I85PREP
⌘ K
AI Business Professional study materials
AI Business Professional study resources

AI Business Professional practice questions

One question at a time. Build your AI Business Professional understanding.

First 50 of 200 questions. Each one lists the choices, the correct answer, and an explanation. The checker at the top of this page saves progress in this browser when it is available.

Continue in the practice tool above to load more questions, 50 at a time.

Question 1AI Foundations and Concepts

In machine learning, which term describes the data used to train a model and includes the correct answers the model should learn to predict?

  1. Labeled data Correct answer.
  2. Unstructured data
  3. Synthetic data
  4. Streaming data

Explanation

Supervised learning relies on labeled data, where each example is paired with the correct output (the label). The model learns a mapping from inputs to those known answers.

Question 2AI Foundations and Concepts

A business analyst wants an AI system to sort incoming customer emails into 'billing', 'technical', and 'general' folders based on examples of previously sorted emails. Which type of machine learning is most appropriate?

  1. Reinforcement learning
  2. Rule-based expert system
  3. Supervised learning Correct answer.
  4. Unsupervised learning

Explanation

The task is classification with known category labels from historical examples, which is the defining setup for supervised learning. The model learns the mapping from email content to category.

Question 3AI Foundations and Concepts

Which statement best describes a large language model (LLM)?

  1. A database that stores every document it was trained on for exact retrieval
  2. A rules engine that follows hand-written grammar instructions
  3. A model trained on large text corpora to predict and generate language Correct answer.
  4. A spreadsheet tool that calculates statistics on text

Explanation

LLMs are neural networks trained on large volumes of text, typically to predict the next token, which lets them generate and process language. They do not store verbatim copies of training documents for retrieval.

Question 4AI Foundations and Concepts

In the context of generative AI, what does the term 'hallucination' refer to?

  1. A model producing fluent but factually incorrect content Correct answer.
  2. A model being trained on too little data
  3. A model refusing to answer a question
  4. A model running slower than expected

Explanation

A hallucination is output that is stated confidently but is fabricated or incorrect. It arises because the model generates plausible text rather than verifying facts.

Question 5AI Foundations and Concepts

Which of the following is an example of unstructured data that AI systems commonly process?

  1. A table of monthly sales figures
  2. A collection of recorded customer service calls Correct answer.
  3. A list of product prices in a spreadsheet
  4. A set of numeric survey responses

Explanation

Audio recordings are unstructured because they lack a predefined tabular schema and require models to extract meaning. Tables of numbers are structured data.

Question 6AI Foundations and Concepts

What is the primary purpose of a prompt in an interaction with a generative AI model?

  1. To encrypt the model's responses
  2. To retrain the model on new data
  3. To provide the instruction or input that guides the model's output Correct answer.
  4. To reduce the model's electricity consumption

Explanation

A prompt is the input text that instructs the model and shapes what it generates. Clear prompts improve the relevance and quality of the response.

Question 7AI Foundations and Concepts

Which of these tasks is a classic example of a natural language processing (NLP) application?

  1. Compressing a video file
  2. Classifying customer reviews as positive or negative Correct answer.
  3. Detecting objects in a photograph
  4. Forecasting next quarter's inventory

Explanation

Sentiment classification of text is a core NLP task because it involves interpreting human language. The other options involve vision, numeric forecasting, or file compression.

Question 8AI Foundations and Concepts

Which of the following best describes the role of a 'token' in a large language model?

  1. A user account that grants access to the model
  2. A hardware component that stores the model
  3. A security certificate that authenticates requests
  4. A unit of text, such as a word or word fragment, that the model processes Correct answer.

Explanation

Models break text into tokens, which are words or sub-word pieces, and process sequences of these units. Token counts also drive context limits and pricing.

Question 9AI Foundations and Concepts

A company wants to group customers into segments based on purchasing behavior without any predefined segment labels. Which approach fits best?

  1. Reinforcement learning
  2. Manual rule writing
  3. Supervised regression
  4. Unsupervised clustering Correct answer.

