USEFULPARTSONLY
USEFUL PARTS ONLYAWS AI Business Strategist: The Prep Kit

AIB-C01 Practice Answer Key

Useful Parts Only · Independent practice material. Bank order is fixed here; the mock shuffles questions and options.

When printing or saving as PDF, turn off Headers and footers in the browser print dialog to omit the local URL and timestamp.

Bank item 1 · Domain 1 · Select one

A wholesaler must apply a statutory surcharge of exactly 2% whenever an invoice exceeds £10,000. The rule is explicit, stable, and must be reproducible in an audit. Which solution should the strategist recommend?

  1. Train a classifier on invoices with reviewed surcharge labels.
  2. Apply the threshold through a deterministic calculation rule. · Correct
  3. Prompt a language model with worked surcharge calculation examples.
  4. Use a regression model fitted to previous invoice surcharges.

Why: The exact threshold and calculation are already known. A deterministic rule directly implements them and supports reproducible audit results. Classification and regression would estimate something that can be calculated exactly. Worked prompts do not guarantee exact arithmetic.

Bank item 2 · Domain 1 · Select one

A streaming company wants to discover viewing patterns among new customers. It has viewing histories but has not defined customer segments or labeled examples. Which approach is most appropriate?

  1. Classify customers using assigned preferred-genre training labels.
  2. Summarize viewing histories and rank customers by viewing frequency.
  3. Cluster customers using similarities in their viewing histories. · Correct
  4. Predict each customer's cancellation risk using historical renewal outcome labels.

Why: The objective is to discover behavioral groups without predefined labels, so clustering fits. Preferred-genre classification and cancellation prediction need defined target outcomes and answer different questions. Individual summaries do not directly identify groups.

Bank item 3 · Domain 1 · Select two

A maintenance assistant cites guidance for the wrong equipment version. Both versions are indexed, but chunks omit model identifiers and retrieval searches the entire library. Which TWO changes most directly address the cause?

  1. Keep equipment identifiers attached to their relevant document sections. · Correct
  2. Add examples teaching the model to prefer recently published manuals.
  3. Filter retrieved passages by equipment version and document approval status. · Correct
  4. Train the model to mention version uncertainty in its answers.
  5. Expand the response allowance to accommodate several possible procedures.

Why: The observed failure is retrieval of guidance for the wrong version. Preserving identifiers in chunks and filtering by version and approval status directly improve evidence selection. Recency alone does not identify the applicable equipment; uncertainty wording and longer responses do not fix retrieval.

Bank item 4 · Domain 1 · Select one

A claims team has 25,000 reviewed examples mapping claim descriptions to a fixed schema. It has tested explicit instructions and representative few-shot examples, but the base model still applies the schema inconsistently. Policy facts are already available and current. Which model adaptation is best to evaluate next?

  1. Retrieve more policy references for each claim.
  2. Continue pre-training on unlabeled claims to improve general domain fluency.
  3. Increase the context window without changing the examples or task instructions.
  4. Fine-tune the model on reviewed claim-to-schema pairs. · Correct

Why: Supervised fine-tuning targets the repeated mapping and output behavior. The team has already tested representative examples and current grounding. More passages or context would not directly teach the mapping, and unlabeled pre-training targets broader language familiarity.

Bank item 5 · Domain 1 · Select one

A materials company wants to adapt a foundation model so it better understands specialist vocabulary across tasks, even without reference passages at inference. It has a large collection of unlabeled research papers. Which approach best matches the goal and available data?

  1. Continue pre-training on the unlabeled research corpus. · Correct
  2. Fine-tune on labeled examples of the report format.
  3. Retrieve relevant research passages at each request.
  4. Distill the model using responses from its current version.

Why: Continued pre-training learns from unlabeled text and can improve general familiarity with the domain. Retrieval supplies facts per request, supervised fine-tuning targets the behavior in labeled examples, and distillation mainly transfers behavior from an existing model.

Bank item 6 · Domain 1 · Select one

An assistant reliably summarizes short contracts. When users upload several long contracts together, the service rejects the request before generating an answer because the supported context length is exceeded. What is the best response?

  1. Shorten the requested answer while retaining all uploaded contract text.
  2. Use the same input with a larger maximum output token allowance.
  3. Summarize in stages or retrieve relevant sections. · Correct
  4. Fine-tune on contract examples before resubmitting the complete document bundle.

Why: The service rejects the input because it exceeds context capacity. Staged processing or selecting relevant sections reduces the input in each request. Shorter outputs may free some space, but do not solve inputs that themselves exceed supported capacity. Fine-tuning does not expand that capacity.

Bank item 7 · Domain 1 · Select one

A fraud model generates 200 alerts. Investigators confirm that 50 alerts are fraud; the others are legitimate transactions. The strategist wants the proportion of alerts that are genuinely fraud. Which statement is correct?

  1. Recall is 25%; the other alerts represent missed fraud.
  2. Precision is 25%; calculating recall also requires missed fraud information. · Correct
  3. False-positive rate is 75%; the alert set provides its complete denominator.
  4. Accuracy is 25%; confirmed fraud divided by alerts gives overall correctness.

Why: Precision is confirmed fraud divided by all fraud alerts: 50/200 = 25%. Recall needs the total actual fraud, including missed cases. False-positive rate needs all actual legitimate transactions, not just alerts. Overall accuracy also needs unflagged transactions.

Bank item 8 · Domain 1 · Select two

A procurement assistant will compare suppliers, request quotes through approved tools, and prepare purchase orders. Leaders are considering agentic behavior. Which TWO characteristics distinguish an agentic solution from a text-only question-answering assistant?

  1. Explain why a supplier fits the purchasing criteria.
  2. Summarize supplier terms for a side-by-side comparison.
  3. Retrieve procurement policies relevant to the requested purchase.
  4. Call approved supplier tools to request quotes. · Correct
  5. Plan purchasing steps and adapt when one fails. · Correct

Why: Authorized tool execution and multi-step goal-directed action distinguish the proposed agentic behavior. Fluent explanation, summarization, and retrieval can all be provided by a question-answering assistant without autonomous action. Agent capabilities still require appropriate boundaries.

Bank item 9 · Domain 1 · Select one

Three call centers label identical complaint types differently. A model trained on the combined labels routes similar complaints to different teams. Which action should leaders prioritize?

  1. Agree common labeling definitions and review inconsistent examples across all centers. · Correct
  2. Train separate models and compare their average regional routing accuracy.
  3. Add more historical complaints while preserving existing regional label definitions.
  4. Tune routing thresholds using the combined labels as the reference standard.

Why: Similar complaints have contradictory labels, so leaders should first agree the target meaning and review consistency. More examples preserve the contradiction. Separate models may reproduce regional inconsistencies, while threshold tuning against unreliable labels does not establish a shared routing standard.

Bank item 10 · Domain 1 · Select one

A model understands customer emails but produces inconsistent summaries. The team has no agreed summary structure or definition of what to include. Which is the most appropriate low-cost first step?

