AI Governance And Risk practice questions

Absence of automated model deployment pipelines

Why it fails: Deployment pipelines concern release automation, not the metadata needed to trace model versions, lineage or changes, so they leave the inventory's versioning gap open. Pipeline tooling is tempting because it feeds the inventory automatically; it would be the right control when the risk is unauthorised or untracked model releases.

Failure to define the RPO for the primary database

Why it fails: RPO is a business continuity requirement, not specifically related to model inventory metadata.

Insufficient budget for cloud storage of historical data

Why it fails: Budget is a resource concern, not a structural risk management control gap.

None of the above

Why it fails: This is not a valid answer for a mandatory compliance gap.

Update the office building safety guidelines

Why it fails: Building safety does not mitigate third-party AI risk.

Ask the vendor for their marketing presentation

Why it fails: Marketing materials are insufficient for risk assessment.

Ignore the missing document if the vendor is well-known

Why it fails: Well-known brands can still pose significant AI risks.

Vertex AI Model Monitoring for data drift

Why it fails: This monitors drift, not PII leakage during inference.

Vertex AI Feature Store access logs

Why it fails: This would log features but does not provide a control to prevent PII logging during inference.

Cloud IAM role binding for Vertex AI

Why it fails: IAM controls access permissions but does not scrub content from inference logs.

Increased training latency

Why it fails: This is a performance risk, not a fundamental security or governance risk.

Lack of access to training hardware

Why it fails: Hardware access is a capacity planning issue, not an architectural risk.

Centralized data breaches

Why it fails: Federated learning is designed to avoid centralizing data, reducing this specific risk.

The programming language used by the vendor

Why it fails: A vendor's programming language is an implementation detail with no bearing on governance outcomes such as data handling, model transparency or accountability. It is tempting because technical stack choices feel relevant to risk, and would matter for integration or support compatibility assessments instead.

Vendor's model update and retraining frequency

Why it fails: Retraining frequency is a model lifecycle and performance concern, not a governance concern; governance addresses accountability, transparency, risk and compliance for third-party AI. It is tempting because frequent retraining can shift model behaviour, but that is a change-management or validation issue rather than a governance control over the provider relationship.

The physical location of the vendor's corporate office

Why it fails: A vendor's corporate office location is a commercial/administrative detail, not a governance control over data handling, model training or accountability. It tempts because data-residency rules can hinge on geography, but those attach to processing and storage locations, not the head office address.

Access control lists for the model production environment

Why it fails: Access control lists govern who may read or modify artefacts in the production environment; they log authorisation decisions, not the derivation path of features into training data. ACLs are the right control when the finding is unauthorised access to models, not a missing lineage chain between feature store and pipeline.

Encryption of data at rest in the data warehouse

Why it fails: Encryption at rest protects stored data confidentiality but records nothing about how features were derived, transformed or moved, so it cannot reconstruct the broken chain. It is the right control when the audit finding concerns unauthorised disclosure of warehouse data rather than traceability between feature store and training pipeline.

Load balancing configuration for the inference API

Why it fails: Load balancing distributes inference requests across replicas for availability and throughput; it captures no metadata about feature provenance or transformation steps, so it cannot repair lineage. It would be the correct control where the audit finding is inference latency or single-point-of-failure in the serving tier.

Conducting weekly penetration tests on the vendor's office premises

Why it fails: Penetration testing the vendor's office premises addresses physical building security, which is unrelated to the AI service's governance, data processing or model behaviour. It is tempting because penetration testing is a recognised assurance activity, and would be correct if the policy covered physical site security of the vendor.

Requesting the vendor's employee turnover rates for the last fiscal year

Why it fails: Employee turnover rates say nothing about how the vendor's AI system is governed, so they cannot evidence alignment with the firm's policy. It is tempting because staff stability can signal organisational health, and would be correct if the policy specifically required vendor workforce-continuity assurances.

Requiring the vendor to submit copies of all their marketing brochures

Why it fails: Marketing brochures are promotional material, not evidence of the vendor's AI risk controls, data handling or model governance, so they cannot demonstrate policy compliance. It is tempting because vendor documentation is easy to request, and would be correct if the policy required reviewing published product claims.

The physical location of the cloud server hosting the model

Why it fails: Server geography is fixed infrastructure metadata and does not change as model behaviour degrades; it affects latency, residency and compliance instead. It would be the correct consideration when choosing a deployment region, not when monitoring reliability drift over time.

The number of lines of code in the model deployment script

Why it fails: Deployment script length is a static code attribute, unaffected by live inference traffic, so it cannot signal reliability decay. It would matter when reviewing maintainability or technical debt, not when detecting drift, data quality shifts or accuracy decline in production.

The number of users accessing the model API

Why it fails: API user counts measure adoption or load, not predictive correctness; traffic can rise while accuracy falls, or drop while the model stays sound. It would be the right metric for capacity planning or licensing, not for spotting drift, bias or degradation in model outputs.

