PMI Certified Professional in Managing AI (PMI-CPMAI)

Last checked October 22, 2026 against the official PMI exam page.

1. Question 1Manage AI Model Development and Evaluation

   A fraud model reports 99.5 percent accuracy on data where only 0.5 percent of transactions are fraud. Which evaluation is most appropriate?

   1. AThe training loss curve alone, because it converges smoothly and indicates the model has learned
   2. BAccuracy alone, because it is the simplest metric to explain to executives reviewing the results
   3. CPrecision, recall, and the cost of missed fraud measured against a simple baseline
   4. DThe number of features used compared with the previous model's feature count from the last release
   Show answer and explanationHide answer

   Correct answer: C

   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.

2. Question 2Operationalize AI Solution

   Which element is most important for monitoring a deployed model over time?

   1. AA one-time accuracy check performed on the day the model is first deployed to production
   2. BA single monthly report of server uptime, with no model-level indicators included in it
   3. CA copy of the training notebook stored in a shared folder for future reference by the team
   4. DMonitoring for data drift, performance decay, and business KPI changes, with retraining triggers defined
   Show answer and explanationHide answer

   Correct answer: D

   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.

3. Question 3Identify Data Needs

   A healthcare analytics team combines records from three hospitals that code diagnoses differently. Which step should come FIRST before any modeling?

   1. AProfile the diagnosis codes and map them to one documented definition
   2. BTrain a baseline model on the largest hospital's data and check the other two hospitals later
   3. CRemove all diagnosis fields so the model cannot learn from inconsistent coding
   4. DChoose the model architecture that handles categorical codes most naturally
   Show answer and explanationHide answer

   Correct answer: A

   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.

4. Question 4Identify Business Needs and Solutions

   A bank's sponsor wants an AI model to cut loan-processing time and says the team should 'use AI for everything in the pipeline.' What should the project manager do FIRST?

   1. AConfirm the business problem, success measures, and whether AI is the right fit compared with simpler automation
   2. BSelect a deep learning framework so the data science team can start building the model immediately
   3. CAsk the vendor for a demo of its generative tool so that the scope is set by what the product can do
   4. DAnnounce a launch date for the model so that funding for the full year is protected before analysis begins
   Show answer and explanationHide answer

   Correct answer: A

   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.

5. Question 5Support Responsible and Trustworthy AI Efforts

   A lending model performs well overall, but its false-rejection rate is clearly higher for one demographic group. What should the team do?

   1. ARemove the demographic field and assume that the disparity disappears from the outcomes
   2. BReport only the overall accuracy to the regulator because it meets the agreed target threshold
   3. CMeasure error rates by group and agree mitigations with governance before release
   4. DRetrain with more epochs so that the disparity shrinks as the model converges further
   Show answer and explanationHide answer

   Correct answer: C

   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.