  1. Supply examples of customer history without changing the requested output.
  2. Increase model capacity while keeping the summary instructions unchanged.
  3. Fine-tune on existing summaries before defining the desired structure.
  4. Define the summary structure and test prompts with examples. · Correct

Why: The team has not defined a consistent target. Explicit instructions, an agreed template, and representative examples establish a low-cost baseline. More context or a larger model does not resolve disagreement about what to include. Training on inconsistent summaries can reproduce inconsistent behavior.

Bank item 11 · Domain 1 · Select one

Procurement wants a vendor to demonstrate an independently certified AI management system. Which standard is the relevant certification target?

  1. ISO/IEC 23053, covering the framework for AI systems using machine learning.
  2. ISO/IEC 27001, covering the management of information security within an organization.
  3. ISO/IEC 42001, covering requirements for an AI management system. · Correct
  4. ISO 9001, covering requirements for an organization-wide quality management system.

Why: The requested certification is specifically for an AI management system: ISO/IEC 42001. ISO/IEC 23053 is a framework for AI systems using ML. Information-security and general quality management standards address related but different scopes.

Bank item 12 · Domain 1 · Select one

A retailer applies an existing demand model to next week's planned promotions. No learning procedure runs and model parameters remain unchanged. What is this activity?

  1. Training, because the inputs are new.
  2. Inference, because an existing model produces predictions. · Correct
  3. Validation, because forecasts are being compared with known weekly demand outcomes.
  4. Fine-tuning, because the next promotion cycle requires adapting the learned parameters.

Why: Inference applies an existing model to new inputs without updating its learned parameters. Training and fine-tuning change parameters. Validation compares model behavior with reference outcomes; the scenario instead describes generating predictions for a future week.

Bank item 13 · Domain 2 · Select one

A distributor must choose its first AI pilot. A forecasting use case has an accountable owner, usable historical data, measurable inventory costs, and strong strategic alignment. A public-facing advisor promises greater revenue but lacks approved data and has unresolved risk controls. What should leaders do?

  1. Pilot the advisor first, then address data readiness.
  2. Run technical demonstrations and select the highest-accuracy solution.
  3. Postpone both pilots until a common enterprise data platform can support every proposed business use case.
  4. Pilot forecasting against value gates; address the advisor readiness gaps separately. · Correct

Why: The forecasting proposal combines measurable value, usable data, ownership, and strategic fit. Testing it while closing the advisor gaps is proportionate. Speculative revenue alone does not outweigh readiness risks, and technical accuracy alone is not the portfolio objective. A complete common platform is not stated as a prerequisite for forecasting.

Bank item 14 · Domain 2 · Select one

Over a two-year evaluation period, an AI initiative will generate £540,000 in total benefits. Implementation costs are £180,000 and total operating costs over the same period are £120,000. Ignoring discounting, what is the ROI?

  1. 80% · Correct
  2. 180%
  3. 44.4%
  4. 120%

Why: Total costs over the two years are £180,000 + £120,000 = £300,000. Net benefits are £540,000 − £300,000 = £240,000. ROI = £240,000 / £300,000 = 80%. The benefit-to-cost ratio is 180%, which does not subtract costs.

Bank item 15 · Domain 2 · Select one

An AI workflow requires a £240,000 initial investment. It produces £30,000 in monthly gross savings and incurs £10,000 in monthly operating costs. Benefits start immediately and remain constant. What is the simple payback period?

  1. 8 months
  2. 10 months
  3. 12 months · Correct
  4. 24 months

Why: Net monthly savings are £30,000 − £10,000 = £20,000. Dividing the £240,000 initial investment by £20,000 gives 12 months. Eight months uses gross savings and omits the operating costs.

Bank item 16 · Domain 2 · Select two

A service team plans to use AI to reduce time spent resolving support cases without increasing rework. No pilot has started. Which TWO measurement decisions should it make now so the team can evaluate the result?

  1. Use forecast savings from the business case as the starting point.
  2. Record resolution time and reopen rates before rollout. · Correct
  3. Count assistant-generated drafts as completed cases after the rollout.
  4. Define consistent case categories and measurement windows for both periods. · Correct
  5. Compare early satisfaction survey responses with last year’s annual finance results.

Why: A documented baseline captures the starting state. Consistent definitions and comparable windows make the before-and-after comparison interpretable. Forecasts, drafts and unrelated reporting windows are not substitutes for a baseline. A comparable baseline is useful evidence of change, but does not on its own prove attribution.

Bank item 17 · Domain 2 · Select one

A demand planning pilot is six weeks from evaluation. It depends on eight regional data feeds. Which metric is the best leading indicator of delivery readiness?

  1. Forecast error on a prototype running on sample regional feeds.
  2. Expected inventory savings across the regions in the pilot business case.
  3. Share of regional teams registered for training on the new workflow.
  4. Share of required regional feeds fully integrated and verified against the pilot’s quality criteria. · Correct

Why: Integrated, quality-checked feeds are a direct prerequisite for evaluation. Prototype accuracy on sample feeds does not prove integration; projected savings and training registrations do not measure feed readiness.

Bank item 18 · Domain 2 · Select one

A small finance team needs a standard invoice-processing capability in three months. It has no ML engineering capacity, and several vendor products satisfy its mandatory security and accuracy requirements. The capability is not a competitive differentiator. Which choice best fits?

  1. Build an internal model and engage temporary engineers to meet the initial deadline.
  2. Buy a suitable product after integration and cost checks. · Correct
  3. Commission a bespoke solution with ownership transferred to the team after deployment.
  4. Extend the timeline to develop internal modeling capabilities before selecting a solution.

Why: Buying fits a standard need with suitable products, little internal capacity, a short deadline, and no stated differentiation benefit from customization. Integration, cost, and contracts still need review. Bespoke or internal development introduces additional delivery and capability burdens without a matching strategic requirement.

Bank item 19 · Domain 2 · Select two

A seasonal campaign assistant will have uncertain usage and no established baseline. Finance wants low upfront commitment without losing cost control. Which TWO actions best fit?

  1. Reserve peak capacity upfront.
  2. Start with usage-based pricing so spending can follow actual requests without a long commitment. · Correct
  3. Budget from average token use without testing campaign traffic spikes.
  4. Set budget alerts and model several traffic scenarios, including unexpected peaks in demand. · Correct
  5. Buy annual seats for expected users.

Why: Consumption pricing can avoid large initial capacity commitments when demand is uncertain. Budgets and multiple demand scenarios help manage cost exposure. Peak reservations or broad annual seat commitments create commitments without a reliable baseline; a single average assumption hides variability.

Bank item 20 · Domain 2 · Select one

An AI tool saves analysts an estimated 2,000 hours annually. The CFO asks whether those hours should immediately be booked as payroll savings. Which answer is most accurate?

