Free CPMAI Practice Questions

Question 1

Invoice processing has two distinct tasks: transcribing fields from scans in hundreds of supplier layouts, then routing invoices according to complete written rules. Transcription is the bottleneck; routing requires no discretion once the fields are verified. The sponsor wants faster processing without changing the rules. Which proposed design uses AI where it adds value?

Accepted Answer

Recognize invoice fields, then apply the written routing rules to the verified values.

Answer

Write template-specific extraction scripts, then train a classifier to decide the destination of every processed invoice.

Answer

Ask a language model to extract fields and infer the routing destination.

Answer

Learn destinations directly from historical invoices and previous staff routing decisions.

Question 2

An online retailer plans to score return risk when an order is paid. In a retrospective extract, support_contact_count includes every contact from payment through the end of the 30-day return window. It is the strongest predictor. The live checkout service can supply only contacts recorded before payment. What requirement must change before the team can rely on the development results?

Accepted Answer

Define features using information available at payment and rebuild the historical extract to match that timing.

Answer

Keep the existing feature definition, split the extract chronologically, and reserve the most recent orders for final evaluation.

Answer

Preserve the full-window contact count and apply stronger regularization so the model becomes less dependent on that predictor.

Answer

Replace unavailable checkout counts with the training median so that production records conform to the historical dataset schema.

Question 3

Two classifiers were evaluated on the same 1,000 payments with independently confirmed outcomes. Model A missed 12 fraudulent payments and incorrectly flagged 20 legitimate payments. Model B missed 4 fraudulent payments and incorrectly flagged 70 legitimate payments. A missed fraud costs $500; an incorrect flag costs $20. Both models meet all release requirements; other costs are equal. For comparable future batches, which selection minimizes the projected cost of these errors?

Accepted Answer

Model B: $3,400 per batch, even though it makes more classification errors.

Answer

Model A: $6,400 per batch, because its smaller error count makes it the better choice.

Answer

Model A: $6,000 per batch, excluding incorrectly flagged legitimate payments.

Answer

Model B: $2,000 per batch, counting the missed-fraud losses alone.

Question 4

A fairness review of an account-security model reports the following for two language groups: confirmed takeovers detected, 92% in both groups; legitimate logins incorrectly flagged, 3% in Group A and 14% in Group B. Samples are large and representative. A manager writes, "Equal detection rates settle the fairness assessment." What has that interpretation overlooked?

Accepted Answer

Matching true-positive rates do not remove the greater false-positive burden on legitimate users in Group B.

Answer

Group B must have a higher false-negative rate because more of its legitimate logins are flagged.

Answer

Language is established as the cause of the disparity and should be removed as a model feature.

Answer

Equalized odds requires matching detection rates, leaving the other difference outside the fairness assessment.

Question 5

For six weeks, a replenishment model has scored live orders without changing any purchase decisions. Its technical, security, and model-performance checks have passed. Operations has a tested rollback, and the business has approved a limited live pilot to measure stockouts and excess inventory. Which deployment step would produce the missing evidence while keeping exposure bounded?

Accepted Answer

Let the candidate direct replenishment for a small eligible group, monitoring outcomes and rollback triggers.

Answer

Substitute planner estimates of stockout reductions for a live pilot.

Answer

Switch all warehouses to the candidate, retaining the incumbent environment for recovery if business outcomes deteriorate.

Answer

Keep recommendations hidden from planners and extend shadow scoring to measure its effect on stockouts.

Question 6

The target is cancellation within 60 days of signup. A proposed training extract labels every account without a recorded cancellation as negative, including accounts opened yesterday. Older cohorts have complete follow-up and adequate, representative samples. For a binary classifier with this target, how should the team proceed?

Accepted Answer

Use cohorts with complete 60-day follow-up for training and evaluation; keep the newer, unresolved outcomes out of the negative class.

Answer

Use every account, keeping recent noncancellations negative and adding account age as a predictor.

Answer

Keep the recent negative labels but give them less weight to compensate for shorter follow-up.

Answer

Suspend model development until all accounts in the extract complete the full follow-up period.

Question 7

An AI assistant requires $80,000 to implement and $20,000 a year to operate. At full use, it would produce $125,000 in annual gross cash savings. A pilot supports 60% sustained adoption, with savings proportional to adoption and operating costs unchanged. Finance defines two-year ROI as net benefit divided by total project cost, with implementation counted once and no discounting or residual value. Which result belongs in the revised business case?

Accepted Answer

25% ROI, after adjusting savings for adoption and including both years of operating costs.

Answer

125% ROI, expressing the expected two-year cash savings as a percentage of the total project cost.

Answer

108.3% ROI, using the assistant's full-use savings against the complete cost of the two-year project.

Answer

87.5% ROI, treating the implementation outlay as the investment base for the entire two-year evaluation period.

Question 8

Customer purchase histories contain random identifiers instead of names. A controlled lookup table retained by the same organization can reconnect each identifier to a customer. The modeling team cannot access that table. A proposal would remove personal-data safeguards from the histories. How should the data be classified when assessing that proposal?

Accepted Answer

Pseudonymized: the retained mapping can reconnect purchase records to individual customers.

Answer

Aggregated: replacing names with identifiers combines the customer-level purchases into nonidentifiable group information.

Answer

Anonymous: the modeling team has no access to the lookup table.

Answer

Encrypted: random identifiers protect the names by encoding them during modeling.

Question 9

A service organization combines two case systems to predict successful resolution. In one system, 'closed' means the customer confirmed a solution. In the other, it also includes cases closed after the customer stopped responding. Both extracts are complete and pass schema checks. The analyst proposes using 'closed' as the positive label in both. Who should resolve the issue, and what must they establish?

Accepted Answer

Service-domain experts and the data steward should agree on a common outcome definition.

Answer

Quality reviewers should accept both labels once their sources meet an identical completeness target.

Answer

Data scientists should retain both label meanings and add a source-system identifier as a feature.

Answer

Data engineers should standardize the status codes to one numeric encoding.

Question 10

Review this validation plan for a vision model intended to inspect previously unseen manufactured parts: photograph each part several times; create rotated copies of every photograph; shuffle all images; split into training and test sets. The resulting test accuracy is excellent. Which revision should precede a new evaluation on unseen parts?

Accepted Answer

Assign physical parts to partitions first, keep all images of each part together, and augment only the training partition.

Answer

Keep the image-level split but stratify it to match defect prevalence across partitions.

Answer

Remove exact duplicate files, then randomly split the remaining images into training and test sets.

Answer

Group rotations with their original photograph, while splitting different photographs of the same part independently.

Free CPMAI Practice Questions

Domain 1: Support Responsible and Trustworthy AI Efforts 15% of exam

Question 1

Question 2

Domain 2: Identify Business Needs and Solutions 26% of exam

Question 3

Question 4

Domain 3: Identify Data Needs 26% of exam

Question 5

Question 6

Question 7

Domain 4: Manage AI Model Development and Evaluation 16% of exam

Question 8

Question 9

Domain 5: Operationalize AI Solution 17% of exam

Question 10

That's 10 of 1,030