STAGING — Test environment. No production data or live payments.

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.

Question 1
Support Responsible and Trustworthy AI Efforts

An AI system has been running well for a year. The project manager learns that an early design decision, documented at the time, means the model performs noticeably worse for a small subgroup of users. No complaints have been received. What is the most appropriate course of action?

  1. A.Take no action, since the decision was documented and no harm has been reported
  2. B.Raise it with the accountable owner, quantify the impact on that subgroup, and decide on remediation
  3. C.Correct it quietly in the next release without drawing attention to it
  4. D.Wait until a complaint is received, then address it as an incident
Show answer

Correct answer: B

Absence of complaints is not evidence of absence of harm, particularly for a small or marginalised subgroup.

  • A: Documentation records that a decision was made; it does not make the consequence acceptable. Knowing about a disparity and choosing inaction is a materially different position from not knowing.
  • B (correct): Quantifying the impact converts a known weakness into a decision someone with authority can take, on the record. The remedy may be a fix, a scope restriction or an accepted risk, but it should be chosen deliberately rather than by default.
  • C: A quiet fix may help future users and leaves those already affected unidentified and unremedied, and it removes any record of a decision the organisation may later need to account for.
  • D: Small subgroups are exactly the users least likely to complain, often because they have least standing to. Waiting for a complaint makes the response depend on the affected group's ability to advocate for itself.
Question 2
Support Responsible and Trustworthy AI Efforts

An organisation is procuring a model from a third-party supplier for a customer-facing decision. Which two contractual provisions most directly support the organisation's responsible AI obligations? Select two.

Select 2.

  1. A.A right to audit the supplier's testing evidence and model documentation
  2. B.A service credit payable if model availability falls below target
  3. C.An obligation on the supplier to notify the organisation of material changes to the model
  4. D.A commitment that the supplier will use the latest available model architecture
  5. E.A cap on the supplier's total liability under the agreement
Show answer

Correct answer: A, C

You remain accountable for a bought model, so contract for the visibility that accountability requires.

  • A (correct): Without a right to see how the model was tested and what it was tested on, the organisation is asserting responsibility for a system it cannot examine. This is the provision that makes any assurance verifiable.
  • B: Availability credits compensate for downtime. They are ordinary service management and say nothing about whether the model behaves responsibly while it is up.
  • C (correct): A model that changes silently invalidates every assessment the organisation has made of it. Notification of material change is what keeps the deployer's own governance from becoming fiction.
  • D: Newest is not a quality standard, and committing to it invites unreviewed change, which is the opposite of what governance needs.
  • E: A liability cap allocates financial risk between the parties. It is important commercially and does nothing for the people the system affects.
Question 3
Support Responsible and Trustworthy AI Efforts

A delivery team has completed its responsible AI training and agrees with the organisation's policy. Reviewing their last three releases, the project manager notices that none of the practices the training covered appear anywhere in how the team actually works: no fairness check ran, and no limitation was recorded. What is the most effective way to change that?

  1. A.Circulate the organisation's responsible AI policy to the team at project kickoff
  2. B.Make specific responsible AI checks part of the team's existing working agreements and definition of done
  3. C.Schedule an annual responsible AI training session for all team members
  4. D.Appoint one team member as the responsible AI representative
Show answer

Correct answer: B

Practice becomes routine when it is part of how work is already done, not an additional obligation alongside it.

  • A: A circulated policy is read once, if at all. It creates awareness, not behaviour, and nothing about the team's daily work changes.
  • B (correct): Embedding concrete checks into the definition of done means the work cannot be called finished without them. It makes the practice unavoidable and visible, and it puts the checks where the decisions are actually made.
  • C: Annual training provides background knowledge and decays quickly. It is a useful supplement and a poor primary mechanism.
  • D: A single named representative concentrates the responsibility in one person and gives everyone else permission to stop thinking about it, which is the opposite of the intent.
Question 4
Support Responsible and Trustworthy AI Efforts

A junior analyst reports that the bias assessment required by the organisation's own policy was skipped in order to protect the release date. The team lead dismisses it, saying the assessment can be done after launch and the schedule does not allow for it now. What should the project manager do?

  1. A.Support the team lead's judgement, since schedule pressure is a legitimate constraint
  2. B.Have the concern assessed on its merits and, if substantiated, escalate the schedule implication rather than the concern
  3. C.Ask the analyst to gather evidence and raise it again if they can prove it
  4. D.Note the concern in the risk register and proceed as planned
Show answer

Correct answer: B

A schedule constraint is a reason to escalate a finding, never a reason to decide it does not apply.

