50 IBM A1000-077 - Assessment: Foundations of AI Practice Questions: Question Bank 2025

A retail team wants to describe the difference between AI, machine learning (ML), and deep learning (DL) to non-technical stakeholders. Which statement is most accurate?

A hospital wants an AI system to categorize incoming emails into "Billing", "Appointments", and "Medical Records" based on the email text. What type of ML problem is this?

A product team wants to quickly add a conversational interface to answer common questions like store hours and return policy, without building an NLP model from scratch. Which IBM capability best fits this need?

A bank is reviewing an AI-based loan approval tool. Which action best supports responsible AI and governance?

A team built a model with 98% accuracy to detect fraudulent transactions, but fraud is rare (about 1% of transactions). Which metric is typically more informative for this situation?

A team is preparing data for a customer churn model. They notice some customers have missing values for "MonthlyCharges". What is a recommended best practice before training?

A company wants employees to ask natural-language questions like "What were last quarter’s top-selling products?" and get answers grounded in internal documents and reports. Which combination is most appropriate?

A team is deploying an AI feature that recommends job candidates for recruiters. Which practice best helps mitigate bias and supports fairness?

A data scientist reports excellent evaluation results, but the model performs poorly in production. Investigation shows the training pipeline calculated "DaysSinceLastPurchase" using the full dataset, including records from after the prediction date. What is the primary issue?

A customer-support chatbot must answer with policies that may change and must cite the source text used. The team wants to reduce hallucinations by grounding responses in an approved knowledge base. Which architecture pattern best meets this requirement?

An HR team wants an AI solution to automatically route employee questions (benefits, payroll, onboarding) to the right internal knowledge articles, with minimal coding. Which approach best fits this requirement?

A data scientist evaluates a binary classifier for fraud detection where fraud is rare. Accuracy is 98%, but many fraud cases are missed. Which metric is most appropriate to focus on to reduce missed fraud cases?

Which statement best describes the difference between narrow AI and artificial general intelligence (AGI)?

A team is building a model for loan approval. They find that the training dataset contains historical bias against a protected group. Which action is the best first step to address fairness before deploying the model?

A retail company wants to forecast weekly demand for each store using past sales, promotions, and holidays. Which machine learning problem type best matches this requirement?

A team notices their model performs much better on the training set than on the validation set. Which is the most likely issue and a common mitigation?

A customer support bot must identify the user's intent (e.g., "reset password", "track order") and extract entities (e.g., order number) from messages. Which IBM Watson capability is most directly aligned with this need?

A model is deployed to production and initially performs well, but performance degrades over months as customer behavior changes. Which governance/operations practice best addresses this issue?

A regulated organization must provide explanations for individual credit decisions made by an ML model. Which approach best supports per-decision interpretability while balancing model performance?

A bank trains a model to predict default risk. During evaluation, it discovers that a feature representing "days since last address change" indirectly encodes sensitive socioeconomic patterns, causing disparate impact. What is the best remediation strategy that maintains compliance and model utility?

A team is building an AI feature to categorize incoming customer emails into a small set of routing queues (e.g., Billing, Technical Support, Account Changes). They have historical emails already labeled by queue. Which learning approach is most appropriate?

A data scientist is evaluating a binary classifier for a rare disease screening program. Missing a true positive is much more costly than raising an extra false alarm. Which metric should be prioritized?

A product owner asks why an AI model that performs well during training sometimes performs poorly on new, unseen data. Which concept best explains this?

A company wants to extract key phrases and overall sentiment from customer reviews to track brand perception. Which IBM Watson service capability best matches this need?

A team is training a model to predict house prices. They notice the model performs extremely well on the training set but poorly on validation data. Which action is the best next step to reduce this issue?

An organization deploys an AI model for loan approvals. Regulators require that decisions be explainable to applicants. Which approach best supports this requirement?

A chatbot built for customer support frequently returns confident answers that are not supported by the company’s documentation. The team wants to ground responses in verified content. Which architecture pattern best addresses this?

A company is using a dataset where one sensitive group is underrepresented. After deployment, that group receives significantly worse outcomes from the model. Which action is most appropriate to mitigate this issue?

A machine learning pipeline shows strong performance in offline testing, but after deployment the model’s accuracy steadily declines over several weeks. Data collection and labeling processes remain unchanged. What is the most likely cause?

A healthcare organization wants to use an external AI service to analyze patient text notes. They must ensure that only the minimum necessary data is processed and that access is tightly controlled. Which design choice best aligns with this requirement?

A customer support team wants to automatically route incoming emails into categories such as 'billing', 'technical issue', and 'account access'. They have historical emails already labeled with the correct category. Which approach is most appropriate?

Which statement best describes overfitting in a machine learning model?

A team is using a large language model to draft customer-facing responses. Which control most directly helps reduce the risk of the model producing sensitive personal information in its outputs?

Which description best matches IBM Watson Natural Language Understanding (NLU)?

A model predicts whether a transaction is fraudulent. Only 1% of transactions are fraud. The business is concerned about missing fraud cases, even if it means investigating more false alarms. Which metric should be prioritized?

A retail company trained a demand-forecasting model. Recently, predictions became significantly less accurate, even though the code and model version have not changed. Which issue is the most likely cause?

A team wants to build a conversational assistant that can answer questions from a set of internal policy documents. They want the assistant to cite the most relevant passages and reduce hallucinations. Which architecture is the best fit?

A data scientist notices a model has much higher error rates for one demographic group than for others. Which action is the best first step in a responsible AI workflow?

A logistic regression model for loan approval has high training accuracy but performs poorly on new applicants. The team also notices the model coefficients are extremely large in magnitude. Which remediation is most appropriate?

An organization plans to deploy an AI system that provides medical triage recommendations. Which governance practice is most important to implement before production use?

A business stakeholder asks whether an AI system is "learning" when it uses a fixed set of if/then rules created by developers. Which statement best describes this system?

A team is building a binary classifier to detect fraudulent transactions. Only 1% of transactions are fraud. Which evaluation metric is generally most informative for this imbalanced dataset?

A project manager wants to reduce risk before deploying an AI model that may affect customers (e.g., loan approvals). Which action is a foundational best practice for responsible AI?

A team deploys an image classifier and notices performance drops when users upload darker, low-light photos. Which issue is most likely causing the drop?

A support organization wants an assistant that answers questions using internal policy PDFs while citing sources. They want to reduce hallucinations and keep responses grounded in their documents. Which approach best fits?

An ML engineer sees strong training performance but poor validation performance on a neural network. Which action is most likely to improve generalization?

A bank wants to use AI to assist human agents in customer calls by suggesting next-best actions. They must ensure the system's suggestions do not unintentionally discriminate against protected groups. What is the most appropriate governance step before rollout?

A developer integrates an IBM-hosted generative AI service into a chatbot. They observe that the model sometimes follows user instructions to reveal confidential internal procedures. Which mitigation is most effective at the application layer?

A company trained a model to predict employee attrition using features that include "manager satisfaction score". Later they discover the score was computed using exit interview outcomes, which are unavailable at prediction time. What is the primary issue?

An enterprise wants to deploy a generative AI assistant that answers questions about proprietary engineering designs. Requirements: (1) data must remain within the organization’s controlled environment, (2) access must be restricted by user role, and (3) responses should be auditable. Which architecture best meets these requirements?