Weeks 1–24

Full Timed Final Mock

A standalone, integrated practice exam of multi-concept scenarios spanning the complete 24-week syllabus and testing practitioner judgement.

Time remaining90:00

A multilingual benefits-enquiry assistant uses enterprise RAG over policy and claimant documents, summarises uploaded evidence, and recommends a case routing that a caseworker can accept or change. Who most clearly counts as an 'affected person' in this scenario?

The claimant, even though they may never see the AI system directlyOnly people who log into the case management systemOnly the caseworker who clicks acceptOnly the vendor's engineering team

In the same scenario, the caseworker remains formally responsible for the final decision but always accepts the AI's routing without reviewing the underlying evidence. What governance risk does this illustrate?

Purpose limitation failureAutomation bias undermining meaningful human oversightDatabase rights infringementA GPAI transparency failure

The retrieval corpus for this assistant contains both official policy and claimant-submitted documents, and performance varies noticeably by language. What is the most appropriate immediate governance response before scaling?

Assume the variation will resolve itself with more usageIgnore the variation since overall accuracy is highTest and document performance separately by language and require a defined minimum standard before scaling to lower-performing languagesRestrict the tool to English-speaking claimants only, without disclosure

A supplier plans to update the underlying model during the pilot phase. What is the strongest contractual and governance safeguard against this introducing unassessed risk?

Terminating the contract immediately upon any mention of a model updateIgnoring model versioning entirely since the interface looks the sameTrusting that supplier updates only ever improve performanceA change-notification clause plus a requirement to revalidate acceptance criteria before the updated model is relied upon in production

A credit union's AI affordability tool recommends 'decline' for an applicant with no human review before the applicant is notified. Under UK GDPR Article 22A–22D, what must be assessed first?

Whether this is a solely automated decision producing a significant/legal effect, which would trigger mandatory safeguardsWhether the applicant consented to receive marketing emailsWhether the model was trained in the UKWhether the applicant has good credit history

A recruitment platform operating in the EU auto-rejects candidates below a configurable score threshold with no human review. Under the EU AI Act, this use case most likely falls into which category, and why?

Minimal risk, because recruitment tools are explicitly exemptedHigh-risk, because employment/recruitment decisions affecting access to work are an Annex III high-risk use caseProhibited practice in all circumstancesLimited-risk, triggering only a transparency notice

For that same recruitment platform, the company that built and markets it under its own name is best classified as which actor?

Distributor only, because it uses a third-party GPAI model for parts of the pipelineDeployer onlyProviderAffected person

A claims-triage insurer wants one proportionate AI governance operating model across financial, healthcare-adjacent and internal-copilot use cases. What is the most defensible design principle?

Govern only generative AI and exempt classic AI models entirelyDelegate all governance decisions to the AI vendorApply an identical checklist to every AI use case regardless of riskUse proportionate risk-tiering with common intake, approval gates and monitoring, layered with sector-specific reviews (e.g. fair lending, patient safety, data leakage) where relevant

During due diligence, a vendor cannot confirm whether customer prompts are used to further train its model, nor its data retention period. What should this gap trigger?

A documented risk in the vendor review, likely requiring contractual clarification (training use, retention, confidentiality) before proceedingNo action, because vendor practices are outside the organisation's controlAutomatic termination of all vendor relationshipsImmediate approval, since most vendors behave responsibly

A model card for a deployed fraud-triage model is six months old and does not reflect a recent supplier-side update. An assurance reviewer is asked whether the system is ready to scale from 2 to 12 sites. What is the most defensible conclusion?

Reasonable assurance, because the model card existsLimited assurance — the evidence base is stale and does not cover the current model version, so scale-up should wait for updated testing evidenceNo conclusion possible, so proceed as originally plannedFull assurance, since uptime metrics look healthy

A HR chatbot occasionally invents a policy that does not exist in the source handbook when asked an edge-case question. What is the most direct control to reduce this specific risk?

Allowing the model unrestricted internet accessIncreasing the model's temperature settingGrounding answers in retrieved source documents with citations, and validating that the answer is actually supported by the cited sourceRemoving all logging to reduce latency

A public housing authority's AI tool prioritises repair requests using tenant descriptions and uploaded images. Analysis shows the tool may infer urgency partly from writing style, which correlates with English proficiency. What is the most appropriate next step?

Remove all image uploads to fix the issue immediately without further analysisConclude there is no issue because the model's average accuracy is highLaunch as planned since the tool is not explicitly using protected characteristicsInvestigate whether this creates a proxy for a protected characteristic and test for disparate impact across tenant groups before scaling

An internal audit team wants to confirm that a claimed monthly control — 'a reviewer checks a sample of AI-assisted decisions' — actually happened for the last quarter. Which evidence is strongest?

Reperformance or inspection of the actual review records/logs for a sample of months, corroborating the claimThe AI vendor's assurance that the control is unnecessaryA verbal assurance in a meeting with no supporting documentationAn email from the reviewer stating they always check decisions

A generative copilot used internally by customer-service staff is found to have received pasted customer personal data directly into free-text prompts. Which is the most complete governance response?

Ban all copilot use permanently without further analysisAssess the privacy risk and data flow, apply input controls/guidance to prevent sensitive data entry, and confirm the copilot provider's data handling and retention termsTake no action since staff are trustedOnly inform the marketing team

A model is measured with 91% overall accuracy but a 3x higher false-decline rate for one age group in a lending use case. Applying the risk mitigation hierarchy, what should be evaluated before considering 'accept'?

