Free PMI-CPMAI Exam Flashcards

Memorize 50 essential terms and definitions for the PMI Certified Professional in Managing AI (PMI-CPMAI). See the term, recall the definition, then flip to check yourself.

50 Flashcards
5 Topics
100% Free
TermClick to flip

How does data minimization reduce privacy risk in an AI project?

Tap to reveal definition
Card 1 of 50Responsible and Trustworthy AI

Filter by Topic

Jump to Card

About These PMI-CPMAI Flashcards

These 50 flashcards are designed to help you memorize key terms and definitions for the PMI Certified Professional in Managing AI (PMI-CPMAI). Each card shows a term on the front and its definition on the back—the classic flashcard format for vocabulary memorization. Use these alongside our practice questions to build both recall and comprehension.

Topics Covered

Responsible and Trustworthy AI7 cards
Business Needs and Solutions13 cards
Data Needs13 cards
Model Development and Evaluation8 cards
Operationalizing AI Solutions9 cards

Complete Flashcard Reference

Review every term in this set. Open any term to reveal its definition.

How does data minimization reduce privacy risk in an AI project?

Collect and retain only the information needed for the intended use. If a delivery forecast works at the postal-code level, leaving street addresses out of the dataset reduces exposure.

How does least privilege apply to an AI training workspace?

Give each person and service only the access it needs. A labeling contractor can review assigned records through a controlled interface without getting bulk export access.

What makes a model selection decision transparent to reviewers?

Document the options considered, the selection criteria, the evidence used, and the trade-offs behind the choice. Reviewers can then trace how the team reached its decision.

What should an explanation of a recommendation help an affected user understand?

Explain in plain language which factors mattered to that recommendation and what it means for the user. Check that the explanation faithfully reflects the system’s behavior.

Why check model error rates across relevant population groups?

A good overall score can hide a higher error rate for a specific group. Compare errors that matter to the use case, then investigate differences and the quality of the evidence behind them.

How should a team respond when an applicable AI policy changes mid-project?

Review the change with legal or compliance, identify affected controls and approvals, and update the evidence and review plan before proceeding with affected work.

What should an auditor be able to reconstruct from an AI release record?

Which model and data versions were reviewed, what evidence supported the decision, and who approved it. Link those records to the actual release so the team can answer for a specific deployment.

What makes an AI project problem statement useful before solution design?

Name the affected users, the workflow problem, and the consequence. For example, service staff lose time rerouting misfiled requests. This gives the team an outcome to investigate before picking a tool.

Which AI pattern fits forecasting call volume to inform staffing?

Predictive analytics and decision support. A forecast gives supervisors evidence for scheduling decisions while they remain responsible for assigning staff.

When should a deterministic rule be the first option to evaluate?

When the decision follows stable, explicit criteria. Start by testing whether a rule meets the business goal with a lower operating burden, then use that result as a baseline for any AI option.

What should an early AI feasibility test focus on?

Test an assumption that could stop the project, such as whether the available signals can distinguish urgent requests from routine ones. Use the result to decide whether to invest in a larger build.

How can a risk assessment turn a possible harmful model output into an action?

Describe the failure and its impact, estimate likelihood and severity, and assign an owner to a specific control. Give the owner a way to check whether the control reduces the risk.

What should the scope statement say about an AI tool that proposes refunds?

Define where it operates and what decisions it can make. For example, the pilot may suggest refund amounts to agents while agents retain approval authority.

Which costs belong in an AI solution’s total cost of ownership?

Estimate costs across the intended operating life, including data and development, integration, inference, monitoring, retraining, and support. Recurring costs can change which option makes business sense.

How can you tell whether people are adopting an AI tool after training?

Measure use in the intended workflow. For example, track the share of eligible cases where staff review the recommendation. Use feedback to find friction when usage falls short.

What should you verify about an existing workflow before adding an AI decision step?

Check where inputs come from, where the output must arrive, and how much delay the workflow can tolerate. A strong model adds little value if its recommendation reaches staff after the decision.

