Support Responsible and Trustworthy AI Efforts
15%of exam
Identify Business Needs and Solutions
26%of exam
Identify Data Needs
26%of exam
Manage AI Model Development and Evaluation
16%of exam
Operationalize AI Solution
17%of exam
Quick Facts
- Exam
- PMI-CPMAI
- Questions
- 120 total, 100 scored, 20 unscored
- Time
- 160 minutes
- Breaks
- None scheduled
- Before scheduling
- Complete PMI's 21-hour prep course
- Prior experience
- None required
- Delivery
- Pearson VUE center or online
- Renewal
- 30 PDUs every 3 years
- Exam outline
- September 2025
Remember Trust Checks
Protect, disclose, test, comply, document
Transparency vs Explainability
Transparency
- Disclose data use and limits
- Document model selection rationale
Explainability
- Describe factors behind an output
- Make reasoning understandable to users
Disclose process, explain an outcome
Choose a Safeguard
- Personal data enters workflow→Assess privacy impact(Minimize collection and restrict access)
- Errors cluster in one group→Check subgroup fairness(Inspect data, model, and algorithm)
- Users ask why outcome changed→Explain decision factors
- Applicable policy changes→Reassess compliance with legal team
- Approval history is missing→Gather records and log gaps
Privacy and Security
- PII
- Personally identifiable information
- Data minimization
- Collect only necessary data
- Privacy impact assessment
- Assess risks to individuals
- Least privilege
- Grant only needed access
- Encryption at rest
- Protect stored data
- Encryption in transit
- Protect data during transfer
- Retention policy
- Set storage and deletion rules
Fairness and Accountability
- Representation check
- Find underrepresented groups in data
- Subgroup evaluation
- Compare errors across relevant groups
- Bias mitigation
- Address source and retest fairness
- Human oversight
- Review and escalate risky decisions
- Compliance monitoring
- Track applicable rules and changes
- Audit trail
- Trace versions, decisions, and approvals
- Chain of custody
- Record who handled which data
Remember the Phase Order
Define, inspect, prepare, build, evaluate, operate
Business KPI vs Model Metric
Business KPI
- Measures operational or business outcome
- Example: average handling time
Model metric
- Measures prediction quality
- Example: classification recall
Check whether model quality delivers value
Scope the AI Project
- Sponsor proposes a model first→Define problem and affected users
- Stable rules solve the problem→Compare conventional automation(Assess cost and achievable value)
- Value claim lacks evidence→Measure current business baseline
- Users resist workflow changes→Test adoption barriers with users
- Early feasibility remains uncertain→Test assumptions through scoped pilot(Agree measurable success criteria first)
- Project scope exceeds available resources→Narrow initial scope
Six CPMAI Phases
- I. Business Understanding
- Define problem, value, and feasibility
- II. Data Understanding
- Identify and assess available data
- III. Data Preparation
- Transform data into model inputs
- IV. Model Development
- Select, train, and tune models
- V. Model Evaluation
- Check agreed criteria before release
- VI. Model Operationalization
- Deploy, monitor, and improve
Business Case and Scope
- Problem statement
- Who is affected and how
- Scope statement
- Boundaries, deliverables, assumptions, constraints
- Feasibility
- Check business, data, implementation viability
- Solution draft
- Outline architecture and data flow
- Resource plan
- Identify skills, compute, and suppliers
- Success criteria
- Baseline, target, and acceptance threshold
- Business case
- Justify investment with evidence
- Return on investment
- Net benefit ÷ investment cost
- Total cost of ownership
- Build, run, and maintain costs
- Adoption plan
- Train users and measure uptake
- Integration risk
- Disruption to systems or workflows
AI Risk Assessment
- Failure mode
- Describe how harm could occur
- Risk exposure
- Assess likelihood and impact
- Risk owner
- Assign responsibility for the response
- Mitigation
- Reduce likelihood or impact
- Contingency trigger
- Define when to activate fallback
Seven AI Patterns
- Hyperpersonalization
- Tailor recommendations to each person
- Conversational and human interaction
- Support natural dialogue or translation
- Recognition
- Identify objects, speech, or handwriting
- Pattern and anomaly detection
- Find structure or unusual behavior
- Predictive analytics and decision support
- Forecast outcomes to inform decisions
- Goal-driven systems
- Optimize actions toward an objective
- Autonomous systems
- Act with limited human intervention
Remember Go/No-Go Decisions
Feasible idea? Suitable data? Ready model?
