Free PMI CPMAI Practice Questions
These free PMI CPMAI practice questions are written in the scenario style of the real exam: you are managing an AI initiative and must choose the best next step. Each question comes with the correct answer and an explanation of the reasoning.
Read each scenario, pick your answer, then tap Show answer to check it. For the exam format, domains and a study plan, see our PMI CPMAI exam guide.
A model that allocates maintenance crews to buildings is found to route fewer visits to one neighbourhood. Investigation shows the model was trained on past maintenance requests, and that neighbourhood historically reported fewer faults because residents had less confidence the reports would be acted on. Which two actions are most appropriate? Select two.
Select 2.
- A.Change the training target from reported faults to an independent measure of building condition
- B.Increase the model's confidence threshold so that fewer allocations are made overall
- C.Add a correction factor that increases predicted need for that neighbourhood by a fixed amount
- D.Collect condition data through proactive inspection rather than relying on resident reports
- E.Present the disparity to the service owner and ask them to override individual allocations
Show answer
Correct answer: A, D
When the data records who asked rather than who needs, the fix is to measure need directly.
- A (correct): Reported faults measure reporting behaviour as much as building condition. Retargeting the model at condition itself removes the distortion at its source rather than patching the output.
- B: Fewer allocations overall does nothing for the distribution between neighbourhoods and reduces service for everybody. It suppresses the symptom and worsens the outcome.
- C: A fixed uplift is a guess with no basis in measured need. It is unjustifiable to anyone who asks how the number was chosen, and it breaks as soon as conditions change.
- D (correct): Proactive inspection generates the unbiased data the retargeted model needs. It answers the question the reports could never answer, and is the practical route to fixing the input rather than the output.
- E: Case-by-case override puts the burden on one person to correct a systematic problem, produces no record of why decisions changed, and leaves the model generating the same skewed allocations underneath.
A team proposes to satisfy an explainability requirement by generating a plain-language justification for each decision using a separate language model that reads the prediction and the input features. What is the most significant weakness of this approach?
- A.The generated text may sound convincing while not reflecting what the model actually computed
- B.Generating text for every decision will be too slow at production volumes
- C.The language model will need its own training data
- D.Regulators generally prefer numerical explanations to narrative ones
Show answer
Correct answer: A
An explanation that was not derived from the model's actual reasoning is a plausible story, not an explanation.
- A (correct): The narrator has no access to why the model decided as it did; it sees an input and an output and writes something that fits. Fluent, confident and unfaithful is the worst possible combination, because it defeats scrutiny rather than enabling it.
- B: Latency is a solvable engineering problem and would not by itself disqualify the design. It is a cost, not a flaw in the reasoning.
- C: Whether the narrator needs additional training or configuration is an implementation question. However it is built, it still has the fidelity problem: it sees an input and an output, not the reasoning, so it can only produce something plausible.
- D: Regulators generally want explanations a person can understand, so narrative form is often preferred. The problem is truthfulness, not format.
A project manager is preparing to communicate a model's performance to a non-technical executive audience. The model achieves 0.81 area under the curve. What is the most useful way to present this?
- A.Report the figure and explain how area under the curve is calculated
- B.Translate it into what it means operationally: how many cases it catches, how many it misses, and what that costs
- C.Compare it against published benchmarks for similar models
- D.Omit the metric and report only that testing was successful
Show answer
Correct answer: B
Executives decide on consequences, so give them the consequence rather than the coefficient.
- A: Teaching the metric spends the audience's attention on statistics rather than on the decision in front of them, and they will still have to translate it themselves.
- B (correct): Expressed as cases caught, cases missed and the cost of each, the same number becomes something an executive can weigh against the alternative and decide on. It also surfaces the trade-off honestly.
- C: Benchmarks give useful context and answer a different question: how the model compares to others, not whether it is good enough for this purpose.
- D: Reporting only that testing passed withholds the information needed to make the decision, and invites the audience to approve something they do not understand.
A manufacturing director asks for a model to predict which machines will fail next month. During discovery the project manager finds that maintenance staff already know which machines are unreliable, but cannot get parts approved in time. What should the project manager conclude?
- A.The model should be built anyway, since a prediction would strengthen the case for parts approval
- B.The model's scope should be widened to include parts inventory forecasting
- C.The stated problem is a procurement delay, and a prediction model would not address it
- D.The project should proceed but with a reduced scope covering only the most critical machines
Show answer
Correct answer: C
If the answer is already known and the bottleneck is acting on it, prediction adds nothing.