Managing the day-to-day procurement of data labeling services

Why it fails: Procuring data labelling services is operational execution, whereas board-level oversight sets direction, risk tolerance and accountability. It is tempting because labelling quality directly affects AI outcomes, and would be correct when assessing an operational or procurement function rather than the governance committee's oversight responsibilities.

Reviewing individual code commits for AI model training

Why it fails: Reviewing individual code commits is hands-on engineering review, not board-level oversight, which addresses strategy, risk and accountability. It is tempting because code quality underpins model safety, and would be correct when assessing engineering assurance or peer review processes rather than the governance committee's remit.

Selecting the machine learning framework used by data scientists

Why it fails: Choosing an ML framework is an engineering implementation decision, not board-level oversight, which concerns risk appetite, strategy and accountability. It is tempting because framework selection genuinely matters for governance outcomes, and would be the right answer when assessing the data science team's technical standards rather than the committee's oversight role.

Describing the history of artificial intelligence research

Why it fails: Historical narrative about AI research assigns no owner, approval path or oversight duty, so it cannot establish accountability. Policy accountability needs defined roles and responsibilities. A research-history section suits an awareness or training document introducing AI concepts to staff, not a governance policy.

Listing the hardware specifications of the training clusters

Why it fails: Hardware specifications describe infrastructure capacity, not who owns AI decisions or answers for outcomes. Accountability demands named roles, responsibilities and escalation paths. Listing training-cluster specs belongs in technical architecture or capacity-planning documentation, where engineers size compute for model training.

Providing the contact information for the company's PR department

Why it fails: A PR contact routes media enquiries; it assigns no ownership of AI risk decisions, approvals or monitoring. Accountability requires named roles and responsibilities. Publishing press contacts belongs in a communications or incident-response plan, where external enquiries about an AI incident need a single spokesperson.

Inability to detect model drift in production environments

Why it fails: Drift detection is a monitoring control comparing production inputs and outputs against baselines; independent validation assesses design, data and assumptions before or during use. It is tempting because both are model-risk controls, but drift monitoring would be the correct answer if the stem concerned production performance monitoring.

Non-compliance with internal cybersecurity patching cycles

Why it fails: Independent validation addresses model accuracy, bias and conceptual soundness, not patch cadence, so this misidentifies the control being omitted. It is tempting because MRM sits within broader governance and cybersecurity programmes, and patch cycles are a genuine compliance obligation, but they are governed by separate frameworks.

Operational silos between the AI team and the IT department

Why it fails: Independent validation is a governance control performed by a party separate from model developers; organisational silos are a collaboration symptom, not the direct consequence of omitting it. It is tempting because validation does require cross-team engagement, but silos stem from operating-model design rather than this specific MRM gap.

Creating a repository of all AI ethical guidelines published by NGOs

Why it fails: Collecting NGO ethical guidelines gathers external reference material; it does not state the organisation's own willingness to accept AI risk. A repository would be useful when benchmarking against industry frameworks, but risk appetite must be an internally agreed, board-approved position with defined limits, not a library of third-party publications.

Providing the board with a list of all AI vendors used in the organization

Why it fails: A vendor inventory documents procurement, not tolerance for risk. AI risk appetite requires the board to articulate acceptable levels of risk across AI activities, with measurable thresholds. Vendor lists support third-party due diligence and supply-chain oversight, which would be the right artefact when the requirement is visibility of external providers rather than appetite definition.

Restricting all AI development to on-premises environments

Why it fails: Restricting development to on-premises constrains deployment architecture, not risk tolerance. AI risk appetite is a governance statement defining how much risk the organisation will accept in pursuit of AI objectives. On-premises hosting would be relevant where data residency or sovereignty requirements dictate infrastructure, but it expresses no appetite threshold.

Employee performance review scores

Why it fails: Performance review scores describe staff, not models, so they record no model purpose, owner or validation status. It is tempting because HR records are a real governance input, and a workforce-capability or resourcing assessment would correctly use them — but model risk inventories require model-specific metadata instead.

List of office furniture used by the team

Why it fails: Furniture inventories carry no model lineage, performance metrics or risk controls, so they cannot document AI risk exposure. It is tempting because asset registers are a genuine governance artefact, and a facilities or procurement audit would legitimately list office equipment — but that is an IT asset question, not model risk.

Marketing budget for the model launch

Why it fails: A marketing budget is a financial planning artefact with no bearing on model risk; an inventory records models, owners, intended use, training data, evaluation results and monitoring status. It is tempting because launch planning touches the model, but budgeting belongs to product finance, not risk governance.

The programming language used to build the model

Why it fails: The programming language is an implementation detail that does not influence whether outcomes are equitable across groups; fairness depends on data representation, label quality and error-rate parity. It is tempting because language choice affects library availability and performance tuning, and would matter when selecting a stack for reproducibility or deployment constraints, not for assessing bias.

The number of CPU cores allocated for training

Why it fails: CPU core allocation affects training throughput and cost, not the distribution of outcomes across groups, so it cannot measure Fairness. It is tempting because compute provisioning genuinely matters when sizing training infrastructure or diagnosing slow convergence. Fairness instead requires evaluating disparities in model performance across protected groups, which hardware allocation never captures.