Explanation

Clustering is an unsupervised technique that discovers natural groupings in data when no labels exist. It groups similar records together based on their features.

Question 10AI Foundations and Concepts

Which statement accurately distinguishes training from inference in a machine learning lifecycle?

  1. Inference adjusts model parameters; training applies the model to new inputs
  2. Training adjusts model parameters; inference applies the trained model to new inputs Correct answer.
  3. Training only applies to unsupervised models; inference only applies to supervised models
  4. Training and inference are the same process performed at different times

Explanation

Training is the phase where the model learns parameters from data, while inference is using the trained model to make predictions on new inputs. They are distinct phases with different computational goals.

Question 11AI Foundations and Concepts

A model performs very well on its training data but poorly on new data. What is this situation called?

  1. Data leakage
  2. Normalization
  3. Underfitting
  4. Overfitting Correct answer.

Explanation

Overfitting occurs when a model memorizes training data, including noise, and fails to generalize to unseen examples. The gap between training and validation performance is the signal.

Question 12AI Foundations and Concepts

Which practice best helps a team evaluate how well a model will generalize before deployment?

  1. Removing all preprocessing steps
  2. Testing only on the training dataset
  3. Increasing the learning rate until loss reaches zero
  4. Holding out a validation or test set not used in training Correct answer.

Explanation

Evaluating on data the model never saw during training gives an unbiased estimate of generalization. A held-out set simulates how the model will behave on new inputs.

Question 13AI Foundations and Concepts

What does the 'context window' of a large language model refer to?

  1. The physical memory of the server hosting the model
  2. The maximum amount of text, measured in tokens, the model can consider at once Correct answer.
  3. The number of users who can access the model simultaneously
  4. The time it takes for the model to respond

Explanation

The context window is the limit on how many tokens of input and output the model can handle in a single interaction. Content beyond that limit cannot be attended to.

Question 14AI Foundations and Concepts

A team wants the model to answer questions using an internal knowledge base that changes frequently, without retraining the model. Which approach is most appropriate?

  1. Switching to an unsupervised clustering model
  2. Fine-tuning the model on the entire knowledge base each day
  3. Reducing the model's context window
  4. Retrieval-augmented generation that supplies relevant documents at query time Correct answer.

Explanation

Retrieval-augmented generation fetches relevant documents at query time and includes them in the prompt, so updated knowledge can be used without retraining. This is efficient when the knowledge base changes often.

Question 15AI Foundations and Concepts

Which statement best explains the concept of 'temperature' when sampling from a language model?

  1. It controls the randomness of token selection, with lower values producing more deterministic output Correct answer.
  2. It measures the model's confidence in factual accuracy
  3. It sets the maximum number of tokens in the output
  4. It determines the model's training data cutoff date

Explanation

Temperature scales the probability distribution used during sampling. Lower temperatures make high-probability tokens more likely, yielding more consistent output, while higher values increase variety.

Question 16AI Foundations and Concepts

A model is described as 'multimodal'. What does this mean?

  1. It produces multiple answers to every prompt
  2. It can run on multiple servers at once
  3. It has been trained on multiple languages only
  4. It can process more than one type of input, such as text and images Correct answer.

Explanation

Multimodal models handle several data modalities, for example combining text, images, and audio. This lets them perform tasks that require understanding across input types.

Question 17AI Foundations and Concepts

Which of the following is the best example of a classification task rather than a regression task?

  1. Predicting a house's sale price in dollars
  2. Predicting tomorrow's temperature in degrees
  3. Forecasting monthly revenue
  4. Predicting whether a loan applicant will default Correct answer.

Explanation

Classification predicts a discrete category, such as default versus no default. Regression predicts a continuous numeric value.

Question 18AI Foundations and Concepts

What is the main purpose of a system prompt when building an application on top of a language model?

  1. To store the user's payment details securely
  2. To set persistent instructions and behavior for the assistant across a conversation Correct answer.
  3. To measure the model's processing speed
  4. To compile the model's source code

Explanation

A system prompt provides high-level, persistent guidance such as role, tone, and constraints that shape all responses in the session. It is separate from individual user turns.

