer

Use a pre-trained language model from a trusted vendor to guarantee ethical behavior.

Question 6

A Salesforce administrator is preparing data for an AI model that predicts customer lifetime value. They notice that the 'AnnualRevenue' field on Account records has many outliers, such as negative values and extremely high numbers. What should the administrator do to address this data quality issue?

Accepted Answer

Remove all records with negative revenue and cap extremely high values at a reasonable threshold.. Negative revenue values are likely errors and should be corrected or removed, while extremely high values can skew model training. Capping high values at a reasonable threshold reduces their influence without discarding data. This cleaning improves the quality of the 'AnnualRevenue' field for predicting customer lifetime value.

Answer

Replace all negative values with zero and leave high values unchanged.

Answer

Leave the outliers as they are, because AI models are robust to extreme values.

Answer

Convert the 'AnnualRevenue' field to a text field to avoid numerical issues.

Question 7

A service agent is on a call with a customer. The agent wants to see relevant knowledge articles automatically suggested based on the case subject and description. Which Einstein feature should be enabled?

Accepted Answer

Einstein Article Recommendations. Einstein Article Recommendations is the only feature that automatically suggests knowledge articles to service agents by analyzing the case subject and description. It helps agents resolve cases faster by providing relevant content. The other features either classify, score, or log activities, but none recommend articles.

Answer

Einstein Case Scoring

Answer

Einstein Activity Capture

Answer

Einstein Case Classification

Question 8

An HR department uses an AI tool to screen resumes for a software engineering position. The tool was trained on resumes of past successful hires, who were predominantly male. The tool has been in use for three months, during which only 10% of candidates shortlisted for interviews are female, even though 40% of applicants are female. The hiring managers are satisfied with the quality of candidates shortlisted, as most perform well in interviews. However, the company's diversity and inclusion officer raises an ethical concern. What should the company do to address this bias?

Accepted Answer

Retrain the model with a balanced dataset that includes more female candidates and remove gender-correlated features.. The bias stems from training data reflecting predominantly male past hires, so retraining with a balanced dataset and removing gender-correlated features directly addresses the root cause of the disparate shortlisting rate (10% female vs. 40% female applicants). This approach aligns with standard fairness remediation in AI hiring: mitigate biased training data and proxy variables rather than only patching outcomes. Option B does not fix the biased model and merely adds an appeal process, which still disadvantages female candidates at scale. Option C ignores the ethical and legal disparate-impact concern because interview performance does not justify a biased screening stage. Option D may reduce bias but is a blunt, non-scalable workaround that discards the AI tool rather than correcting it.

Answer

Use the AI tool for initial screening but allow candidates to appeal the decision.

Answer

Continue using the current tool since it selects high-quality candidates.

Answer

Manually review all resumes without using the AI tool.

Question 9

A global e-commerce company deploys Einstein Bots in multiple countries. The bot uses natural language processing to handle customer returns. In one region, customers frequently complain that the bot does not understand their local dialect and incorrectly rejects valid returns. The company wants to maintain consistent customer experience while respecting regional diversity. The bot's language model was trained mainly on English data from the US and UK. The AI ethics board is concerned about fairness and transparency. They consider four options: (A) use a single, centrally-trained model with fallback to human agents for non-English queries, (B) deploy separate models fine-tuned on each dialect but with centralized monitoring, (C) disable the bot in regions with dialect issues, (D) use a translation layer to convert all inputs to English before processing. What is the best ethical approach?

Accepted Answer

Deploy separate models fine-tuned on each dialect with centralized monitoring. Deploying separate models fine-tuned on each regional dialect directly addresses the root cause of the misclassification—the US/UK English training data—while centralized monitoring preserves consistent oversight, fairness auditing, and transparency across regions. Fine-tuning on local dialect data improves the bot's ability to correctly recognize valid returns, respecting regional diversity without sacrificing a consistent customer experience. Option A does not fix the model's dialect gap and merely routes non-English queries to humans, which is costly and inconsistent. Option B's translation layer loses dialect-specific nuance and can introduce errors, while Option C removes service entirely and harms customers in those regions.