What should an early AI solution sketch make visible?

Show how data enters, where the model processes it, and how its output reaches a user or another system. Reviewing the flow exposes unresolved interfaces and handoffs.

How do you turn a business goal into a measurable AI success criterion?

Define the outcome metric, its baseline, a target, and a measurement window. A pilot might aim for a 15% reduction in median request-routing time versus the current workflow.

What should an AI business case show when benefits depend on uncertain adoption?

Compare the costs, benefits, and risks of viable options. Show how projected value changes under different adoption assumptions so decision-makers can judge the investment.

How can a capability gap affect an AI project’s plan?

A missing skill, such as production data engineering, can become a delivery dependency. Identify who will fill the gap through staffing, training, or contracting and reflect that choice in the schedule.

What should a team specify about the target label before collecting examples?

Define the outcome the model should learn and how to identify it consistently. For product returns, specify which events count as returns and the observation window used to assign the label.

What does data granularity describe?

It is the level of detail represented by a record. Daily totals may support capacity planning, while hourly coverage decisions need data at a finer time interval.

Why involve a data subject matter expert early?

A domain expert can explain how a field is generated and what it means in practice. A blank closure date, for example, might indicate a reopened case or an incomplete migration.

What should a data source inventory capture before collection?

List each source’s location, owner, contents, refresh cycle, and access path. This lets the team trace dependencies and identify which sources to verify first.

What makes a workspace ready to handle sensitive model-development data?

Provide approved compute and storage, defined access controls, and a workspace where the team can track changes and recover data. Verify those controls before transferring the dataset.

How can you verify that an extraction captured the intended records?

Reconcile counts and date ranges against the source and investigate differences. A feed that runs successfully can still omit a region or reporting period.

What should a team verify before using a vendor dataset for model training?

Check whether the license or agreement permits the planned training and downstream use. A dataset that a team can download may still have restrictions on how it can be used.

What record helps trace an unexpected value through several data transformations?

Use data lineage. It shows where the data originated and what changed at each step, which helps locate whether the bad value came from the source or the pipeline.

How do you test data accuracy when a dataset has no blank fields?

Compare sampled values with a trusted reference or verified observations. Filled records demonstrate completeness for those fields, while the comparison tests whether the recorded values are correct.

What makes a dataset fresh enough for the intended use?

Its update timing and lag must suit the decision. A weekly snapshot can inform quarterly planning, while same-day inventory decisions may require a more frequent feed.

Why can a large dataset still fail a representativeness check?

It may contain few examples from the users or conditions the system will face. A million weekday transactions give limited evidence about a workflow that also runs on weekends.

What should happen when the available data fails a critical readiness requirement?

Hold the proposed modeling work while the team closes the gap or revises the use case. Record the missing evidence and the condition that would allow work to proceed.

What belongs in a data readiness update for leadership?

Summarize what the team can use, the gaps that threaten the business outcome, and the cost or time to address them. Recommend a decision and state when new evidence will be available.

When is a simpler model the stronger choice?

When it meets agreed performance and risk criteria and is easier to explain or maintain. Compare the options under the same evaluation conditions so the trade-off is visible.

What makes a model test result traceable to the configuration tested?

Record the model artifact and the versions of its data, code, and configuration, then attach the test result to that record. This lets reviewers identify or reproduce the tested build.

What is the role of a validation set during model development?

Use it to compare candidate models and tune choices. Reserve a separate test set for the final performance estimate after those choices are settled.

Why can duplicate records across training and test sets inflate performance?

The model has already seen some test examples during training. Group or remove duplicates before splitting so the final test measures performance on genuinely unseen examples.

Why should serving use the same preprocessing as training?

The same raw inputs should produce features with the same meaning in each environment. Learn transformation settings from the training data and apply those fixed settings in validation, testing, and serving.

Why check label agreement before approving training data?

Inconsistent interpretations give the model conflicting examples of the target. Review disagreements against the labeling rules and resolve material gaps before approving the data for training.