Question 19AI Foundations and Concepts

Which statement best captures the difference between AI, machine learning, and deep learning?

  1. AI is a subset of deep learning, which is a subset of machine learning
  2. Machine learning is a subset of deep learning, which is a subset of AI
  3. Deep learning is a subset of machine learning, which is a subset of AI Correct answer.
  4. They are three names for exactly the same technology

Explanation

AI is the broad field of intelligent systems; machine learning is a subset that learns from data; deep learning is a further subset using multi-layer neural networks. The relationship is nested from broad to narrow.

Question 20AI Foundations and Concepts

A company is evaluating an AI vendor's claims about a model's accuracy. Which evidence most directly supports the claim?

  1. Results on a held-out benchmark that matches the company's use case Correct answer.
  2. The number of parameters listed on the vendor's website
  3. The size of the vendor's engineering team
  4. A marketing brochure describing the model as state of the art

Explanation

Performance on a relevant, held-out benchmark provides direct evidence of accuracy for the intended task. Marketing claims and model size are weaker proxies.

Question 21AI Foundations and Concepts

A model trained to approve loans shows much lower approval rates for one demographic group even though that group is not a direct input. What is the most likely underlying cause?

  1. The temperature setting is too low
  2. The model's context window is too small
  3. The model uses too few tokens per request
  4. Proxy variables in the training data correlate with the excluded group Correct answer.

Explanation

When a protected attribute is removed, other features can act as proxies that encode the same information, producing disparate outcomes. This is a well-known fairness risk in supervised models.

Question 22AI Foundations and Concepts

An LLM consistently cites plausible-sounding but nonexistent sources when asked for references. Which mitigation most directly addresses this?

  1. Shortening the prompt
  2. Requiring the model to ground answers in retrieved documents and cite only those Correct answer.
  3. Switching to a smaller model
  4. Increasing the temperature

Explanation

Grounding responses in retrieved source documents constrains the model to verifiable material and reduces fabricated citations. This is a core motivation for retrieval-augmented approaches.

Question 23AI Foundations and Concepts

A team notices that adding a feature available only at prediction time (but derived from the outcome) boosts validation accuracy dramatically. What is the most likely explanation?

  1. The dataset is too large
  2. The model is underfitting
  3. Target leakage has inflated the measured performance Correct answer.
  4. The learning rate is too high

Explanation

Using information that encodes the outcome constitutes target leakage, which produces unrealistically high validation scores that will not hold in production. The feature must be removed or the evaluation redesigned.

Question 24AI Foundations and Concepts

Which scenario best illustrates the difference between a model's capability and its reliability?

  1. A model has a large context window and many parameters
  2. A model is hosted on cloud infrastructure
  3. A model can summarize long reports but sometimes omits critical figures Correct answer.
  4. A model was trained on a large corpus of text

Explanation

Capability is what the model can do; reliability is how consistently it does so correctly. A model that can summarize but sometimes drops critical details shows high capability with imperfect reliability.

Question 25AI Foundations and Concepts

A product team wants to reduce hallucinated answers in a customer-facing assistant without retraining the base model. Which combination is most effective?

  1. Increasing the context window without changing instructions
  2. Higher temperature with longer outputs
  3. Lower temperature with retrieval grounding and a verification step Correct answer.
  4. Removing the system prompt and allowing free-form answers

Explanation

Lower temperature reduces randomness, retrieval grounding supplies verifiable facts, and a verification step catches errors before display. Together these directly target fabrication without retraining.

Question 26AI Foundations and Concepts

Which statement best describes supervised learning?

  1. The model learns by receiving rewards after taking actions in an environment.
  2. The model is trained only on unlabeled text to predict the next word.
  3. The model learns from labeled input-output pairs that provide a target signal. Correct answer.
  4. The model discovers clusters in data without any target values.

Explanation

Supervised learning relies on a dataset where each input is paired with a known correct output, and the model is optimized to reduce the gap between its predictions and those labels. The other approaches describe unsupervised, reinforcement, and self-supervised learning respectively.

