AI Associate AI Fundamentals Practice Test 9 — 15 Questions

Question 1

A company wants to use AI to automatically extract invoice numbers, dates, and totals from scanned invoices. Which AI capability is MOST relevant?

Accepted Answer

Computer vision with entity extraction. Computer vision can process images of documents, and entity extraction (often part of NLP) identifies specific data points. Sentiment analysis is for emotions, intent detection for purpose, and generative AI for content creation.

Answer

Intent detection

Answer

Generative AI to create invoices

Answer

Sentiment analysis

Question 2

A data scientist notices that a churn prediction model has high variance: small changes in training data cause large changes in predictions. Which technique is BEST to address this?

Accepted Answer

Apply regularization techniques. High variance — where small changes in training data cause large swings in predictions — is the classic signature of overfitting, and regularization (L1/L2, dropout, early stopping) is the standard remedy. Regularization penalizes model complexity, constraining the hypothesis space so the model generalizes rather than memorizing noise.

Answer

Reduce the amount of training data

Answer

Add more features to the model

Answer

Increase the number of training epochs

Question 3

Which statement best describes 'inference' in the context of machine learning?

Accepted Answer

The process of using a trained model to make predictions on new data. In machine learning, inference (also called prediction or scoring) is the phase where a previously trained model is applied to new, unseen data to generate outputs such as classifications, probabilities, or regression values. It is distinct from training, which learns parameters from labeled data, and from data preparation or evaluation, which occur at different stages of the ML lifecycle. Inference is what delivers business value — the model's learned patterns are used in production to make decisions.

Answer

The process of training a model on labeled data

Answer

The process of collecting and preparing data

Answer

The process of evaluating a model's accuracy

Question 4

A company wants to generate personalized marketing email content for each customer, including product recommendations and tailored copy. Which AI approach is BEST?

Accepted Answer

Use predictive AI to score customers and generative AI to create email copy based on the scores. The best approach combines predictive AI to score or segment each customer (for example, propensity to buy, churn risk, next-best-product) with generative AI to produce personalized copy and recommendations grounded in those scores. This hybrid pattern gives both the targeting intelligence and the natural-language personalization the scenario demands. Predictive-only or generative-only approaches each miss half the requirement.

Answer

Use an unsupervised learning model to cluster customers and send generic emails per cluster

Answer

Use only predictive AI to send the same email to all customers with top products

Answer

Use only generative AI to create emails without any customer data

Question 5

A company is developing a sentiment analysis model for customer reviews. The team wants to ensure the model is fair and does not exhibit bias. Which TWO actions are MOST effective? (Choose two.)

Accepted Answer

Audit the training data for imbalanced representation across demographic groups. Option B is correct because auditing the training data for imbalanced representation across demographic groups directly detects and quantifies bias before it is learned by the model, allowing the team to apply mitigation techniques such as resampling, reweighting, or collecting additional data. Option D is correct because ensuring the training data includes reviews from a diverse range of customers gives the model representative examples across demographics, which reduces the risk of the model systematically underperforming or producing skewed sentiment predictions for underrepresented groups. Options A, C, and E are not appropriate: a more complex model (A) can amplify bias rather than fix it, using historical data as-is (C) preserves existing imbalances and is the opposite of a fairness action, and removing all negative reviews (E) introduces severe label and selection bias by distorting the true sentiment distribution.

Answer

Use a more complex model to capture nuanced patterns

Answer

Use only historical data as-is without modification

Answer

Remove all negative reviews from the training data to avoid negative bias

Question 6

A sales team uses an AI model to prioritize leads. The model's predictions are not improving despite adding more data. Which THREE factors could explain this? (Choose three.)

Accepted Answer

The new data is noisy or of low quality. Option A is correct because adding noisy or low-quality data introduces label errors and irrelevant variance that degrade generalization, so more data cannot improve predictions and may even hurt them. Option B is correct because an underfitting model lacks the capacity (e.g., too few parameters or too simple a hypothesis class) to capture the underlying relationship, so additional data alone will not raise accuracy. Option D is correct because if the input features have little or no predictive signal for lead conversion, no amount of extra data can teach the model a useful mapping. Option C is not correct here because overfitting would typically show strong training performance with poor validation, and the scenario describes stagnant predictions rather than a train/validation gap. Option E is not correct because merely adding more features is not a guaranteed fix; irrelevant or redundant features can worsen performance, and the question asks about factors explaining the lack of improvement.

Answer

The model is already overfitting to the training data

Answer

The team has not added enough features to the model

Question 7

A company wants to build a customer service chatbot that answers questions about their internal policy documents. The documents are updated monthly, and the team cannot afford to retrain a model each time. Which approach is MOST appropriate?

Accepted Answer

Use Retrieval-Augmented Generation (RAG) with the policy documents indexed in a vector store. RAG (Retrieval-Augmented Generation) allows the LLM to retrieve relevant document sections at inference time, so knowledge stays current without retraining. The other options either require expensive retraining for each update or lack document grounding.