- A: This assumes the approval process is blocked by a lack of evidence. The scenario says staff already know which machines are unreliable, so the constraint is elsewhere and a model would produce information nobody is short of.
- B: Expanding the scope adds work without testing whether the original premise holds. It moves further from the actual obstacle rather than closer to it.
- C (correct): The organisation is not failing to identify at-risk machines; it is failing to get parts approved in time to act. Naming that plainly is more valuable than delivering a model that predicts what people already knew.
- D: Narrowing the scope makes a project that does not address the problem smaller. Fewer machines does not change the fact that prediction is not the constraint.
A sponsor wants a model that predicts which service contracts are likely to be cancelled at renewal, so the retention team can act in advance, and asks the project manager to confirm early whether this is feasible. Which consideration is most decisive at this stage?
- A.Whether the organisation has recorded which past contracts were cancelled at renewal
- B.Whether the organisation has data scientists available
- C.Whether similar systems exist elsewhere in the industry
- D.Whether the infrastructure can support model training
Show answer
Correct answer: A
A prediction can only be learned where the thing being predicted has been recorded.
- A (correct): Learning to predict cancellation requires past cases labelled as cancelled or renewed. Where the organisation has never recorded the outcome, or records it inconsistently, that gap governs everything else and is often the point at which the use case has to be reshaped. Whether the recorded history is sufficient and representative is a later question.
- B: Staffing determines whether the work can be resourced, which is solvable by hiring or contracting. It does not determine whether the problem is tractable.
- C: Industry precedent raises confidence and tells you little about your own position, since their data situation may be nothing like yours.
- D: Infrastructure can be procured. It rarely decides feasibility for a business application.
A vendor offers a model that performed well in a demonstration using the vendor's own sample data. The project manager must decide whether to recommend it. What is the most informative next step?
- A.Ask for references from other customers in the same industry
- B.Negotiate a trial period with the option to withdraw
- C.Request the vendor's internal accuracy benchmarks
- D.Require evidence that the product meets the organisation's agreed business requirements on representative cases of its own before recommending purchase
Show answer
Correct answer: D
A demonstration shows the product at its best; evidence on your own representative cases shows it against your requirement.
- A: References describe other organisations' experience, filtered through customers the vendor chose to offer. Useful context, weak evidence about your data.
- B: A trial with an exit is sound commercial protection and it postpones the evaluation rather than performing it. Without agreed criteria, the trial produces an impression rather than a finding.
- C: Vendor benchmarks are measured on data the vendor selected. They tell you what the product can do under favourable conditions, which the demonstration already suggested.
- D (correct): The organisation's own representative cases, judged against requirements agreed before the results are seen, are what answer the question being asked. Agreeing the requirements first is what stops the outcome being interpreted to fit the decision already leaning one way.
An insurer needs weather history to model claim surges after storms. Internal records cover only the last two years, which the team agrees is too short to capture severe events. A commercial provider sells forty years of the same measurements at a price the project can afford. What should the project manager confirm before committing to the purchase?
- A.That the provider's file format can be loaded by the team's existing tooling
- B.That the purchase price fits inside the current phase's allocated budget
- C.That the licence permits the data to be used for model training and the intended commercial deployment
- D.That the provider is used by other insurers working on comparable problems
Show answer
Correct answer: C
Bought data carries the terms it was sold under, and those terms can forbid the use you planned.
- A: Loading problems are real but solvable with conversion work, and they delay the team rather than stopping the project. A permission problem can invalidate the model after it has been built.
- B: Budget fit is necessary for the purchase to proceed at all, and it is easy to check. It says nothing about whether the data may lawfully be used the way the project intends.
- C (correct): Licences may restrict redistribution, derived works, and commercial deployment of anything trained on the data. Confirming the grant covers both the training use and the intended deployment is what prevents a model that cannot lawfully ship once it is built.
- D: Other insurers using the provider suggests the data is credible. Their licences may differ from yours, and their permitted uses tell you nothing about your own grant.
A team removes customer names and account numbers from a dataset and reports to the project manager that the data is now anonymised and can be handled with reduced controls. The fields remaining in the extract include full postcode, exact date of birth and current employer. What should the project manager point out?