Answer

Use a single centrally-trained model with fallback to human agents for non-English queries

Answer

Use a translation layer to convert all inputs to English before processing

Answer

Disable the bot in regions with dialect issues

Question 10

A sales team is using Einstein Lead Scoring, but the scores for new leads seem inconsistent and not reflecting recent conversion patterns. The admin checks the model and finds it was trained three months ago. Which action should the admin take to improve model accuracy?

Accepted Answer

Retrain the Einstein Lead Scoring model with the latest lead data.. The correct action is A: retrain the Einstein Lead Scoring model with the latest lead data. Einstein Lead Scoring relies on a machine learning model that must be periodically retrained so it reflects current conversion patterns; a model trained three months ago will produce stale, inconsistent scores as lead behavior and win rates change. Retraining uses recent lead and opportunity outcomes to recalibrate the scoring algorithm and restore accuracy. Option B is wrong because manually overriding scores does not improve the model and can corrupt the training signal. Option C is wrong because field history retention affects audit/history tracking, not model training data. Option D is wrong because field-level security controls user access, not the fields available to the scoring model.

Answer

Manually override the lead scores for a sample of leads.

Answer

Increase the field history retention period for lead fields.

Answer

Adjust field-level security to allow the model to access more fields.

Question 11

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 12

A data engineer is preparing a dataset for an Einstein Discovery model that predicts customer churn. The dataset contains missing values in several fields. Which two methods are appropriate for handling missing values to maintain model accuracy? (Choose two.)

Accepted Answer

Impute missing numerical values with the mean of the column.. Imputing missing numerical values with the mean and missing categorical values with the mode are both standard, effective methods to handle missing data without discarding records. They preserve dataset size and maintain distributions, which supports model accuracy. Other methods like row deletion or arbitrary zero-filling can introduce bias or errors.

Answer

Remove all rows with any missing values.

Answer

Set all missing values to zero.

Answer

Ignore missing values and proceed with model training.

Question 13

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 14

A Salesforce administrator is integrating external data into Data Cloud to enhance an AI model. The external data contains customer addresses with inconsistent formatting (e.g., 'St.' vs 'Street', 'Ave' vs 'Avenue'). What is the most effective approach to prepare this data for the model?

Accepted Answer

Use Data Cloud's data transformation capabilities to standardize address formats before ingesting into the model.. Inconsistent address formatting introduces noise that can mislead the model if address-derived features are used. Standardizing formats during data transformation in Data Cloud ensures that similar addresses are recognized as such, improving the quality of features. This preprocessing step is essential for accurate predictions and is a best practice in data preparation for AI.

Answer

Exclude address fields from the model to avoid the complexity of standardization.

Answer

Ingest the data as-is and rely on the AI model to automatically handle inconsistencies.

Answer

Create a custom formula field in Salesforce to concatenate all address components into a single string.

Question 15

A data scientist notices that a Salesforce Einstein model's performance degrades over time. The model was trained on data from the last year. What is the most likely cause?

Accepted Answer

Concept drift. Concept drift (A) is the correct answer because the statistical properties of the target variable or its relationship with input features change over time, causing a model trained on last year's data to become less accurate as real-world patterns evolve. This is a classic scenario in production ML where temporal data shifts degrade model performance even though the model was valid at training time. Underfitting (B) refers to a model too simple to capture patterns in the training data, which would show poor performance from the start rather than degrading over time. Data leakage (C) is a training-time flaw where future or target information contaminates features, typically inflating validation metrics rather than causing gradual production decay. Overfitting (D) means the model memorized training data and fails to generalize, but that would also manifest immediately on new data, not as a slow degradation tied to temporal change.

Answer

Underfitting

Answer

Data leakage

Answer

Overfitting

Question 16

A financial institution uses an AI model to approve credit. The model shows disparate impact against a protected group. Under Salesforce's ethical AI principles, what is the most appropriate action?

Accepted Answer

Apply a bias mitigation technique such as reweighing.. Salesforce's ethical AI principles emphasize responsible use and mitigating harmful bias, so applying a bias mitigation technique such as reweighing directly addresses the disparate impact while preserving the model's utility. Reweighing adjusts sample weights during training to reduce bias against the protected group, which is a targeted, proportionate remedy for a model that is otherwise functional. Discontinuing the model immediately (A) is an overreaction that discards a viable system instead of remediating it, and increasing the decision threshold for all applicants (B) does not correct group-specific bias and may worsen outcomes. Publishing predictions publicly (C) exposes sensitive financial data and does not itself mitigate the bias, so it is not the appropriate action.