Answer

Use a larger foundation model with a longer context window and paste all documents into each prompt

Answer

Train a custom model from scratch on the policy documents each month

Answer

Fine-tune a base LLM on the policy documents monthly

Question 8

Which type of machine learning is used to predict customer churn based on historical labeled data?

Accepted Answer

Supervised learning. Supervised learning trains on labeled historical data where each example has a known outcome (churned vs. not churned), allowing the model to learn the mapping from features to labels. Customer churn prediction is a classic binary classification problem solved with supervised algorithms like logistic regression, random forests, or gradient boosting.

Answer

Reinforcement learning

Answer

Unsupervised learning

Answer

Self-supervised learning

Question 9

A data scientist trains a lead scoring model that achieves 99% accuracy on training data but only 65% accuracy on a held-out test set. What is the most likely issue?

Accepted Answer

Overfitting due to model complexity or insufficient regularization. A large gap between training accuracy (99%) and test accuracy (65%) is the classic signature of overfitting: the model has memorized training data, including noise, rather than learning generalizable patterns. This typically results from excessive model complexity (too many parameters, deep trees) or insufficient regularization (no dropout, L1/L2, or early stopping). The fix is to reduce capacity or add regularization, not to add complexity.

Answer

Underfitting due to insufficient model complexity

Answer

Label noise in the training data

Answer

Data leakage from the test set into training

Question 10

A sales team wants to prioritize leads that are most likely to convert. They have historical data on lead attributes and conversion outcomes. Which AI technique should be used?

Accepted Answer

Supervised learning to build a lead scoring model. Supervised learning is the correct technique because the problem involves predicting a binary outcome (conversion or not) based on historical labeled data (lead attributes and conversion outcomes). A supervised learning model, such as logistic regression or a decision tree, can be trained on past leads to score new leads by their likelihood to convert. This is a classic classification problem.

Answer

Unsupervised clustering to group leads by similarity

Answer

Natural language processing to analyze lead emails

Answer

Computer vision to analyze lead profile pictures

Question 11

A company uses AI to automatically categorize customer support tickets into 'Billing', 'Technical', or 'General'. The model is trained on thousands of past tickets labeled by agents. What type of AI task is this?

Accepted Answer

Text classification. This is a text classification problem using supervised learning, where each ticket is assigned to a predefined category based on labeled training data.

Answer

Intent detection

Answer

Sentiment analysis

Answer

Named entity recognition

Question 12

A predictive model for opportunity scoring shows high precision but low recall. Which business impact is most likely?

Accepted Answer

The sales team misses some convertible opportunities. High precision means that when the model predicts an opportunity will convert, it is usually right — so the sales team is not wasting time on false positives. Low recall means the model misses many opportunities that would actually convert, so the sales team fails to pursue some winnable deals. The business impact is therefore missed convertible opportunities.

Answer

The model is overfitting to the training data

Answer

The sales team wastes time on many low-quality opportunities

Answer

The model overestimates the likelihood of conversion for most opportunities

Question 13

Which of the following is an ethical concern when using AI to make decisions about customers?

Accepted Answer

Bias in training data may lead to unfair treatment of certain customer groups. Bias in training data is a well-known ethical concern in AI, as models learn patterns from historical data that may contain societal biases. When AI is used to make decisions about customers (e.g., loan approvals, hiring, pricing), biased training data can lead to unfair treatment of certain groups, perpetuating discrimination. This is a core ethical issue because it can cause harm and violate fairness principles. Options A, C, and D are incorrect because they are either false statements or not ethical concerns.

Answer

AI models always improve over time without intervention

Answer

AI models require no human oversight

Answer

Deep learning is too complex for business use

Question 14

A retailer wants to recommend products to customers based on their purchase history and browsing behavior. Which AI approach is most suitable?

Accepted Answer

Supervised learning to predict purchase probability for each product. Recommending products based on purchase history and browsing behavior is a classic supervised learning problem: you train a model on historical user-item interactions labeled with purchase outcomes, then predict purchase probability for each candidate product and rank recommendations. Supervised learning (classification or ranking) directly fits this pattern.

Answer

Reinforcement learning with real-time rewards

Answer

Natural language generation to create product descriptions

Answer

Computer vision to analyze product images

Question 15

A model trained on CRM data predicts customer lifetime value (CLV) with high accuracy, but when deployed, predictions are significantly off for new customer segments. What is the most likely cause?

Accepted Answer

The training data was not representative of the new segments. The model performs well on the original CRM data but fails on new customer segments because those segments were not represented in the training data. This is a classic case of sampling bias or covariate shift, where the training distribution does not match the deployment distribution. Since the model learned patterns specific to the original segments, it cannot generalize to unseen segments with different characteristics. The correct answer directly addresses this distribution mismatch.

Answer

Feature engineering was insufficient for the original segments

Answer

The model is overfitting to the training data

Answer

The model architecture is too simple