Which metric measures the share of actual defects a classifier catches?

Recall. Divide detected defects by all actual defects, including any the model missed. Use it when missed detections matter to the use case.

What evidence supports a go decision for model operationalization?

Check that the model meets the agreed acceptance criteria on suitable evaluation data. Confirm that production infrastructure, monitoring, and support procedures are ready before approving deployment.

What should a phased rollout specify about rollback?

Set a measurable trigger, name who can act, and document how to restore a known good model or service. Test the recovery path before widening the rollout.

What does a post-deployment smoke test verify?

It checks that a controlled request can pass through the production workflow and reach the intended destination. It can catch integration or configuration failures that an offline model test misses.

What does a production input drift alert tell you?

The distribution of one or more inputs has shifted relative to a reference. Investigate the reason and check outcomes where labels are available before deciding whether to update the model.

What should happen before a retrained model replaces the live version?

Evaluate it against release criteria, record its lineage and results, and use the normal approval process. Keep the current version available if the replacement needs to be rolled back.

Why monitor business outcomes alongside technical model metrics?

They show different parts of the service. A routing model may still classify cases accurately even as time to resolution worsens because the receiving queues are understaffed.

What makes a production monitoring alert actionable?

Give it a defined trigger, a responsible responder, and a first diagnostic action. For a surge in stale inputs, the runbook might direct the responder to check the source feed’s last successful refresh.

What makes a lessons-learned entry useful for the next AI project?

Record the observed outcome, what contributed to it, and a practice the next team can apply. Include enough context to judge whether the same change fits another project.

What should operations accept at an AI service handover?

Confirm ownership for monitoring, incidents, and maintenance. Provide runbooks, access, and training, then verify that the receiving team can perform a routine check and follow an escalation procedure.

How does a contingency plan keep a workflow running during an AI outage?

Use a tested fallback, such as sending requests to a manual queue. Define who activates it, how users are informed, and the checks for safely restoring the AI service.

Frequently Asked Questions

How many questions are on the PMI-CPMAI exam?

PMI gives you 160 minutes to answer 120 questions. Of these, 100 count toward the result and 20 are unscored pretest questions. The exam has no scheduled breaks. Pearson VUE offers test-center and online-proctored delivery.

What is the passing score for PMI-CPMAI?

PMI uses psychometric analysis to set the passing standard and reports a pass or fail outcome. PMI publishes neither a numerical passing threshold nor a candidate pass rate for this exam.

What do I need to do before I schedule the PMI-CPMAI exam?

You must be at least 18 and complete PMI’s 21-hour PMI-CPMAI Exam Prep Course. Then use your myPMI dashboard to schedule with Pearson VUE. PMI requires no prior project management, technical, or AI experience or certifications.

What is the PMI-CPMAI retake policy?

You have one year from purchase to obtain the certification, with up to three total attempts in that eligibility period. PMI suggests 30 days of preparation before a retake and charges for each additional attempt. After three unsuccessful attempts in the eligibility period, wait one year from your last exam before reapplying.

Which PMI-CPMAI exam domains carry the most weight?

Business needs and solutions and data needs are 26% each. The other domains cover operationalizing AI solutions (17%), model development and evaluation (16%), and responsible and trustworthy AI (15%).

How do the CPMAI phases fit together?

The six phases are Business Understanding, Data Understanding, Data Preparation, Model Development, Model Evaluation, and Model Operationalization. Teams can revisit an earlier phase when new evidence changes what the project needs.

How do I maintain a PMI-CPMAI certification?

Earn and report 30 professional development units (PDUs) per three-year cycle, then pay PMI’s renewal fee. At least 18 PDUs must be from education, including at least four in each PMI Talent Triangle skill area. Up to 12 may come from giving back to the profession.

Same family resources

Explore More PMI Certifications

Continue into nearby exams from the same family. Each card keeps practice questions, study guides, flashcards, videos, and articles in one place.