  1. Treat hours saved as capacity; link cash savings to spending changes. · Correct
  2. Record saved hours as payroll reductions using the fully loaded employee hourly rate.
  3. Recognize the full estimated saving once staff confirm that drafting became faster.
  4. Classify the benefit as revenue growth using the value of additional available hours.

Why: Time savings free capacity, but do not automatically reduce payroll or generate revenue. Cash savings require a realization mechanism such as reduced spending or avoided hiring, while revenue claims need evidence that the capacity generated additional sales. A credible case separates these effects.

Bank item 21 · Domain 2 · Select one

Two competitors can license the same foundation model. Which investment is most likely to create a durable advantage for one of them?

  1. Negotiate exclusive pilot access to that model.
  2. Increase model size to improve the sophistication of responses across every business workflow.
  3. Combine proprietary data, distinctive workflows, and customer feedback. · Correct
  4. Accelerate launch using a standard workflow and compete primarily through the AI announcement.

Why: Proprietary data, integrated workflows, and accumulated feedback can be difficult to reproduce and can reinforce customer value. Temporary access advantages, model size, or publicity alone do not establish a durable advantage when competitors can use similar underlying models.

Bank item 22 · Domain 2 · Select two

A bank is replacing its AI assistant platform. A successful demonstration used only public documents, while production includes restricted documents and critical workflows. Which TWO requirements should gate the transition?

  1. Compare provider service-level commitments.
  2. Benchmark candidate platforms on public documents.
  3. Move documents in batches while retaining the old platform’s broad access defaults.
  4. Test whether restricted documents stay protected and critical workflows meet quality and performance requirements. · Correct
  5. Stage the production cutover, keep the old route available, and rehearse rollback before expanding traffic. · Correct

Why: The demonstration excluded the restricted data and critical workflows that matter in production. Representative tests and a rollback-capable transition address permission enforcement, quality, performance, and continuity. An SLA or general benchmark cannot replace those checks, and permission reconciliation must not wait until after exposure.

Bank item 23 · Domain 2 · Select one

After two iterations, a pilot meets accuracy targets but misses its agreed user-value and cost gates. The sponsor argues that it must scale because £200,000 has already been spent. What should the steering group do?

  1. Expand gradually so more teams can absorb the sunk investment.
  2. Pause expansion and reassess the forward-looking business case. · Correct
  3. Continue until the delivery budget ends.
  4. Approve rollout against the accuracy gate and review cost once adoption stabilizes.

Why: Past spending is sunk. The missed value and cost gates justify pausing expansion and deciding whether a credible revision can meet them. Technical accuracy does not replace those gates.

Bank item 24 · Domain 3 · Select one

A group uses AI for internal meeting summaries and for recommendations affecting credit access. Its policy applies identical reviews to both. What is the best governance improvement?

  1. Apply the highest-risk review to every use case.
  2. Set review requirements from the vendor’s model rating.
  3. Tier controls by the potential harm and autonomy of each use case; revisit them when scope changes. · Correct
  4. Use one launch checklist and review exceptions after incidents.

Why: Risk depends on how a system is used and its possible impact, so use cases need proportionate controls across the lifecycle. Provider-level ratings miss application context. Applying maximum controls everywhere can impede low-risk work, while waiting for incidents leaves preventable risks unmanaged.

Bank item 25 · Domain 3 · Select one

A lending model meets the overall accuracy target, but testing reveals materially worse error rates for one customer group. Launch is scheduled next week. What should leaders do before approving the affected use?

  1. Pause approval, investigate the disparity, and validate proportionate mitigation. · Correct
  2. Launch with subgroup monitoring.
  3. Launch to a smaller audience and assess group outcomes during the next review cycle.
  4. Raise the aggregate accuracy gate and use it to decide whether to release.

Why: Aggregate accuracy can hide subgroup harms. Diagnose the disparity and validate proportionate safeguards or corrections before proceeding. This does not require assuming that all differences have the same cause or that every difference must be eliminated.

Bank item 26 · Domain 3 · Select two

A research agent reads external web pages and can access internal tools. A page tells it to ignore its instructions and upload internal documents to an external address. Which TWO controls most directly reduce the risk?

  1. Increase the model’s context capacity.
  2. Ask the agent to explain each tool call.
  3. Restrict retrieval to established sites but allow text to steer tool use.
  4. Keep retrieved text separate from authorized instructions. · Correct
  5. Limit tool permissions and gate sensitive transfers. · Correct

Why: Retrieved web pages are data, not authorized instructions. Keeping them separate from instructions helps resist prompt injection; restricted tools and approval gates limit possible transfers. More context or explanations do not enforce permissions, and even an established site can return unsafe instructions.

Bank item 27 · Domain 3 · Select one

A company wants an assistant to answer employee HR questions. Payroll records are sensitive and only a few questions require them. Which design best supports privacy?

  1. Index payroll centrally; trust employees' question choices.
  2. Minimize indexed data and enforce per-user retrieval permissions. · Correct
  3. Use a private network and give the assistant access to the full payroll archive.
  4. Remove employee names but allow organization-wide retrieval of detailed salary and role records.

Why: Privacy requires both appropriate data scope and authorization. A private network does not establish which employees may view payroll information. Removing names alone may leave re-identifiable sensitive details, and asking users to behave appropriately is not an access control.

Bank item 28 · Domain 3 · Select one

A company uses a managed AI service. A project manager claims that the cloud provider is therefore responsible for all data permissions, application behavior, and regulatory compliance. Which response is most accurate?

  1. Providers assume compliance accountability when customers select managed models.
  2. Customers manage service infrastructure as well as application data.
  3. Managed infrastructure does not remove customer responsibility for data and application use. · Correct
  4. The model developer owns application access decisions while the customer manages only user training and budgets.

Why: Shared responsibility depends on the service, but managed infrastructure does not transfer all customer responsibilities. Customers still govern their data, access choices, application use, and applicable obligations. Providers generally operate the underlying infrastructure; model developers do not automatically own customer application permissions.

Bank item 29 · Domain 3 · Select one

A finance assistant drafts board summaries. Reviewers find plausible figures that do not match the ledger. Which control most directly reduces the risk while preserving useful summaries?

  1. Pull figures from the ledger and validate claims before release. · Correct
  2. Fine-tune on approved reports to improve the accuracy of typical figures.
  3. Reconcile figures across independent generations and keep values that recur most often.
  4. Require citations for figures and verify that each cited document exists.

Why: The ledger is authoritative. Retrieving its figures and checking numerical claims directly addresses mismatch. Fine-tuning can reproduce old figures, agreement among generated answers can repeat the same error, and citations do not prove that cited numbers were accurately reproduced or calculated.

Bank item 30 · Domain 3 · Select one

An insurance assistant explains eligibility under a policy with explicit decision rules. It sometimes cites the correct clause but draws a conclusion that contradicts those rules. Which validation control most directly catches this failure before release?