Question 27AI Foundations and Concepts

In a neural network, what is the primary function of an activation function?

  1. To normalize the size of the input dataset
  2. To introduce non-linearity so the network can model complex relationships Correct answer.
  3. To count the number of layers in the model
  4. To store the training data between layers

Explanation

Activation functions transform the weighted sum of inputs at each neuron, and their non-linear nature is what allows stacked layers to approximate complex, non-linear functions. Without them, a deep network would collapse into an effectively linear model.

Question 28AI Foundations and Concepts

Which description matches a generative AI system rather than a purely predictive one?

  1. It ranks existing documents by relevance to a query.
  2. It flags transactions that exceed a fixed threshold.
  3. It produces new text, images, or audio content in response to an input. Correct answer.
  4. It assigns a spam score to an incoming email.

Explanation

Generative systems create new content such as text, images, or audio, whereas purely predictive systems output a label, score, or ranking for existing data. The distinction is about producing new artifacts versus scoring inputs.

Question 29AI Foundations and Concepts

What does 'grounding' mean when applied to a large language model application?

  1. Tying model outputs to retrieved or provided source material so responses are verifiable Correct answer.
  2. Compressing the model so it runs on smaller hardware
  3. Fine-tuning the model on a very large general corpus
  4. Restricting the model to answer only in a single language

Explanation

Grounding connects a model's response to specific source content, such as retrieved documents or supplied passages, so that claims can be traced and verified. It reduces unsupported generation by anchoring answers in evidence the application controls.

Question 30AI Foundations and Concepts

A retailer wants to forecast next month's revenue in dollars from historical sales data. Which model type fits this task?

  1. A classification model that outputs a category label
  2. A clustering model that groups similar months
  3. A ranking model that orders products by popularity
  4. A regression model that predicts a continuous value Correct answer.

Explanation

Revenue is a continuous numeric quantity, so a regression model is appropriate because it estimates numeric outputs rather than discrete categories. Classification would only make sense if the target were a label such as high or low revenue.

Question 31AI Foundations and Concepts

A team needs a model to summarize support tickets, but they have only a few hundred labeled examples. Which approach is most practical?

  1. Collect no additional data and rely on random guessing
  2. Train a large language model from scratch on the small dataset
  3. Use a pretrained foundation model and adapt it with prompting or light fine-tuning Correct answer.
  4. Discard the labeled examples and use only rule-based keyword matching

Explanation

Pretrained foundation models already encode broad language ability, so a small labeled set is usually enough to adapt them through prompting, few-shot examples, or parameter-efficient fine-tuning. Training from scratch requires far more data and compute than a few hundred examples provide.

Question 32AI Foundations and Concepts

Which example best illustrates a recommendation system using collaborative filtering?

  1. Suggesting items whose descriptions share keywords with a viewed item
  2. Suggesting items that users with similar purchase histories also bought Correct answer.
  3. Suggesting items that are newest in the catalog
  4. Suggesting items sorted by their current price

Explanation

Collaborative filtering bases recommendations on patterns of interaction across many users, such as similarity in purchase or rating behavior. Content-based and simple sorting methods rely on item attributes or metadata rather than cross-user patterns.

Question 33AI Foundations and Concepts

What is the purpose of splitting data into training, validation, and test sets?

  1. To tune hyperparameters on validation data while estimating final performance on unseen test data Correct answer.
  2. To increase the total amount of data available to the model
  3. To reduce the storage cost of the dataset
  4. To ensure every data point is used equally often in every phase

Explanation

The validation set supports model selection and hyperparameter tuning, while the test set provides an unbiased estimate of performance on data the model has never seen. Keeping them separate prevents optimistic bias from tuning on the same data used to report results.

Question 34AI Foundations and Concepts

A company deploys a model that flags fraudulent transactions. Which metric matters most when missing a fraud case is far costlier than a false alarm?

  1. Overall accuracy across all transactions
  2. The number of features used by the model
  3. The average training time of the model
  4. Recall for the fraud class Correct answer.