- A.The remaining fields will reduce the model's predictive accuracy
- B.The remaining fields can still identify individuals when combined
- C.The removal should have been applied after the train and test split
- D.The removal should be documented in the project's data catalogue
Show answer
Correct answer: B
Identifiability survives the deletion of obvious identifiers.
- A: Accuracy may move in either direction when fields are removed. The concern raised here is about the privacy claim being wrong, not about model performance.
- B (correct): Full postcode, exact date of birth and employer can substantially narrow the possible individuals and may permit re-identification when combined. Calling the result anonymised therefore understates the remaining identification risk.
- C: Split ordering matters for preventing leakage between training and test sets. It does not affect whether the retained fields identify people.
- D: Cataloguing the change is good practice and aids traceability. Documenting an incorrect anonymisation claim records the error rather than correcting it.
A team is specifying data for a model that will score insurance claims for fast-track handling at the moment each claim is submitted. A strong candidate field is the loss adjuster's assessment, which is recorded several days after submission. What should the project manager conclude?
- A.The field should be included because it is strongly associated with the outcome of interest
- B.The field will not exist at the moment the model is required to score, so it cannot be an input
- C.The field should be included provided its values are complete in the historical record
- D.The field should be included and the model rerun once the assessment is recorded
Show answer
Correct answer: B
A field that does not exist when the model runs cannot be an input, however predictive it looks.
- A: Strength of association in historical data is what makes the field tempting. It is exactly the wrong basis for selection when the value is not yet known at the point of use.
- B (correct): A model can only use what is available at the point it runs. The adjuster's assessment is created days after submission, so at scoring time the field is empty for every claim the model must judge, however predictive it looks in historical data.
- C: Completeness in history says nothing about availability at scoring time. A fully populated historical column can still be entirely absent when a new claim arrives.
- D: Rescoring later describes a different product serving a different decision. The fast-track decision is needed at submission, so a model that waits for the assessment does not address it.
Before a fraud detection model goes live, the project manager wants the cut-off set deliberately rather than by default. The team asks what information is needed from the business to make that decision. Select two.
Select 2.
- A.The algorithm the team used to produce the fraud scores
- B.The relative cost of a missed fraud and a wrongly flagged transaction
- C.The share of transactions that were fraudulent in each of the last three years
- D.The number of flagged cases the review team can handle in a day
- E.The total number of features the model uses to generate each score
Show answer
Correct answer: B, D
Setting a cut-off requires knowing what each error costs and how much review capacity exists.
- A: The method behind the scores does not determine where the cut-off belongs. The same cut-off question arises whatever algorithm produced them.
- B (correct): The cut-off balances the two errors, so their relative cost sets where the balance should fall. Without it the choice has no basis beyond convention.
- C: Historical prevalence is useful background for interpreting the scores. It does not by itself say what balance of errors the business should accept.
- D (correct): A cut-off that flags more cases than the team can review produces a backlog, and the unreviewed cases receive no action. Capacity puts a practical ceiling on the choice.
- E: Feature count describes the model's construction and has no bearing on where the decision boundary should sit.
Want the full question bank?
Full-length 120-question CPMAI mock exams with a real exam timer, domain-wise scoring and detailed explanations.
A team reports that raising a fraud model's recall from sixty to eighty per cent caused its precision to fall from seventy to forty-five per cent. A stakeholder asks whether the change is an improvement. What should the project manager say determines the answer?
- A.Whether the sum of precision and recall is higher after the change than it was before
- B.What a missed fraud costs compared with a review of a legitimate case
- C.Whether the model's overall accuracy also improved when the change was made
- D.Whether the recall figure now exceeds the precision figure by a clear margin
Show answer
Correct answer: B
Whether a precision and recall trade is worth making depends on what each kind of error costs.
- A: Adding the two measures treats the errors as interchangeable. They are not, and the sum has no meaning in terms of what the organisation loses.
- B (correct): The trade moves error from one type to the other. Only the relative consequences of a missed fraud and a review of a legitimate case can establish whether the new balance is preferable.
- C: Overall accuracy combines both error types into a single figure and can move very little when either measure changes. It cannot arbitrate a trade between them.
- D: There is no general principle that recall should exceed precision. Which matters more depends entirely on the costs involved.