  1. Verify that each cited clause exists in the current policy document.
  2. Screen generated responses for sensitive data and disallowed topics.
  3. Compare the explanation’s wording with previously approved customer responses.
  4. Check the proposed conclusion against the formal eligibility rules. · Correct

Why: The issue is rule consistency: an answer can quote the right clause and still reach the wrong conclusion. A rule check validates the conclusion; citation, content filtering, and wording comparisons address different concerns.

Bank item 31 · Domain 3 · Select two

A marketing agency plans to use a vendor's generative image service for commercial campaigns. Which TWO due-diligence actions most directly address intellectual property and contractual risk?

  1. Confirm the provider offers private hosting.
  2. Review commercial-use rights and contractual responsibility for claims. · Correct
  3. Screen proposed artwork for infringement and brand risk before publication. · Correct
  4. Check supported formats and turnaround times.
  5. Retain provenance records to document how each published asset was created.

Why: Contract review clarifies commercial-use rights and who bears responsibility for claims. Publication review addresses the risks of the actual image and its intended use. Private hosting, production fit, and provenance can help with other decisions, but do not establish those rights or clear an image for publication.

Bank item 32 · Domain 3 · Select one

An agent can change supplier bank details. A mistaken action could redirect large payments. What is the most appropriate initial control model?

  1. Permit small changes and review account updates afterward.
  2. Authorize changes using model confidence and notify the finance team after each update.
  3. Require approval, limit permissions, and retain audit and recovery controls. · Correct
  4. Use two agents to agree on changes before automatically updating the supplier payment details.

Why: Bank-detail changes can redirect payments, so constrained permissions, independent authorized approval, audit, and recovery are appropriate. Payment-size thresholds can miss later consequences of a changed account. Model confidence or agreement between agents does not replace authorization.

Bank item 33 · Domain 3 · Select one

A cross-functional AI council reviews proposals, but no one owns decisions or accepts residual risk. Projects stall and approvals are inconsistent. Which change should leaders make?

  1. Assign decision rights and residual-risk acceptance authority. · Correct
  2. Require council unanimity for every proposal.
  3. Let delivery leads waive compliance concerns.
  4. Ask the supplier to provide a standard approval template for the council to follow.

Why: The gap is ownership of decisions and accepted risk. Explicit authority, decision rights, and escalation resolve it. Unanimity can increase delay without clarifying accountability; templates provide documentation but not authority; a delivery lead cannot assume compliance waiver powers by default.

Bank item 34 · Domain 3 · Select two

After a model update, an AI assistant starts giving materially incorrect customer advice. Which TWO actions belong in the organization's response?

  1. Increase alert sampling while the updated model continues serving all customers.
  2. Compare provider benchmarks with reports before deciding whether to escalate the incident.
  3. Correct reported responses and keep the update active for the rest of the audience.
  4. Contain the incident through escalation and a safe fallback. · Correct
  5. Investigate impact and verify the fix before restoring service. · Correct

Why: Observed materially incorrect customer advice warrants timely containment and investigation. Recovery should follow validated correction. Benchmarks do not override observed failures, waiting can prolong harm, and correcting individual answers without investigating scope can leave the cause active.

Bank item 35 · Domain 3 · Select one

A benefits agency uses AI recommendations in decisions affecting applicants. Applicants need to understand adverse decisions and challenge mistakes. Which requirement best supports this need?

  1. Publish the model inputs, validation results, and overall accuracy for all applicants.
  2. Give each applicant an understandable reason for the decision and a clear route to request human review. · Correct
  3. Explain model confidence and provide a summary of the training and validation process.
  4. Publish the technical documentation and direct applicants to the model supplier for details.

Why: Affected applicants need usable reasons for their decision and a way to challenge errors through accountable review. General inputs, accuracy, confidence, or technical documentation can support transparency but do not themselves provide individual explanations and recourse.

Bank item 36 · Domain 4 · Select one

A company has modern cloud infrastructure and strong data scientists, but leaders disagree about objectives and no one owns AI risk. A board member says it is ready because the technology is ready. What is the best response?

  1. Validate infrastructure at expected usage levels.
  2. Address compute and integration gaps first, then assign business objectives and risk ownership during rollout.
  3. Use the data science capability assessment as the primary readiness gate for the proposed enterprise program.
  4. Align leaders on outcomes and assign accountable risk ownership before launch. · Correct

Why: Readiness includes objectives, leadership, people, processes, data, and governance. The stated disagreement and missing risk ownership are critical organizational issues. Capacity and data science strength are necessary in some settings but cannot substitute for alignment and accountability.

Bank item 37 · Domain 4 · Select one

Regional units use different customer identifiers and disagree about who can authorize data sharing. An enterprise AI pilot cannot obtain a consistent customer view. Which intervention is most appropriate?

  1. Centralize copies; let models reconcile regional identifiers.
  2. Choose one region as the default standard and require other regions to adopt its identifiers immediately.
  3. Assign data ownership, common definitions, and authorized sharing processes. · Correct
  4. Keep each regional dataset separate and combine model answers without resolving the inconsistent customer identifiers.

Why: Common meaning, accountable owners, and approved access are prerequisites for a consistent customer view. Moving copies does not settle definitions or permissions. Imposing one region standard without agreement and combining inconsistent outputs can preserve the underlying conflict.

Bank item 38 · Domain 4 · Select two

Claims staff are worried that an AI rollout will undermine their roles. A technically successful pilot has low voluntary usage. Which TWO actions best support adoption?

  1. Involve staff in workflow design and discuss expected role changes. · Correct
  2. Publish forecast productivity gains before asking for feedback.
  3. Reward account activation and use it as the primary measure of workforce readiness.
  4. Offer a common vendor introduction to all roles.
  5. Offer practical training, feedback channels, and credible peer support. · Correct

Why: Participation, honest communication, task-relevant training, feedback, and local support address both trust and capability. Productivity forecasts and activation incentives do not resolve concerns about work design. A generic presentation may inform staff but is insufficient as practical adoption support.

Bank item 39 · Domain 4 · Select one

A team proposes a six-week pilot but has not defined success, maximum acceptable cost, or what happens at the end. What should the sponsor require before it begins?

  1. Commit to scaling after reaching technical accuracy.
  2. Define outcome measures, limits, ownership, and a decision date. · Correct
  3. Purchase enterprise licenses ahead of pilot completion.
  4. Select the production platform and finalize all enterprise integrations before testing the business hypothesis.

Why: The sponsor needs an actionable experiment: measurable outcomes, cost and risk limits, an owner, and a decision date. Accuracy alone does not define business success. Enterprise purchases and fully completed integrations can create commitments before value has been tested.

Bank item 40 · Domain 4 · Select one

A proof of concept performs well on a curated dataset. The team now plans to serve a real business unit. Which requirement most clearly distinguishes production launch from experimentation?