Requiring the data scientist to publish the code on an internal repository

Why it fails: Publishing code on an internal repository addresses reproducibility and peer review, not the explainability a black-box model lacks in high-stakes lending. Repository publication would be correct for collaborative development or version control, but an AI audit framework requires interpretability evidence, such as documented feature attributions or a surrogate model.

Allowing the model if it achieves 99% accuracy on the test set

Why it fails: Accuracy on a test set measures predictive performance, not explainability, so it cannot satisfy an audit requirement for reasoned, contestable decisions in lending. It tempts because benchmark accuracy is the usual model-selection metric, and would suffice where outcomes carry no regulatory duty to explain.

Asking the marketing team to verify the model's fairness

Why it fails: Marketing staff hold no mandate over credit-model fairness, and their review produces no documented, auditable evidence of bias testing. It tempts because cross-functional review sounds like governance, and would fit where the concern is brand or campaign impact rather than regulated lending decisions.

To reduce the time spent on manual data entry

Why it fails: HITL concerns oversight of model outputs, not clerical workload; it typically adds review steps rather than removing data-entry effort. Reducing manual entry is the aim of straight-through processing or RPA, where no human judgement is needed.

To replace all human employees with automated systems

Why it fails: HITL inserts human review at decision points to catch errors and preserve accountability; it does not remove staff. Automation that eliminates human roles is full autonomy, the opposite design goal, so this misreads the control's purpose entirely.

To increase the speed of model training

Why it fails: Training speed is governed by compute, batch size and optimiser settings, not by human oversight. HITL inserts human review or approval into an AI workflow to catch errors and handle edge cases before action is taken. It would be the right choice when decisions carry legal, safety or financial risk requiring human sign-off.

A list of all employees who have taken an AI training course

Why it fails: A training attendance list records who completed awareness courses; it neither defines audit boundaries nor states control objectives. AI audit scope is set by the AI governance policy, risk assessment and audit charter, which identify systems and controls to examine. A training roster would support evidence gathering for competency controls, not scoping.

The company's social media posting guidelines

Why it fails: Social media posting guidelines govern public communications, not AI system boundaries, risk criteria or control objectives. An AI audit's scope derives from the AI governance policy and applicable standards, which define systems, processes and controls in scope. Posting guidelines would be relevant only when auditing marketing content approval workflows.

The employee handbook for the engineering department

Why it fails: The engineering employee handbook covers HR matters such as conduct, leave and disciplinary process, not AI system scope, risk classification or control requirements. AI audit scope comes from the AI governance framework and audit charter. A handbook would be the correct source only when auditing workforce policy compliance, not AI systems themselves.

The Model Building Standards Act

Why it fails: This is not a real regulation.

The Sarbanes-Oxley Act

Why it fails: The Sarbanes-Oxley Act governs financial reporting integrity and corporate accounting controls, not the processing of personal data, so unanonymised PII in training data falls outside its scope. It is tempting because SOX does mandate controls over data accuracy and retention, making it the right answer when the scenario concerns financial records rather than privacy.

The Internal Server Maintenance Code

Why it fails: This is an invented internal IT procedure, not a regulation, and it concerns server upkeep rather than data protection. It is tempting because internal codes sound authoritative, yet PII misuse in training data is governed by privacy legislation such as the GDPR.

The history of the company's founding

Why it fails: Company founding history has no bearing on detecting, containing or reporting AI incidents, so it satisfies no incident-response control. It is tempting as background context, but compliance requires elements such as defined roles, escalation paths, notification timelines and evidence retention tied to the plan's procedures.

A log of all office supply purchases

Why it fails: Office supply purchases are unrelated to AI incident handling, so recording them fulfils no regulatory requirement within the response plan. It is tempting because logging and record-keeping are compliance staples, but that logic applies to model versioning, incident timestamps and decision audit trails, not procurement records.

A list of all cafeteria menus in the building

Why it fails: Cafeteria menus have no bearing on detecting, containing or reporting AI incidents, so they cannot satisfy any compliance obligation in an AI incident response plan. It is tempting because documented procedures are auditable artefacts, but that reasoning applies to incident classification, notification timelines and evidence retention, not building amenities.

Increasing the number of neurons in the hidden layers

Why it fails: Adding neurons expands model capacity, which addresses underfitting, not the changing statistical relationship between inputs and target over time. Retraining on recent data is what counters drift; widening hidden layers suits a model that cannot capture the complexity already present in stationary data.

Moving the model to a larger GPU cluster

Why it fails: Extra GPU capacity raises throughput and shortens training time; it does not alter the input distribution or retrain the model, so predictions still degrade as the relationship between features and labels shifts. Larger clusters suit scaling compute for training runs, not correcting drift.

Restricting access to the training dataset

Why it fails: Restricting dataset access is a governance and privacy control; it neither detects nor corrects the shift in input distribution or label relationship that defines drift. Scheduled retraining on fresh data is the remedy; access restriction suits protecting sensitive training records from unauthorised use.