A model is approaching a deployment decision. The project manager finds that its intended use, the population it was evaluated on and the conditions under which it should not be relied upon have not been written down anywhere. The team considers this documentation to be a post-deployment task. What is the risk of deferring it?
- A.The model may be applied outside the conditions it was evaluated for
- B.The team will have forgotten the details of the evaluation by the time it writes them down
- C.The deployment will be delayed while the documentation is prepared after release
- D.Auditors will require the documentation to be rewritten in a prescribed format later on
Show answer
Correct answer: A
Undocumented limits become invisible, and a model may be applied where it was not shown to work.
- A (correct): Without documented limits, downstream users may apply the model outside the population or conditions for which it was evaluated, with nothing available to tell them where it does not apply.
- B: Recollection may fade, and records can be reconstructed from the evaluation artefacts. The consequential risk is misuse in the meantime.
- C: Documentation written after release does not delay the release. The concern is what happens in the period without it.
- D: Format requirements are a compliance detail. They do not describe what goes wrong operationally when limits are unrecorded.
A model that passed evaluation is ready to go live for all branches at once. The project manager proposes releasing it to a small number of branches first. The team argues that evaluation already showed the model works and that a staged release only delays the benefit. What does the staged release provide that evaluation did not?
- A.Evidence of how the model behaves within the live process, at limited exposure
- B.Another measurement on the same held-out evaluation records before wider release
- C.An opportunity to collect more training records before the model reaches every branch
- D.Assurance that the model's infrastructure can carry the full production load
Show answer
Correct answer: A
Evaluation measures the model; a staged release measures the model inside the live process around it.
- A (correct): Evaluation scored predictions against held-out records. A staged release exercises the integrations, the handoffs and the people acting on the output, with a small population carrying any problems.
- B: Re-running the original evaluation repeats a measurement already taken on records the model has already been scored against. It exercises none of the live process around the model.
- C: Records gathered during a staged release can be useful later. Gathering them is not the reason to limit exposure at go-live.
- D: Load behaviour at a few branches says little about full volume, so a staged release is weak evidence for capacity. Capacity is tested separately.
Monitoring shows that the mix of incoming records for a live model has shifted markedly over six weeks, although the model's measured accuracy has not yet moved. The team proposes taking no action until accuracy declines. What is the argument for acting on the input shift before then?
- A.Outcomes arrive after the decisions are made, so waiting may expose live decisions before any performance effect becomes observable
- B.An input shift always causes a fall in accuracy, so the decline is certain to follow
- C.The measured accuracy figure is unreliable whenever the input mix has changed
- D.Retraining is cheaper to perform before a decline appears than after one
Show answer
Correct answer: A
A shift in inputs can precede a performance effect that outcome data will only reveal later.
- A (correct): The input shift is observable now; any accuracy effect becomes measurable only once outcomes for the new mix are known. Waiting concedes that interval to a change whose consequences are not yet visible.
- B: A model can be insensitive to a shift in fields it depends on weakly, so a decline is a risk rather than a certainty. The case for acting rests on the delay in observing it.
- C: The figure remains a valid measurement of the records it covers. The issue is what it cannot yet cover, not that it is wrong.
- D: Cost may well be similar either way, and the reason to act early is exposure during the delay rather than the price of the work.
Operational staff can override a live model's recommendation. Over the first quarter, overrides are recorded as a simple count with no reason captured. The count is rising. The project manager wants to understand what the rise means. What should be changed in how overrides are handled?
- A.Require approval from a supervisor before an override can be recorded
- B.Remove the override facility, since a rising count suggests it is being misused
- C.Capture the reason for each override, so the pattern behind the rise can be seen
- D.Compare the override count with the count recorded by a similar organisation
Show answer
Correct answer: C
An override is evidence about the model, but only if the reason for it is captured.
- A: An approval step suppresses overrides rather than explaining them, and may discourage staff from correcting the model at all.
- B: Nothing indicates misuse, and removing the facility would leave staff unable to correct a wrong recommendation.
- C (correct): A rising count is consistent with a degrading model, a shift in cases, or growing confidence among staff in exercising judgement. Without reasons the three are indistinguishable.
- D: Another organisation's rate reflects its own model, cases and staff. It cannot explain the pattern here.
Which statement correctly describes how standard K-means assigns a data point after clusters have been formed?