  1. Broaden the demonstration and validate it with examples selected by business stakeholders.
  2. Test how quickly the platform can onboard the planned number of new users.
  3. Approve production on the basis of the pilot report and early feedback.
  4. Assign service ownership and establish monitoring, support, incident escalation, and a tested fallback before production begins. · Correct

Why: Moving to production creates service obligations and requires accountable operations, controls, monitoring, and recovery. More demonstration examples improve evidence but do not supply operational arrangements. Ownership and support should be ready before real business dependency grows.

Bank item 41 · Domain 4 · Select one

Several departments have useful AI pilots but duplicate evaluations and use incompatible governance templates. Leaders want to scale while keeping departments accountable for their business outcomes. What should a center of excellence primarily do?

  1. Provide shared standards and reusable tools while departments own their outcomes. · Correct
  2. Move business-case ownership into the center and let it prioritize departmental outcomes.
  3. Keep each department’s evaluations separate until all deployments are complete.
  4. Standardize every department on one model and workflow regardless of its use case.

Why: A center of excellence enables teams with common patterns, standards, and advice. Departments still own their use cases and outcomes. Centralizing delivery or imposing one workflow can create bottlenecks or poor local fit.

Bank item 42 · Domain 4 · Select two

A contact-center assistant succeeds in one country. Leaders want deployment in five more countries with different languages, workflows, and requirements. Which TWO actions best support responsible scaling?

  1. Apply global thresholds with local translation checks.
  2. Finish all regional deployments before gathering enough data to compare local quality and adoption.
  3. Validate local data, performance, workflows, and applicable requirements. · Correct
  4. Stage deployment with local owners, monitoring, and rollback criteria. · Correct
  5. Gate rollout on overall model accuracy.

Why: New regions introduce different languages, workflows, data, and obligations. Local validation plus staged rollout and local ownership address these differences. Translation checks and global averages can conceal local problems; waiting until every region is deployed delays correction.

Bank item 43 · Domain 4 · Select one

An organization offers the same advanced model-building course to executives, customer-service staff, and data scientists. Completion is low and employees cannot apply the content. What should the learning lead change?

  1. Require advanced-course completion before granting access.
  2. Provide role-based learning with practical responsible-use exercises. · Correct
  3. Replace the course with optional vendor demonstrations covering the latest available model features.
  4. Keep the course unchanged but reward attendance to increase the percentage of employees completing it.

Why: Training should match each role and its real tasks. Practical exercises and responsible-use guidance build applicable capability. Attendance incentives do not fix content mismatch; optional demonstrations are not a complete development program.

Bank item 44 · Domain 4 · Select one

An enthusiastic practitioner runs AI workshops, but cross-department funding and policy conflicts remain unresolved. Which leadership action is most appropriate?

  1. Expand workshops; let departments negotiate funding.
  2. Generate further evidence through additional prototypes.
  3. Appoint a supplier relationship manager to coordinate decisions across internal departments without changing decision authority.
  4. Assign an executive sponsor with cross-department authority. · Correct

Why: The unresolved issues require organizational authority over priorities, resources, and conflicts. A credible practitioner supports peers but does not necessarily have that authority. More demonstrations or coordination without decision rights does not close the leadership gap.

Bank item 45 · Domain 4 · Select one

A service leader wants to retain useful AI drafts to save time, but every response to a distressed customer must receive human review before sending. The assistant may misread emotion or unusual circumstances. Which workflow meets both requirements?

  1. Automatically send familiar replies and sample them later.
  2. Require staff to rewrite drafts for every sensitive reply.
  3. Use drafts while staff review and own sensitive replies. · Correct
  4. Approve responses automatically when model confidence passes a threshold.

Why: AI can produce a usable first draft, while staff apply empathy, verify context, and approve sensitive replies before sending. Post-send sampling is too late for these cases; mandatory rewrites give up useful drafting, and model confidence does not replace required review.

Bank item 46 · Domain 4 · Select one

Six months after launch, an assistant has many registered users but managers cannot tell whether it improves work. What measurement change is most useful?

  1. Compare adoption, output quality, and business outcomes against the pre-launch baseline. · Correct
  2. Track registrations and repeat visits.
  3. Measure generated responses and total tokens as direct evidence that employees are completing more valuable work.
  4. Use department-level satisfaction surveys as the sole indicator of realized operational and financial value.

Why: Business value needs evidence of useful adoption, outcomes, quality, and comparison with the original baseline. Accounts, sessions, and request volumes measure activity. Satisfaction is useful but cannot alone establish productivity or financial effects.

Bank item 47 · Domain 1 · Select one

A media company evaluates two systems. System X learns to classify photographs by subject. System Y drafts a new caption for each photograph. Which classification BEST describes these systems?

  1. Both classify inputs: X labels images and Y assigns a caption label.
  2. Both generate content: X writes a subject description and Y writes a caption.
  3. X classifies images by subject; Y generates new captions. · Correct
  4. X generates captions and Y assigns images to existing categories.

Why: Classification assigns an existing label; generating a new caption creates content. Both are within AI.

Bank item 48 · Domain 1 · Select one

A manufacturer wants to forecast spare-part demand from dated sales tables and summarize technicians' free-text repair notes. Which data assessment BEST supports these two use cases?

  1. Forecast from the sales tables and summarize the free-text notes. · Correct
  2. Use repair notes to forecast demand and sales tables to summarize repairs.
  3. Use sales tables for both forecasting demand and describing repair work.
  4. Use repair notes for both tasks and exclude the dated sales history.

Why: Defined sales fields support forecasting, while free-text notes are unstructured content suited to summarization.

Bank item 49 · Domain 1 · Select one

Two business units use different definitions for model, training, and inference, so their AI proposals cannot be compared. Leadership wants a common framework for describing ML-based AI systems. Which reference BEST fits this need?

  1. ISO/IEC 27001.
  2. ISO 9001.
  3. ISO/IEC 23053. · Correct
  4. ISO/IEC 42001.

Why: ISO/IEC 23053 provides a common framework for describing AI systems that use machine learning. ISO/IEC 42001 concerns AI management systems, ISO/IEC 27001 concerns information security management, and ISO 9001 concerns quality management.

Bank item 50 · Domain 1 · Select one

A warranty model remains available and responds quickly, but its prediction error has risen over six months as product designs change. Which response BEST addresses this problem?

  1. Check drift against the baseline and evaluate retraining. · Correct
  2. Tune alert thresholds without checking outcomes for the redesigned products.
  3. Expand compute capacity and recheck response times under peak load.
  4. Use the original holdout set to confirm that launch accuracy is reproducible.

Why: Availability does not establish prediction quality. Changing product relationships require performance monitoring and an appropriate update.

Bank item 51 · Domain 1 · Select one

An employee wants to use a new AI writing tool for public drafts, but the tool has not completed security review. The company permits documented, time-limited exceptions only after a risk owner signs off. What should the AI governance team do?

  1. Record it as under evaluation and require sign-off for any public-data exception. · Correct
  2. Approve public use by default while the vendor review is still open.
  3. Apply the existing approval for a different writing tool to this one.
  4. Block the tool indefinitely without an owner or route to review.

