Free PMI Certified Professional in Managing AI (PMI-CPMAI) practice test: 30 questions with answers

With rare positives, accuracy is misleading because a model that never flags fraud would score similarly, so precision, recall and the cost of misses against a baseline are needed. Loss curves and feature counts do not show how well the model finds fraud.

Promotion should rest on evaluation with a held-out test set and on the acceptance criteria agreed in business understanding. Removing the validation set, tuning on production test data and relying on a demo each make the result unreliable.

Selecting on the test set means it no longer estimates performance on unseen data, which inflates the reported score. A validation set or cross-validation should guide selection, and the test set should be reserved for a final check.

Accuracy and latency are both requirements, so the team should agree on the trade-off with stakeholders and then retest both together. Dropping a requirement or deploying a known breach to watch it fails to meet the agreed constraints.

Cross-validation and separate holdout data estimate generalization, and removing leaking features stops the model from learning shortcuts. Training longer, choosing on training accuracy and repeatedly testing on the test set all push the model toward overfitting or optimistic results.

Recording data version, code, parameters and seeds lets any run be repeated and compared. Personal laptops, overwritten files and chat threads leave no reliable record of what produced a given result.

Production monitoring tracks drift, performance decay and business outcomes, with triggers that start investigation or retraining. A one-time check, uptime alone and an archived notebook do not reveal when a model starts to fail.

A rollback plan and named owners ensure a bad release can be reversed and that someone is accountable for changes and incidents. Announcements, naming and long-term capacity reservations do not protect the service during operation.

Automated retraining needs gates that check each candidate against the current model before it is promoted, so a worse model cannot silently replace a better one. Automatic promotion without checks, date-named files and hidden schedules all weaken control.

Low adoption usually reflects workflow fit, timing or interface problems, so the team should observe the work and change the design while measuring adoption. Retraining with more layers or sending emails treats the symptom, and removing the feature abandons the investment.

A handoff needs the exact versioned artifact with its data reference, and monitoring thresholds with an incident contact so operations can act. Marketing plans, speculative features and shared working notebooks are not operational handoff elements.

Retraining should follow measured performance or drift against agreed thresholds, with someone accountable for acting on the trigger. Fixed calendars, staffing changes and library releases are not evidence that the model has degraded.

Profiling the codes and mapping them to one documented definition resolves the inconsistency at the data level, which every later model depends on. Choosing an architecture or training on one hospital first builds on flawed inputs, and dropping the fields throws away useful clinical signal.

When inputs in production no longer resemble the training data, the model's assumptions break, which is data drift. Storage, normalization and identifier counts are real issues, but they do not explain a change in the customer population over time.

Missing labels can bias a rare-class model, so the cause must be understood and the treatment applied under documented rules. Imputing the majority class or dropping rows hides the bias, and random synthetic labels create noise the model will learn.

A feature that is only known after the outcome leaks the answer into training, so offline metrics look strong and production results disappoint. Training time, scaling and class counts are not the central problem with this feature.

Data minimization, pseudonymization and purpose-limited access reduce exposure while still supporting the model. Copies on laptops, broad shared access and hiding data only from documentation each increase risk without protecting the data itself.

A data inventory or data sheet records ownership, lineage and permitted uses for each source, which governance and privacy reviews rely on. A confusion matrix, GPU reservations and tuning logs describe modeling activity rather than the data sources.

The first step is to confirm the business problem and measurable outcome, and to test whether AI adds value over rules or simpler automation. Choosing a framework, scoping around a vendor demo or fixing a date all happen before anyone knows what success looks like.

CPMAI is iterative: each phase feeds learning back into earlier decisions, so business goals and data findings are revisited as the work progresses. A waterfall that freezes requirements early, a purely technical start and a one-time build all miss that cycle.

A model that is accurate but unused usually failed on business fit: the decision it supports and the success criteria were never agreed with the users. Tuning, data volume and latency may matter, but they do not explain a model that managers ignore despite high accuracy.

A project charter captures the business objectives, KPIs with baselines, and constraints that define success before modeling begins. A model card, data dictionary and runbook are useful later but do not set the agreed targets.

If the business case no longer holds or the data cannot support the target outcome, stopping early is the sound decision. Provider preference, an unreviewed architecture change and a small extra training delay are not grounds to stop.

Success criteria against a baseline and an agreed decision change are defined during business understanding and guide everything after. Endpoint files, hyperparameters and cost estimates belong to later phases of the project.

Fairness has to be measured by group, with the cause investigated and mitigations agreed through governance before release. Removing a field often leaves proxies in place, and overall accuracy or extra training can hide rather than fix the disparity.

A plain-language explanation of the key factors, paired with a route to human review, gives people understanding and recourse. Publishing code, hiding the model type and exposing raw weights do not explain a specific decision to the person affected.

Grounding answers in approved sources with citations and routing uncertain cases to a person addresses the root cause, which is answers without evidence. Higher temperature increases variation, removing the knowledge base makes answers less reliable, and disabling logs removes the means of auditing errors.

A responsible AI assessment examines risks such as privacy, fairness, safety and accountability for the specific use, and approval depends on it. Marketing plans, cost forecasts and naming histories are useful project records but are not part of the risk assessment.

Recording consent and purpose keeps collection lawful and within scope, and representative sampling reduces the risk of unfair performance. Collecting everything, reusing data without review and deleting lineage each weaken privacy or accountability.

A model card is designed to document purpose, data sources, limitations and evaluation in one place, which is what a regulator needs. Burndown charts, utilization reports and region lists record delivery or infrastructure facts rather than the model's behavior.