Explanation

Recall for the fraud class measures how many actual fraud cases the model catches, which directly reflects the cost of missed detections. Overall accuracy can look high while the model misses most fraud because fraud is rare.

Question 35AI Foundations and Concepts

Which situation best describes a false positive in a content moderation model?

  1. A harmful post is allowed to remain online
  2. A harmless post is correctly allowed to remain online
  3. A harmful post is correctly flagged for review
  4. A harmless post is incorrectly flagged for review Correct answer.

Explanation

A false positive occurs when the model predicts the positive class, here 'harmful', for an item that is actually negative. Flagging a harmless post is exactly that error, while missing a harmful post is a false negative.

Question 36AI Foundations and Concepts

Why do many organizations prefer retrieval-augmented generation over relying solely on a model's internal knowledge for factual queries?

  1. It eliminates all inference costs
  2. It removes the need to evaluate answer quality
  3. It lets answers draw on current, citable documents without retraining the model Correct answer.
  4. It guarantees the model never produces an incorrect statement

Explanation

Retrieval-augmented generation supplies relevant documents at query time, so answers can reflect up-to-date information and cite sources without changing model weights. It reduces but does not eliminate errors, and it adds retrieval and evaluation work rather than removing it.

Question 37AI Foundations and Concepts

A bank's model approves loans accurately overall but denies a qualified applicant whose profile resembles many past denials. Which concept does this illustrate?

  1. The model has learned bias present in historical decisions Correct answer.
  2. The model has too few parameters to function
  3. The model has memorized the training data
  4. The model was trained with an incorrect learning rate

Explanation

When training labels reflect past human decisions that disadvantaged certain profiles, a model can reproduce and amplify those patterns. This is a bias-in-data problem rather than a capacity or optimization issue.

Question 38AI Foundations and Concepts

What distinguishes an agentic AI workflow from a single prompted model call?

  1. It uses a larger context window for every request
  2. It requires no evaluation because it is autonomous
  3. It plans and takes multiple steps, using tools and intermediate results to reach a goal Correct answer.
  4. It always produces shorter responses

Explanation

Agentic workflows decompose a goal into steps, invoke tools or retrieve information, and use intermediate results to decide the next action. A single prompted call produces one response without this iterative planning loop.

Question 39AI Foundations and Concepts

A team fine-tunes a model on its own customer data and later finds it reproduces fragments of that data verbatim. Which risk does this demonstrate?

  1. Underfitting caused by insufficient training epochs
  2. A context window that is too small
  3. A tokenization error in the input pipeline
  4. Data leakage from overfitting to sensitive training examples Correct answer.

Explanation

Models trained on sensitive records can memorize and later emit those records, which is a privacy risk known as training data leakage. Mitigations include deduplication, differential privacy, and output filtering, rather than simply adjusting context size.

Question 40AI Foundations and Concepts

An organization wants to compare two language models on a customer-facing task where many answers are valid but differ in tone and completeness. Which evaluation approach is most appropriate?

  1. Combining rubric-based human review with automated checks for factual support Correct answer.
  2. Comparing the models' parameter counts
  3. Using a single exact-match score against one reference answer
  4. Measuring only the average response latency

Explanation

When many valid outputs exist, exact-match scoring unfairly penalizes acceptable variation, so rubric-guided human judgments paired with automated factuality checks capture quality more faithfully. Parameter counts and latency describe cost and speed, not answer quality.

Question 41AI in Business Operations and Workflows

In a business process, which term describes a defined sequence of tasks that transforms an input into a deliverable of value to a customer?

  1. A cost center
  2. A data warehouse
  3. A workflow Correct answer.
  4. A service level agreement

Explanation

A workflow is the structured sequence of tasks, roles, and handoffs that turns inputs into an output the customer values. The other terms describe data storage, contractual service targets, or accounting units, not task sequences.

Question 42AI in Business Operations and Workflows

A team wants to remove manual data re-entry between its CRM and its billing system. Which approach most directly addresses this goal?