Why: Classification makes it clear that the tool is still under evaluation. Any public-data use should follow the documented exception process and the risk owner’s approval. Neither another tool’s approval nor an informal default permits this tool.

Bank item 52 · Domain 1 · Select one

A team changes the role description, examples, document sources, and output instructions simultaneously. Results improve, but the team cannot identify why. Which evaluation approach BEST supports a reusable prompt?

  1. Change several prompt components together and compare against last week’s responses.
  2. Change one variable at a time on a fixed evaluation set. · Correct
  3. Choose examples from the improved outputs and test only those cases.
  4. Compare new prompts against different tasks to broaden the evaluation pool.

Why: Deliberate comparisons help establish which prompt change improves the required behavior.

Bank item 53 · Domain 2 · Select one

A software company wants AI investment to reduce the time required to release tested changes. Which proposal MOST clearly connects a capability to that business outcome?

  1. Pilot automatic release notes and track how often developers publish them.
  2. Pilot a larger model and track how much test code it generates.
  3. Pilot AI test drafting; track review effort, defects, and release lead time. · Correct
  4. Pilot faster code generation and track completion time for individual coding tasks.

Why: This names a concrete capability, its mechanism, and measures tied to delivery speed and quality.

Bank item 54 · Domain 2 · Select one

An AI pilot produces strong savings, but its budget ends at launch and no team can operate or monitor it. The funding gap is fixable in the next planning cycle. Which portfolio decision is MOST defensible?

  1. Run more pilots to confirm savings before discussing who will operate them.
  2. End the initiative and move its remaining funds to an operational programme.
  3. Pause expansion until ownership and operating funds are secured. · Correct
  4. Expand to more teams so savings can fund future support and monitoring.

Why: Value alone is insufficient; a sustainable initiative must be operated and funded across its life.

Bank item 55 · Domain 2 · Select one

An AI support pilot reduces handling costs, but customers report more repeat contacts. Which KPI approach BEST assesses whether the pilot creates business value?

  1. Compare cost per interaction, handling time, and satisfaction for completed chats.
  2. Track answer quality, response volume, and average time to first reply.
  3. Compare operating spend, queue length, and satisfaction across all support channels.
  4. Track cost per resolved case, repeat contacts, resolution quality, and customer satisfaction across the pilot period. · Correct

Why: A cost reduction can conceal a poorer experience; both financial and nonfinancial outcomes should be visible.

Bank item 56 · Domain 2 · Select one

A generative AI assistant receives more requests than expected. Each request includes long irrelevant histories, and costs rise without improving answer quality. Which action BEST targets the immediate cost factor?

  1. Buy capacity for peak usage.
  2. Switch to a cheaper model without first checking which inputs drive costs.
  3. Trim irrelevant context and verify cost and quality changes. · Correct
  4. Limit requests per user while keeping the same oversized input payload.

Why: Consumption depends on usage and text processed. Removing unnecessary input can reduce cost while checking that value is preserved.

Bank item 57 · Domain 2 · Select one

A regional bank learns that competitors now provide instant routine service answers, while its customers wait until the next business day. What is the MOST useful initial strategic assessment?

  1. Estimate competitor AI spending and compare it with the proposed investment.
  2. Identify the customer expectations competitors are setting and prioritize the service gaps most worth closing. · Correct
  3. Compare competitor feature lists and prioritise parity in the most visible categories.
  4. Assess the capacity of the internal team to build and operate a service model.

Why: Competitive analysis connects rivals' capabilities to customer value and helps select an appropriate response.

Bank item 58 · Domain 2 · Select one

An industrial company uses an AI failure-prediction tool internally. It proposes selling monitored uptime as a subscription instead of selling only equipment. Which statement BEST characterizes the proposal?

  1. It improves the reliability of the existing equipment sales process.
  2. It changes the customer offering and how revenue is earned. · Correct
  3. It extends after-sales support while keeping the existing sales-only model.
  4. It reduces the unit cost of manufacturing and maintaining equipment.

Why: A subscription outcome-based offering changes the business model and revenue mechanism.

Bank item 59 · Domain 2 · Select one

AI use is still emerging in an industry. A company has two promising but untested opportunities and a limited budget. Which investment posture is MOST appropriate?

  1. Fund both initiatives now and compare outcomes after enterprise deployment.
  2. Wait until peer adoption stabilizes before investing in either opportunity.
  3. Fund bounded tests and expand when the evidence warrants it. · Correct
  4. Split the budget equally and use the same rollout targets for both initiatives.

Why: Investment should reflect industry maturity, competitive dynamics, readiness, and evidence of value.

Bank item 60 · Domain 2 · Select one

A team values 2,000 hours saved at $40 per hour and then adds the same $80,000 again as labor cost savings. No payroll has changed, and no additional benefit is demonstrated. What is the MOST defensible correction?

  1. Report the released capacity and payroll reduction as separate monetary benefits.
  2. Count the capacity once; report cash savings only when demonstrated. · Correct
  3. Replace the estimate with the full revenue gain over the same period.
  4. Remove the time estimate from the business case until headcount is reduced.

Why: Valuing the same saved hours twice overstates benefit. Capacity and cash outcomes need honest attribution and distinct evidence.

Bank item 61 · Domain 2 · Select one

A recommendation service learns from customer responses, creating an accumulating data asset. The same loop begins reducing exposure of less common products. Which strategic response BEST preserves value while managing risk?

  1. Track the learning advantage and intervene when exposure skews harmfully. · Correct
  2. Keep learning from clicks and judge the system solely by conversion gains.
  3. Freeze the feedback loop and keep the original ranking model in production.
  4. Retrain more frequently on observed clicks while leaving exposure unchanged.

Why: A feedback loop can support differentiation and also amplify skew. Its benefit does not remove the need for monitoring and appropriate controls.

Bank item 62 · Domain 2 · Select one

A company compares three AI offerings: one charges for input and output text, one charges for provisioned compute time, and one charges per licensed user. Which comparison is MOST useful?

  1. Compare each provider’s quoted unit price before estimating expected demand.
  2. Model each usage driver over a common period and compare total value. · Correct
  3. Annualize first-month spending and assume the mix of users and requests stays constant.
  4. Use a common per-seat estimate derived from current team size for all three offerings.

Why: Estimate input and output text usage, compute hours, and licensed seats for the same workload and time horizon. Compare total costs alongside the expected outcome. The quoted unit prices use different denominators.

Bank item 63 · Domain 2 · Select one

A specialist insurer needs an AI capability within four months. The team lacks the relevant domain expertise, and a provider has established industry workflows and evidence of suitable controls. Which acquisition approach is MOST appropriate to evaluate?

  1. Build internally with temporary staffing and develop the needed industry knowledge in parallel.
  2. Evaluate the specialist partnership against requirements, controls, and cost. · Correct
  3. Buy a general-purpose product and resolve regulatory workflow differences after deployment.
  4. Compare vendors by their quoted licence fees and select the lowest bid.