Answer

Discontinue the model immediately.

Answer

Increase the model's decision threshold for all applicants.

Answer

Publish the model's predictions publicly for scrutiny.

Question 17

A company uses an AI chatbot that automatically responds to customer service inquiries. When a customer questions the bot's response, there is no mechanism for the customer to appeal or speak to a human. What ethical principle is being violated?

Accepted Answer

Accountability. Accountability, because the scenario describes a system where no mechanism exists for a customer to appeal an AI decision or escalate to a human, meaning no party is answerable for the chatbot's responses. Accountability in AI ethics requires that there be identifiable responsibility and recourse channels for automated decisions that affect users. Transparency (B) concerns disclosing how the system works and that AI is being used, which is not the issue here. Privacy (A) relates to handling of personal data, and Fairness (D) concerns bias and equitable treatment, neither of which is described in this scenario.

Answer

Privacy

Answer

Transparency

Answer

Fairness

Question 18

A bank uses Einstein to approve loan applications. The model is trained on data that includes zip codes. Analysis shows that applicants from low-income zip codes are disproportionately rejected, even when their credit profiles are similar. What is the most likely ethical issue?

Accepted Answer

Use of a proxy for protected attributes. Use of a proxy for protected attributes, because zip codes correlate strongly with race and socioeconomic status, so using them as a model feature causes disparate impact even though the protected attribute itself is not explicitly included. This is a classic redlining-style bias, where the model learns to reject applicants from low-income zip codes despite similar credit profiles. Option A does not fit because using zip codes is not inherently a privacy violation in this scenario. Option C does not fit because the issue is biased outcomes, not explainability. Option D does not fit because the model is performing as trained; the problem is fairness, not accuracy.

Answer

Privacy violation due to zip code use

Answer

Lack of transparency in decision-making

Answer

Unreliable model performance

Question 19

A company is building an AI model to score sales leads. They have a dataset with historical leads, including whether they converted. The dataset contains 90% male and 10% female leads. The model will be used to prioritize leads for sales follow-ups. What is the primary ethical concern?

Accepted Answer

The training data is imbalanced, which may cause the model to be less accurate for female leads, leading to unfair prioritization.. The correct answer is A: the training data is imbalanced, with 90% male and 10% female leads, which can cause the model to learn patterns that favor the majority class and perform less accurately for female leads, resulting in unfair prioritization in sales follow-ups. This is a classic class-imbalance and fairness concern in machine learning, where underrepresentation of a group can lead to disparate model performance and discriminatory outcomes. Option B is incorrect because overrepresentation of male leads typically biases the model in favor of, not against, the majority group. Option C is not necessarily true, as dataset size alone is not the primary ethical issue here, and no size threshold is given. Option D is too broad and not the primary concern, since using historical data is standard practice and only becomes unethical if it encodes bias or outdated discrimination.

Answer

The model will be biased against male leads because they are overrepresented.

Answer

The dataset is too small to build a reliable model.

Answer

Using historical data is unethical because it may not reflect current conditions.

Question 20

A data engineer is preparing a dataset for an Einstein Discovery model that predicts customer churn. The dataset includes fields such as 'LastPurchaseDate', 'TotalSpend', 'CustomerServiceCalls', and 'Age'. Which two data preparation steps are essential to ensure the model performs well? (Choose two.)

Accepted Answer

Normalize 'TotalSpend' and 'CustomerServiceCalls' to a similar scale to prevent one from dominating the model.. Converting date fields to numeric durations and scaling numeric features are fundamental data preparation steps for predictive modeling. The date transformation captures recency, while scaling prevents features with larger ranges from dominating the model. These steps ensure that the model can learn from all features fairly and effectively, leading to better churn predictions.

Answer

Remove any records with missing values to avoid errors during model training.

Answer

Encode 'Age' into categorical bins (e.g., 18-25, 26-35) to reduce the impact of outliers.

Answer

Create a new feature that multiplies 'TotalSpend' by 'CustomerServiceCalls' to capture interaction effects.