  • A: The team lead has not judged whether the assessment matters; they have declined to run it on cost grounds. Endorsing that substitutes a scheduling decision for a policy one, and teaches the team that raising such things is futile.
  • B (correct): Establishing the facts separates the two questions. Whether a required assessment was skipped does not change with the deadline, and if it was, the resulting schedule impact belongs with whoever owns the schedule to decide on, not with the team lead to absorb quietly.
  • C: Placing the burden of proof on the most junior person, against their team lead, reliably ensures such reports stop being made.
  • D: Logging and proceeding records that the organisation knew a required control was skipped, which is a worse position than not knowing, and does nothing to establish whether it should have been.
Question 5
Support Responsible and Trustworthy AI Efforts

An AI system will control equipment that moves near people. Engineering has proposed a set of safety constraints enforced outside the model. The project manager is asked what governance this requires beyond the technical design. What is most important?

  1. A.A named individual with the authority to accept the residual risk, on the record, before the system operates
  2. B.A commitment from engineering that the constraints have been tested
  3. C.A record of the model's accuracy on the safety-relevant cases
  4. D.Agreement from the vendor that the constraints are correctly implemented
Show answer

Correct answer: A

Technical controls reduce risk; somebody with authority still has to accept what is left.

  • A (correct): Constraints never reduce risk to zero, so a residual remains. Naming who accepted it, before operation and in writing, is what makes the decision accountable and what gives that person standing to stop it.
  • B: Testing evidence is necessary and is an input to the decision, not the decision. It tells you the control works as built, not whether the remaining risk is acceptable to the organisation.
  • C: Accuracy figures describe the model. The residual risk here comes from what happens when the model is wrong, which is precisely what the accuracy number does not tell you.
  • D: Vendor confirmation is worth having and shifts nothing. The organisation operating equipment near people answers for the harm regardless of who implemented the control.
Question 6
Support Responsible and Trustworthy AI Efforts

An AI system has been operating for eight months. A near miss occurs: the system produced an output that would have caused harm, but a human noticed in time. No one was hurt and the operator resolved it locally. The project manager learns about it informally three weeks later. What is the most significant problem?

  1. A.The operator should have been given clearer instructions for that situation
  2. B.There is no reporting route that brings near misses to the people accountable for the system
  3. C.The system should have been suspended as soon as the output occurred
  4. D.The human oversight arrangement worked, so no action is required
Show answer

Correct answer: B

A near miss is free information about how a system fails, and it is only free if it reaches someone who can act.

  • A: Better instructions might help that operator next time. It treats a systemic blind spot as a training gap, and the next near miss would be resolved locally and go unreported in the same way.
  • B (correct): The organisation has been running for three weeks unaware that its system can produce harmful output. Without a route that carries near misses to the accountable owner, the only incidents anyone learns from are the ones that already caused harm.
  • C: Immediate suspension may or may not be proportionate, and that judgement belongs to whoever is accountable. They could not make it, because nobody told them.
  • D: Oversight caught it once, which is fortunate rather than reassuring. Treating a near miss as a success is how an organisation convinces itself a recurring fault is under control.

Want the full question bank?

Full-length 120-question CPMAI mock exams with a real exam timer, domain-wise scoring and detailed explanations.

Question 7
Support Responsible and Trustworthy AI Efforts

A model has behaved reliably for a year. The vendor announces that the underlying model will be updated automatically next month, with improved benchmark performance. The project manager is asked whether any action is needed. What should the project manager conclude?

  1. A.No action, since the update improves measured performance
  2. B.No action, provided the vendor confirms the interface is unchanged
  3. C.The organisation's own validation and approval must be repeated before the updated model is relied on
  4. D.The update should be refused, since the current model is working
Show answer

Correct answer: C

A model that changes silently invalidates every assessment the organisation made of it.

  • A: Benchmark improvement says the model is better on the vendor's tests. It says nothing about this organisation's data, its edge cases, or the specific harms its own assessment identified.
  • B: A stable interface means the integration will not break. The behaviour behind that interface is exactly what changed, and that is what was validated.
  • C (correct): Every fairness result, performance figure and risk acceptance on file describes the previous model. Re-running validation and re-approving is what keeps the organisation's governance describing the system that is actually running.
  • D: Refusing all updates freezes the system on a model that will not receive fixes, including security ones. The answer is to control the change, not to prevent it.
Question 8
Support Responsible and Trustworthy AI Efforts

An organisation uses a third-party model for a customer-facing decision. Eighteen months into the contract, the project manager wants to confirm the arrangement is still sound. Which check gives the most assurance?