Why: Specialized expertise, time constraints, and control requirements can favor partnership, subject to due diligence.

Bank item 64 · Domain 3 · Select one

A company clearly discloses that an assistant uses AI. However, the assistant performs poorly when users submit unusual abbreviations or slightly malformed requests. Which responsible AI dimension is MOST directly implicated?

  1. Veracity and robustness when inputs vary. · Correct
  2. Transparency about when AI is used.
  3. Explainability of individual system responses.
  4. Privacy in handling personal input data.

Why: The concern is reliable, correct behavior for unexpected inputs. Disclosure addresses transparency but does not establish robustness.

Bank item 65 · Domain 3 · Select one

A team is selecting data sources and deciding which customer actions an AI agent may perform. No code has been built. What is the MOST appropriate governance action now?

  1. Define permitted data, actions, and oversight with the accountable stakeholders now. · Correct
  2. Let engineering set the initial scope and review it once the prototype is complete.
  3. Use the vendor’s default permissions while business owners prepare a later policy.
  4. Approve data sources now and defer agent authority decisions until launch.

Why: These early choices shape risk. Governance can influence the design before changes become expensive.

Bank item 66 · Domain 3 · Select one

A production AI dashboard tracks quality daily, but no one has defined an unacceptable value or who should respond. Which addition MOST directly turns the dashboard into an effective control?

  1. Set quality thresholds, an alert owner, and response steps. · Correct
  2. Publish the dashboard more frequently.
  3. Add a summary quality score to the existing quarterly governance pack.
  4. Collect user complaints in a shared queue and summarize trends monthly.

Why: A control needs measurement, a trigger for action, and a person or team able to act.

Bank item 67 · Domain 3 · Select one

A marketing team plans to publish AI-generated artwork in a paid campaign. It assumes every generated image is automatically owned and unrestricted. What is the MOST appropriate governance response?

  1. Review service terms and publication risks with legal counsel. · Correct
  2. Rely on approval for internal use and check the artwork against brand guidelines.
  3. Require provenance records and treat those records as proof of publication rights.
  4. Check the license for the design tool and assume it covers each generated asset.

Why: Internal approval, provenance and design-tool licensing do not establish whether a specific output is suitable for a commercial campaign. Review the service terms and the proposed use with the appropriate rights and legal owners.

Bank item 68 · Domain 3 · Select one

A company defines employee potential solely as past promotion speed, even though promotion opportunities historically differed across groups. Where does the described risk of bias first enter?

  1. At data collection, if examples are missing from one group.
  2. At evaluation, if aggregate accuracy masks subgroup errors.
  3. At deployment, if staff apply scores inconsistently across teams.
  4. At problem framing, when promotion speed becomes the target. · Correct

Why: Selecting what to predict embeds a business judgment; a proxy can import unequal opportunity into the objective itself.

Bank item 69 · Domain 3 · Select one

A team proposes releasing financial notices automatically whenever a model reports high confidence. Testing shows that some high-confidence responses contain material errors. Which safeguard is MOST appropriate before relying on this routing approach?

  1. Route by confidence after increasing the threshold on the existing scale.
  2. Validate confidence-based routing and independently check notices that cannot be routed safely. · Correct
  3. Allow high-confidence release with retrospective review of a weekly sample.
  4. Use a customer-facing disclaimer and let recipients flag inaccurate notices.

Why: Testing shows that high confidence is not a reliable proxy for correctness here. Validate confidence against observed errors and independently check or review notices where safe routing cannot be demonstrated. Raising the threshold alone does not fix unreliable confidence; retrospective review and disclaimers do not prevent an erroneous notice.

Bank item 70 · Domain 4 · Select one

A company has reliable AI in production in its finance function, while research teams are still running isolated proofs of concept. How should the strategist assess maturity?

  1. Rate the company by its most advanced deployment and apply that rating to all units.
  2. Rate the company by its least mature unit before approving any further work.
  3. Assign a common maturity stage from next year’s enterprise adoption target.
  4. Assess each unit against its evidence of deployment and operation. · Correct

Why: Business units can be at different maturity stages. Production experience in finance is evidence about that unit, while research proofs of concept show experimentation rather than sustained operations.

Bank item 71 · Domain 4 · Select one

A company's roadmap funds advanced personalization next quarter, but its customer identifiers cannot be reconciled across systems. The personalization use case requires a combined customer view. What investment should be prioritized?

  1. Add more behaviour events before checking how customers match across source systems.
  2. Resolve customer identity matching across the required sources. · Correct
  3. Standardize the personalisation interface while retaining separate customer records.
  4. Fund a higher-capacity model to infer profiles without validated customer joins.

Why: Prerequisite sequencing funds the foundation that makes the strategic next step feasible.

Bank item 72 · Domain 4 · Select one

Two datasets are individually accurate and accessible. One has weekly product totals and the other daily totals using different product codes. The business needs a combined product-level view. What should the team assess first?

  1. Whether additional weekly records would increase the coverage of the larger dataset.
  2. Whether the two sources should use the same text export format.
  3. Whether a higher-capacity model can infer a mapping from aggregate totals.
  4. Whether product codes and time intervals can be reconciled reliably. · Correct

Why: Accurate source records do not guarantee a valid join. Product identifiers and time grain must be reconciled before the team can trust a combined view. More rows, common file formats, or a larger model do not establish that mapping.

Bank item 73 · Domain 4 · Select one

A cross-company AI program has enthusiastic local champions, but department heads deny them time and disagree about the program's objectives. Which action BEST sustains momentum?

  1. Appoint champions in more teams and ask them to promote locally selected tools.
  2. Centralise communications while departments retain separate priorities and staffing decisions.
  3. Organise a cross-team showcase to compare departmental use cases and gather feedback.
  4. Align executive sponsors and give champions protected time, information, and a clear mandate. · Correct

Why: Local advocacy needs leadership support and practical empowerment to survive competing priorities.

Bank item 74 · Domain 4 · Select one

An organization plans AI-assisted invoice review. Employees want to know whether they will still approve exceptions. Which communication BEST addresses their concern?

  1. Share rollout dates and affected teams.
  2. Explain when the change starts, which invoice checks AI performs, and who retains authority to approve exceptions. · Correct
  3. Publish accuracy results and invite staff to submit questions after launch.
  4. Explain expected productivity gains and provide a general software training session.

Why: Transparent communication connects the implementation to concrete changes in work and responsibilities.

Bank item 75 · Domain 4 · Select one

Leadership must choose a first AI project. One is a small isolated demonstration with no reusable data or workflow pattern. Another is a bounded use case that can teach patterns needed by several departments. Which is the stronger starting point?

  1. Start with the isolated demo because it is quicker to deliver and communicate.
  2. Choose a broad rollout to gather more feedback across departments.
  3. Choose a bounded use case with reusable patterns and measurable value. · Correct
  4. Run parallel demos and compare how many teams express interest.