  1. Scheduling more status meetings
  2. Increasing the size of the billing team
  3. Integrating the two systems through an API Correct answer.
  4. Adding more approval steps to the billing process

Explanation

System integration lets data flow automatically between applications, eliminating duplicate manual entry. Adding approvals, meetings, or staff does not remove the underlying re-entry work.

Question 43AI in Business Operations and Workflows

Which metric best measures how much of a process's total elapsed time is spent on value-adding work?

  1. Employee utilization rate
  2. Defect density
  3. Process cycle efficiency Correct answer.
  4. Customer acquisition cost

Explanation

Process cycle efficiency is the ratio of value-added time to total lead time, so it directly expresses how much elapsed time adds value. Utilization, acquisition cost, and defect density measure other dimensions of performance.

Question 44AI in Business Operations and Workflows

What is the primary purpose of a business process model?

  1. To document the org chart of a department
  2. To rank employees by productivity
  3. To create a shared, explicit representation of how work flows Correct answer.
  4. To forecast quarterly revenue

Explanation

Process models make the steps, roles, decisions, and handoffs explicit so stakeholders share one view of how work actually flows. They are not org charts, financial forecasts, or performance rankings.

Question 45AI in Business Operations and Workflows

An AI system drafts routine customer email replies and routes any reply expressing dissatisfaction to a human agent. What is this pattern called?

  1. Human-in-the-loop Correct answer.
  2. Batch processing
  3. Data replication
  4. Full automation

Explanation

Human-in-the-loop design keeps a person as a decision point for cases the system should not handle alone. Full automation would send all replies without review, while batch processing and data replication describe unrelated technical patterns.

Question 46AI in Business Operations and Workflows

Which of the following is a leading indicator rather than a lagging indicator of operational performance?

  1. Total revenue last year
  2. Number of customer complaints last quarter
  3. Annual employee turnover
  4. Backlog of unprocessed tickets today Correct answer.

Explanation

Leading indicators reflect current conditions that predict future outcomes, such as today's backlog. Complaints, revenue, and turnover are lagging results recorded after the fact.

Question 47AI in Business Operations and Workflows

What does robotic process automation (RPA) typically do?

  1. Replaces a company's core database
  2. Designs new products from customer feedback
  3. Mimics user actions to execute repetitive rule-based tasks Correct answer.
  4. Trains large language models on company documents

Explanation

RPA software bots replicate keystrokes and screen interactions to perform repetitive, rule-based work across existing applications. It does not replace databases, train language models, or perform creative design.

Question 48AI in Business Operations and Workflows

A manager wants a single agreed number that defines acceptable turnaround time for support tickets. What should the manager establish?

  1. A service level agreement Correct answer.
  2. A key risk indicator
  3. A data dictionary
  4. A change log

Explanation

A service level agreement formally defines the performance target, such as maximum turnaround time, between a provider and its customers. Risk indicators, data dictionaries, and change logs serve different governance purposes.

Question 49AI in Business Operations and Workflows

A finance team automates invoice matching so that invoices within tolerance are approved automatically and exceptions are flagged. Which benefit does this design primarily deliver?

  1. Guaranteed elimination of all duplicate payments
  2. Complete removal of the accounts payable team
  3. Faster handling of routine cases with human focus on exceptions Correct answer.
  4. Elimination of the need for any financial controls

Explanation

Exception-based automation speeds up routine matching while directing human attention to the cases that need judgment. It does not remove the need for controls, guarantee zero errors, or eliminate the function entirely.

Question 50AI in Business Operations and Workflows

A process owner maps each step of an order fulfillment workflow and finds that three approvals add no control value. What is the most appropriate next action?

  1. Remove the non-value-adding approvals before automating Correct answer.
  2. Add an AI model to speed up the approvals
  3. Escalate the finding without changing the process
  4. Automate all three approvals exactly as they are

Explanation

Process simplification should precede automation so that technology does not lock in wasted steps. Automating unnecessary approvals or merely speeding them up preserves the waste, and escalation alone leaves the process unchanged.