- A.It gives the point a probability for every possible hidden state
- B.It assigns the point to the cluster with the nearest centroid
- C.It asks an agent to choose the cluster and then rewards that choice
- D.It predicts the point's label from word frequencies
Show answer
Correct answer: B
K-means makes a hard assignment to the nearest cluster centroid.
- A: Assigning a probability across hidden states describes HMM, not K-means.
- B (correct): Correct. Standard K-means assigns each point to a single nearest centroid.
- C: Reward-driven action selection describes reinforcement learning.
- D: Word-frequency classification is associated with Naive Bayes.
A model must recognise objects in photographs, where what matters is the spatial arrangement of nearby pixels rather than their position in a list. Which neural network design is built for that kind of spatial structure?
- A.Recurrent neural network (RNN)
- B.Hidden Markov Model (HMM)
- C.K-Nearest Neighbor (KNN)
- D.Convolutional neural network (CNN)
Show answer
Correct answer: D
A convolutional network learns features from local spatial regions, which suits image data.
- A: Recurrent networks carry information forward across ordered steps and are built for sequence, not for spatial relationships across an image.
- B: HMM models hidden states behind an ordered sequence, which does not capture the two-dimensional spatial structure of a photograph.
- C: KNN compares a new case against whole stored examples and learns no spatial features within an image.
- D (correct): Correct. Convolutional layers detect features within small neighbouring regions and build up to larger shapes, which is what spatial recognition needs.
Which combination of elements is the strongest indicator of reinforcement learning?
- A.Labels, features, and a fixed target
- B.Clusters, centroids, and a chosen K
- C.Hidden states, observations, and transition probabilities
- D.Agent, environment, actions, and rewards
Show answer
Correct answer: D
Reinforcement learning is identified by an agent acting in an environment and improving from rewards or penalties.
- A: Labels and targets are characteristic of supervised learning.
- B: Clusters and centroids are characteristic of K-means.
- C: Hidden states and sequential transitions are characteristic of HMM.
- D (correct): Correct. This action-feedback loop is the signature of reinforcement learning.
For a use case where an auditor must trace one prediction through explicit steps, why might a single decision tree be preferred over a random forest?
- A.A random forest's individual trees are too shallow to inspect
- B.A forest can be traced by averaging its trees into one tree
- C.A single tree is preferred because it uses fewer input features
- D.A single tree exposes one branch path, while a forest aggregates many trees.
Show answer
Correct answer: D
One decision tree offers direct traceability, while an ensemble of many trees reduces that transparency.
- A: Depth is not the obstacle. Individual tree outputs can be examined; what makes a forest harder to interpret is that its prediction aggregates many trees rather than following one path.
- B: Trees cannot be meaningfully averaged into one equivalent tree. The forest's prediction comes from aggregating separate trees, and that aggregation is what resists tracing.
- C: A single tree may use just as many features as a forest. Its advantage to an auditor is the traceable branch path, not the number of inputs.
- D (correct): Correct. A single branch path is easier to audit than an aggregate across many trees.
A learner sees four clues: 'nearest labeled examples,' 'maximum-margin boundary,' 'membership weights across clusters,' and 'hidden states over time.' Which sequence of algorithms matches those clues?
- A.K-means, Decision tree, GMM, Linear regression
- B.Naive Bayes, KNN, SVM, K-means
- C.Decision tree, Naive Bayes, K-means, GMM
- D.KNN, SVM, Fuzzy C-means, HMM
Show answer
Correct answer: D
The four clues map respectively to KNN, SVM, Fuzzy C-means, and HMM.
- A: None of these four fits its clue: K-means assigns to the nearest centroid rather than comparing stored labeled examples, and linear regression fits a line rather than modelling hidden states over time.
- B: KNN and SVM do match two of the clues, but not the ones they are paired with here. Naive Bayes classifies by feature probabilities, and K-means does not model hidden states over time.
- C: A decision tree traces branching rules rather than comparing nearest labeled examples, K-means makes hard cluster assignments rather than membership weights, and GMM models a static mixture rather than states changing over time.
- D (correct): Correct. KNN uses nearest labeled examples; SVM uses a maximum-margin boundary; Fuzzy C-means uses membership weights; HMM models hidden states across sequences.
Ready for the real thing?
Practise with full CPMAI mock exams that match the real exam: 120 questions in 160 minutes, with explanations for every answer.
CertMocks is an independent exam-prep platform and is not affiliated with or endorsed by PMI. All questions are independently written for practice.