Whether to blame the affected age group for the model's error rateNothing further; 91% accuracy is high enough to accept as-isWhether the disparate impact can be avoided or reduced through design, data or threshold changes, or shared/transferred, before falling back to acceptance with monitoringWhether to hide the subgroup result from the release evidence

An AI agent embedded in a finance workflow can initiate payments above a threshold without further approval. A red-team test shows a crafted document can manipulate the agent into approving a fraudulent payment. What combination of controls is most appropriate?

Remove all logging to avoid revealing the vulnerabilityGive the agent broader permissions so it can self-correctRely solely on the model's built-in safety trainingA mandatory human approval gate for payments above a defined threshold, tool permission scoping, and monitoring/alerting for anomalous agent actions

A DPIA screening for a new AI-driven eligibility tool concludes the processing is likely high risk due to profiling of a large-scale, vulnerable population. What is the correct next step?

Conduct a full DPIA before processing begins, documenting risks, safeguards and consultation as neededSkip the DPIA if the legal team is too busyOnly complete a DPIA after the tool has already launchedProceed without further analysis since screening is just a formality

Which statement best reflects how ISO/IEC 42001, NIST AI RMF and the EU AI Act relate to each other in practice for an organisation operating internationally?

They are entirely redundant with each other and only one needs to be chosenISO/IEC 42001 can provide the management-system backbone, NIST AI RMF can inform risk-management practice, and the EU AI Act imposes specific binding legal obligations where it applies — together they can be aligned rather than duplicatedThe EU AI Act replaces the need for any management systemNIST AI RMF is a legal requirement in the EU

A vendor's contract is silent on what happens if the organisation needs to exit the relationship and migrate to another AI provider. What risk does this create?

This is solely a marketing concernNo risk, since exit terms are rarely relevantVendor lock-in risk, including possible loss of data portability, continuity gaps and reduced negotiating leverageThis only matters for open-source models

A model card states the system's intended use is 'general customer support' but the deployment team is using it for medical symptom triage. What is the most accurate governance conclusion?

Model cards are optional and can be disregarded once a system is liveNo new assessment is needed since accuracy has not been formally measured to be poorThis is acceptable because the model is technically capable of answering health questionsThis is an out-of-scope/prohibited use that requires a fresh risk and impact assessment specific to the medical triage context before continuing

Which best distinguishes an inherent risk from a residual risk in a risk register entry for an AI system?

Inherent risk is the risk level before controls are applied; residual risk is the level remaining after controls are appliedResidual risk is always zero once any control existsInherent risk only applies to cybersecurity risksThey are the same figure recorded twice

A company's AI incident response plan has never been tested. During a live incident where a chatbot leaked another customer's data, the team is unsure who has authority to trigger a kill switch. What does this reveal?

The incident plan is adequate because it exists in writingA gap between documented and operational readiness — authority, technical means and communication routes need to be tested, not just documentedKill switches are unnecessary for chatbot-based systemsThis is solely a technical engineering problem with no governance implication

A practitioner is asked in a viva: 'Why not just rely on the vendor's claim that their model is unbiased?' What is the strongest answer?

Because bias testing is impossible in principleVendor claims are always accurate and sufficientBecause accountability cannot be outsourced — the deploying organisation needs independent evidence, context-specific testing and contractual assurance rights, not just a supplier assertionBecause vendors are legally prohibited from making any claims

A scenario question describes a system with many technical details but the actual decision turns on who is accountable for a launch decision under incomplete evidence. What is the best exam strategy here?

Pick the longest answer option, since it is usually correctSkip the question since it seems too complexFocus only on the most technical-sounding details, since they are usually the answerIdentify the actual decision being tested — accountable ownership and risk acceptance under uncertainty — and filter out details that do not change that judgement

Which best describes the relationship between 'transparency' and 'explainability' as distinct but related concepts?

Transparency is about making relevant information about a system available; explainability is about making the reasons behind a specific output understandableTransparency only applies to code; explainability only applies to marketingNeither concept applies to generative AIThey are identical and interchangeable in every context

A company wants evidence that its 'human review' step is not just theatre. Which test result would most convincingly demonstrate meaningful oversight?

A log showing 100% of AI recommendations were accepted with average review time under 2 secondsA log showing reviewers sometimes overturn recommendations after examining underlying evidence, with documented reasons, within a realistic review timeframeA policy document stating that human review is requiredA vendor's marketing claim that the tool includes 'human-in-the-loop' features

A regulator asks why an organisation classified its recruitment tool as 'limited risk' rather than 'high-risk' under the EU AI Act. Which answer reflects sound practitioner judgement?

'Classification only matters after an incident occurs''We assumed low risk because the vendor said so''We assessed intended purpose, decision impact and use context against Annex III criteria, and can provide the documented reasoning and evidence behind the classification''We didn't classify it; classification isn't necessary if the tool works well'

Comparing an open-source self-hosted model to a proprietary hosted API for a sensitive internal use case, which factor most favours self-hosting?

Elimination of all need for governanceAutomatic compliance with every applicable lawGuaranteed lower total cost in every caseGreater control over data residency, customisation and security posture, at the cost of taking on more operational responsibility

A capstone evidence pack is being assembled for a final approval decision. Which combination best reflects what a defensible pack should contain?

A risk register, control/evidence matrix, testing results, monitoring plan and a clear, reasoned recommendation with a named accountable ownerOnly the original system design document from before any testingOnly positive results, omitting any unresolved risksOnly the final recommendation, with no supporting evidence

Exam centre

Confidence