IBM A1000-075: Foundations of AI Advanced Practice Exam: Hard Questions 2025

A bank is building an AI system to triage inbound customer messages (chat + email) into categories and suggested next actions. During pilot, it performs well on common categories but fails catastrophically on rare but high-risk intents (e.g., fraud reports), sometimes assigning them to “general inquiry.” The team can’t collect many new labeled examples quickly. Which approach most directly mitigates this failure mode while remaining feasible under the constraint?

A healthcare provider wants to deploy a clinical decision support tool. The model outputs a risk score. During validation, the team observes strong discrimination (AUC is high) but poor calibration: predicted probabilities consistently overestimate true event rates in some subpopulations. The provider needs probabilities to support shared decision-making and threshold-based interventions. What is the best next step?

An organization is selecting an AI approach for automating customer-support responses. They need: (1) grounded answers only from internal policy documents, (2) traceability to sources, (3) the ability to refuse when sources are insufficient, and (4) minimal risk of hallucination. Which architecture best fits these constraints?

A global retailer uses IBM Watson services to analyze customer feedback across languages. After rollout, sentiment results are inconsistent: identical translated phrases yield different sentiment depending on the original language, and executives question reliability. Which troubleshooting path is most effective to isolate the root cause?

A financial services firm is integrating an AI assistant into a call-center workflow. Regulatory requirements demand: auditability of interactions, separation of duties (no single component can both generate and approve actions), and the ability to reproduce outputs for an investigation. Which design best meets these requirements?

A manufacturer uses a Watson-based virtual agent to answer product safety questions. Legal requires that when the agent is uncertain or the query concerns safety, the system must provide an approved escalation path and avoid generating speculative guidance. Which control is most appropriate?

A data science team trains a churn model and reports excellent validation metrics. Later, production performance collapses. Investigation shows that a feature “days_since_last_ticket_closed” was computed using ticket closure timestamps that, in training, included closures occurring after the churn label window. What is the most accurate diagnosis and fix?

A utility company trains a model to predict equipment failure. Failures are rare (0.2%). Missing a true failure is extremely costly, but unnecessary maintenance also has a cost. Which evaluation approach best aligns model selection with business impact?

An insurer deploys a claims triage model and notices higher false-positive rates for one protected group, leading to more manual reviews and slower service. The business requires both improved fairness and preserved detection of genuinely suspicious claims. Which action is most defensible and effective as a first step?

A company wants to use customer support transcripts to train and evaluate an AI assistant. Transcripts contain names, addresses, and occasionally health information. The company must minimize privacy risk while still enabling model improvement and auditability. Which governance approach best balances these needs?