  1. A.Confirming the vendor is still trading and the contract has not lapsed
  2. B.Reviewing the vendor's current testing evidence against the organisation's own monitoring of outcomes
  3. C.Asking the vendor whether any issues have been reported by other customers
  4. D.Comparing the vendor's pricing against alternatives in the market
Show answer

Correct answer: B

Assurance comes from comparing what the supplier says with what you can observe yourself.

  • A: Contractual and commercial status is basic supplier management. A solvent vendor with a live contract can still be supplying a model that has drifted badly.
  • B (correct): The vendor's evidence describes the model as they test it; the organisation's monitoring describes how it behaves on real customers. Divergence between the two is the earliest signal that something has changed, and neither source alone would reveal it.
  • C: Other customers' experience is useful context and arrives filtered through the vendor, who has an interest in how it is characterised. It is not a substitute for your own observation.
  • D: Pricing is a commercial review. It has no bearing on whether the model is still fit for the decision it is making.
Question 9
Support Responsible and Trustworthy AI Efforts

A data scientist tells the project manager privately that they believe a model about to be released will disadvantage a group of customers, but they do not want to raise it formally because the release is high-profile and they are new to the organisation. What should the project manager do?

  1. A.Respect the confidence and monitor the model closely after release
  2. B.Encourage them to raise it themselves, since a concern carries more weight from the person who found it
  3. C.Take the concern forward through the escalation route without requiring them to be named as its source
  4. D.Ask them to gather supporting evidence before deciding whether to act
Show answer

Correct answer: C

The concern is the organisation's problem to assess; who raised it is not what makes it valid.

  • A: Keeping the confidence and watching afterwards accepts a known risk on behalf of customers who were never consulted, and post-release monitoring finds the harm after it has happened.
  • B: This puts the cost of speaking up on the most junior person in the room, at the moment they have told you they cannot bear it. It is also how organisations learn that raising concerns is the reporter's problem.
  • C (correct): The escalation route exists so a concern reaches someone who can assess it. Carrying it forward without exposing the source gets it assessed on its merits and demonstrates that the route is safe to use, which is what makes the next person willing to speak.
  • D: Asking for evidence first is reasonable for a vague concern and here it defers assessment while the release approaches, and it signals that a burden of proof rests on the reporter.
Question 10
Support Responsible and Trustworthy AI Efforts

A model will change how a community service is allocated. The project manager is planning engagement and has budget for one substantial activity. Which is most valuable?

  1. A.A briefing for community representatives once the model is built, explaining how it works
  2. B.Working with representatives of the affected community while the model's objectives and success measures are still being set
  3. C.A public consultation after launch, inviting feedback on the service
  4. D.A survey of current service users measuring satisfaction with the existing process
Show answer

Correct answer: B

Engagement changes outcomes only while the decisions it could influence are still open.

  • A: Explaining a finished model is communication, not engagement. The objectives are fixed, the trade-offs are made, and the only available response is acceptance or objection.
  • B (correct): Objectives and success measures encode whose interests the model serves and what counts as a good outcome. Involving the affected community while those are open is the one point where engagement can still change the design rather than explain it.
  • C: Post-launch consultation gathers reaction to decisions already taken and already affecting people. It is worth doing and it is the weakest point at which to start.
  • D: Baseline satisfaction data is useful for measuring whether the change helped. It is research about people rather than engagement with them.
Question 11
Support Responsible and Trustworthy AI Efforts

An organisation deploys a model that another company developed and sold as a general-purpose tool. The organisation configures it for its own hiring process. A stakeholder assumes the developer carries the regulatory obligations. How should the project manager characterise the position?

  1. A.The developer carries them, since they built and trained the model
  2. B.The obligations transfer entirely to the organisation once it configures the tool
  3. C.Both carry obligations, and putting a general-purpose tool to a specific high-stakes use typically attracts duties that sit with the organisation doing so
  4. D.Neither carries obligations until the tool has been used to make an actual decision
Show answer

Correct answer: C

Regulation usually follows the use, and the organisation choosing the use is the one that can answer for it.

  • A: A developer answers for how the tool was built and documented. They know nothing about this hiring process and cannot assess whether the tool is appropriate for it, which is why the duties do not stop with them.
  • B: The developer's build-side obligations do not disappear because someone configured the tool. Assuming a clean transfer leaves the organisation relying on evidence nobody is obliged to supply.
  • C (correct): Deciding to use a general tool for hiring is what creates the risk to candidates, and the organisation making that decision is the only party positioned to assess it, oversee it and answer for it.
  • D: Obligations attach well before the first decision: assessment, documentation and oversight all have to be in place beforehand. Waiting for a decision means the duties were breached by the time anyone looks.

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.