Data Understanding vs Preparation
Phase II: Understand
- Inventory and profile sources
- Check rights and suitability
Phase III: Prepare
- Clean, label, and transform
- Validate and version prepared inputs
Assess availability, then transform
Resolve Data Gaps
- Unclear what data is needed→Specify target, format, and coverage
- Meaning of fields is unclear→Consult data subject matter expert
- License excludes planned use→Obtain rights or replace source
- Target population is underrepresented→Seek representative samples(Recheck coverage before proceeding)
- Available data misses essential criteria→Record no-go and remediation(Report impact and options)
Data Requirements and Sourcing
- Data requirements
- Specify inputs, labels, and coverage
- Sampling plan
- Specify volume and sampling method
- Granularity
- Required level of detail
- Data subject matter expert
- Clarify source context and meaning
- Source inventory
- Record locations, owners, and formats
- Usage rights
- Verify permitted uses and licensing
- Secure AI workspace
- Provision compute, storage, and access
- Data ingestion
- Extract and verify incoming records
- Data refresh
- Set cadence for updated inputs
Data Quality and Readiness
- Data lineage
- Trace origins and transformations
- Exploratory data analysis
- Profile distributions, gaps, and outliers
- Completeness
- Check required fields and records
- Accuracy
- Values match real-world observations
- Consistency
- Values agree across records or sources
- Freshness
- Data current enough for use
- Representativeness
- Coverage reflects intended use population
- Data sufficiency
- Enough suitable data for purpose
- Readiness report
- Summarize gaps and recommend action
Remember the Data Splits
Fit on train, tune on validation, score on test
Validation vs Test Set
Validation set
- Compare configurations
- Tune hyperparameters
Test set
- Reserve for final evaluation
- Estimate performance on unseen data
Fit preprocessors using training data only
Choose a Modeling Approach
- Predict a known target label→Supervised classification
- Predict a continuous value→Supervised regression
- Find groups in unlabeled data→Unsupervised clustering
- Learn sequential actions from rewards→Reinforcement learning
- Labeled training examples are scarce→Assess transfer learning feasibility(Verify rights and domain fit)
- Training score dwarfs validation score→Investigate overfitting(Check split integrity and complexity)
- Model misses release criteria→Hold deployment and diagnose(Re-evaluate after correction)
Data Preparation
- Data cleaning
- Correct errors and remove duplicates
- Imputation
- Replace missing values with estimates
- Labeling
- Assign target annotations to examples
- Feature engineering
- Create useful model inputs
- Scaling
- Adjust numerical feature magnitudes
- Data augmentation
- Create variants of existing examples
- Reproducible pipeline
- Version transformations and inputs
- Prepared-data quality check
- Recheck quality before model training
Precision vs Recall
Precision
- Of flagged cases, truly positive?
- True positives ÷ all predicted positives
- Prioritize when false alarms hurt
Recall
- Of actual positives, how many flagged?
- True positives ÷ all actual positives
- Prioritize when missed positives hurt
Choose metric by error cost
Model Selection and Training
- Supervised learning
- Learn from inputs with targets
- Unsupervised learning
- Discover structure without target labels
- Reinforcement learning
- Learn actions from reward feedback
- Baseline model
- Set performance reference for comparison
- Training set
- Fit model parameters
- Validation set
- Tune and select models
- Test set
- Estimate final generalization
- Experiment log
- Record data, settings, and results
- Configuration management
- Version models and their settings
Model Readiness vs Deployment
Readiness decision
- Check model against acceptance criteria
- Verify operational prerequisites
Deployment
- Put approved version into production
- Verify live system after release
Decide readiness before rollout
Model Evaluation
- Classification accuracy
- Correct predictions ÷ all predictions
- F1 score
- Harmonic mean of precision and recall
- Mean absolute error
- Average absolute prediction error
- Overfitting
- Fits training data, generalizes poorly
- Data leakage
- Model learns from unavailable information
- Robustness
- Check behavior under expected variations
- Model go/no-go
- Check targets, risks, infrastructure, documentation
Remember Post-release Responsibilities
Verify, monitor, respond, improve
Rollback vs Contingency
Rollback
- Revert to last known-good version
- Use defined rollback triggers
Contingency
- Keep workflows functioning during incidents
- Use recovery, escalation, or fallback
Rollback can support service continuity
Respond in Production
- New release needs limited exposure→Use canary rollout
- Release breaches rollback threshold→Execute planned rollback
- Live input distribution shifts→Investigate drift and verify performance
- Predictive quality falls→Diagnose before retraining
- Major outage interrupts service→Activate incident and recovery plan
- Support takes ownership→Confirm runbook, access, escalation contacts
Rollout and Handover
- Deployment plan
- Schedule, integrate, validate, plan rollback
- Canary rollout
- Route limited traffic to new version
- Post-deploy verification
- Confirm live functionality and performance
- Production access
- Provision approved roles and permissions
- Transition plan
- Transfer ownership, knowledge, and support
- Operations runbook
- Document maintenance and incident procedures
- Disaster recovery
- Restore systems and data
- Final report
- Summarize outcomes and lessons learned
Production Monitoring
- Data drift
- Input distributions shift over time
- Performance degradation
- Predictive quality declines in production
- Training-serving skew
- Training and production inputs differ
- Service health
- Watch latency, throughput, and errors
- Business outcomes
- Track realized savings and adoption
- Alert threshold
- Trigger investigation when breached
- Retraining
- Update model when evidence warrants
- Change control
- Validate and approve model updates
Common Traps
Blueprint vs lifecycle
Five weighted exam domains ≠ Six iterative methodology phases
Accessibility vs permission
Data can be accessed ≠ Data may be used
Volume vs coverage
Many records available ≠ Target population represented
Pattern vs technique
Forecast demand to guide decisions ≠ Train a supervised regression model
Aggregate vs subgroup
High overall accuracy ≠ Acceptable errors across groups
Drift signal vs measured impact
Input distributions shift ≠ Predictive quality declines
Technical quality vs uptake
Model meets test targets ≠ Users adopt redesigned workflow
Last Minute
- 1.Trustworthy AI spans all six phases
- 2.Compare AI with simpler alternatives
- 3.Set targets before model development
- 4.Verify data rights and representativeness
- 5.Recheck prepared data before training
- 6.Tune on validation, reserve test data
- 7.Check subgroup errors before deployment
- 8.Recall matters when misses hurt
- 9.Check business targets before rollout
- 10.Confirm monitoring, ownership, rollback triggers
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.
More From This Family
Videos and articles for deeper review.