Why: A useful early win builds credibility and establishes patterns that support later expansion.

Bank item 76 · Domain 4 · Select one

A company will expand AI order routing to all branches. The existing manual process is still reliable. Which rollout plan BEST protects business continuity?

  1. Cut over by region after pilot acceptance, then restore manual work if needed.
  2. Stage rollout with validated capacity, a tested fallback, and rollback triggers. · Correct
  3. Run both workflows briefly, then retire the manual one when adoption reaches target.
  4. Stagger by branch size and review afterward.

Why: The strongest plan checks capacity before expansion and retains a tested fallback with explicit rollback triggers. Briefly running both workflows can help, but adoption alone does not establish whether it is safe to retire the manual route.

Bank item 77 · Domain 4 · Select one

An AI assistant works well for 100 users. It will soon serve 20,000 users and depends on a shared document API. Which validation BEST addresses scaling risk?

  1. Measure model latency at expected peak traffic with document responses supplied by a mock API.
  2. Benchmark shared API throughput under today’s workload and extrapolate capacity for the new assistant.
  3. Load-test the assistant using cached documents and measure response times and cache hit rates.
  4. Load-test the model and shared API together at expected peak traffic, including how the assistant handles retrieval failures. · Correct

Why: The whole workflow must handle scale; an upstream bottleneck can impair the service even when the model itself is sound.

Bank item 78 · Domain 1 · Select two

A procurement agent must gather quotes and prepare draft orders. Company policy says only human buyers can commit funds. Which TWO controls separate the agent’s permissions from the order release workflow?

  1. Let the agent submit orders below the usual manual approval threshold.
  2. Keep the agent on a buyer’s account and log all purchase activity.
  3. Permit order submission once a second model agrees the price is reasonable.
  4. Give the agent read access to quotes and draft-only order permissions. · Correct
  5. Require an authenticated buyer to review and release each draft. · Correct

Why: Tool permissions restrict the agent to preparation. A separate human release step enforces the purchasing policy on each order. Logs and model agreement do not substitute for either boundary; the stated policy does not delegate low-value purchasing.

Bank item 79 · Domain 1 · Select two

A policy assistant often cites the wrong policy because its prompt includes every historical version. The company needs current, verifiable answers without retraining whenever a policy changes. Which TWO changes best address the problem?

  1. Retrieve only relevant passages from effective policy versions. · Correct
  2. Fine-tune on policy citation examples.
  3. Cite retrieved evidence and defer when it is insufficient. · Correct
  4. Rerank retrieved passages by publication date without checking which version is effective.
  5. Expand the input context window.

Why: Filtering to effective policy versions keeps retrieval current. Citing evidence and deferring when it is insufficient makes the answer verifiable. A larger context window leaves obsolete versions in play, fine-tuning on citation examples does not keep policy facts current, and publication date is not the same as effective date.

Bank item 80 · Domain 2 · Select two

A company launches an AI support assistant during a seasonal lull. Average handling time falls, and leadership wants to know whether AI caused the improvement. Which TWO actions make the case more credible?

  1. Check case mix and staffing changes before attributing gains. · Correct
  2. Compare with a similar untreated queue over the same period. · Correct
  3. Compare outcomes with the annual target.
  4. Compare frequent and infrequent users.
  5. Attribute savings to the tool based on adoption and satisfaction during the campaign.

Why: A contemporaneous comparison helps account for shared changes, and checking case mix and measurement guards against false attribution. Activity and timing alone do not establish causal impact.

Bank item 81 · Domain 2 · Select two

An AI document-review pilot is affordable, but the company expects traffic and exception handling to grow sharply in production. Which TWO actions best inform the scale decision?

  1. Estimate unit costs across volumes, including exception review. · Correct
  2. Forecast licence spend under vendor rates.
  3. Extrapolate pilot review time without testing how exception rates change at scale.
  4. Set cost and quality gates for staged rollout. · Correct
  5. Prioritize rollout by department volume.

Why: Scenario-based unit economics expose costs hidden by pilot volumes. Paired cost and quality gates prevent expansion from being driven by throughput alone.

Bank item 82 · Domain 3 · Select two

A customer assistant invents a contractual exception when the retrieved documents are incomplete. The company needs to reduce repeat incidents while preserving useful automated responses. Which TWO controls best fit?

  1. Review unsupported-answer incidents and fix recurring gaps. · Correct
  2. Defer or escalate when evidence is insufficient. · Correct
  3. Add a disclaimer and sample response accuracy during the next regular review.
  4. Expand the knowledge base without checking whether responses are supported by retrieved text.
  5. Use topic filters to restrict requests to contracts already in the system.

Why: Deferring or escalating unsupported answers can prevent an immediate error. Logging and reviewing incidents helps identify and repair recurring gaps. Disclaimers, a bigger document collection, and topic boundaries do not verify the evidence for a specific answer.

Bank item 83 · Domain 3 · Select two

At launch, a claims triage model met the company’s subgroup error thresholds. After expansion to a new region, overall performance is stable but one group’s false-escalation rate has breached its threshold. Which TWO responses best fit the monitoring plan?

  1. Recalibrate the overall threshold using all regions.
  2. Compare adoption across the regions.
  3. Escalate the breach and limit affected routing while investigating. · Correct
  4. Compare subgroup outcomes and input changes by region. · Correct
  5. Wait for the next scheduled review to see whether the aggregate trend changes.

Why: A threshold breach triggers accountable intervention. Limit the affected automated routing and compare regional input and outcome evidence to diagnose whether the change is due to data shift, operating differences, or another cause. Aggregate metrics can mask the affected group.

Bank item 84 · Domain 4 · Select two

A pilot works reliably while its original team is present, but operations will inherit it after launch. Which TWO deliverables most directly reduce handover risk?

  1. A revised business case that forecasts adoption and revenue across three operating scenarios.
  2. A communication plan that promotes the new workflow to users before launch.
  3. A named operational owner and tested escalation procedures. · Correct
  4. A load rehearsal covering failure and fallback. · Correct
  5. A rollout schedule that retires the manual process when participation reaches its target.

Why: Production handover needs explicit ownership and a tested response to realistic failures. Forecasts and communications can support launch, but they do not demonstrate readiness to maintain the service or recover when it fails.

Bank item 85 · Domain 4 · Select two

An assistant will draft summaries for frontline staff and recommend escalations. Staff worry they will be blamed for inaccurate suggestions. Which TWO actions support responsible adoption?

  1. Require completion of tool navigation training.
  2. Reward teams with consistent usage.
  3. Pair new users with experienced staff to demonstrate tool navigation.
  4. Clarify decision ownership and create a safe reporting route. · Correct
  5. Practise when to accept, override, or escalate suggestions. · Correct

Why: Clear decision rights and practice with real workflow choices build confidence and accountability. Usage metrics and mandates do not establish competence.