PMI-CPMAI — PMI Certified Professional in Managing AI Quick Reference

Last revised: June 29, 2026

Compact PMI-CPMAI reference for AI project lifecycle, governance, risk, data, modeling, deployment, and exam decision points.

This Quick Reference is for candidates preparing for the PMI Certified Professional in Managing AI (PMI-CPMAI) exam from PMI, exam code PMI-CPMAI. Use it to connect project management judgment with AI-specific lifecycle, data, model, governance, and operational risks.

AI Project Lifecycle at a Glance

AI initiatives are not managed like ordinary software delivery. The project manager must handle uncertainty in data, model performance, experimentation, governance, stakeholder trust, and production monitoring.

Lifecycle area	Manager focus	Typical artifacts	Exit evidence	Common exam trap
Business understanding	Define the business problem, value, decision context, and measurable outcome	Problem statement, business case, success metrics, stakeholder map, use-case canvas	Clear target outcome, success criteria, constraints, acceptance threshold	Starting model selection before validating the business need
Data understanding	Determine what data exists, who owns it, whether it is fit for purpose	Data inventory, data lineage, data quality profile, access plan, privacy/security review	Data sources identified, quality risks known, permissions understood	Assuming available data is usable data
Data preparation	Clean, transform, label, enrich, engineer, and split data	Data pipeline, labeling guide, feature list, training/validation/test split, data version	Reproducible prepared data with documented assumptions	Creating leakage by using future or target-derived data
Model development	Select approach, train, tune, compare, and document models	Experiment log, model candidates, hyperparameters, feature importance, baseline comparison	Candidate model meets technical targets in controlled evaluation	Treating the first high-scoring model as production-ready
Model evaluation	Validate against business, risk, fairness, security, explainability, and operational criteria	Evaluation report, bias assessment, model card, risk register update, go/no-go decision	Model accepted by business and governance stakeholders	Optimizing only accuracy while ignoring business impact
Operationalization	Deploy, monitor, support, retrain, and manage change	Deployment plan, MLOps pipeline, rollback plan, monitoring dashboard, retraining trigger	Production model monitored with ownership and incident process	Assuming deployment is the end of the AI lifecycle
Continuous governance	Manage drift, compliance, auditability, accountability, and responsible AI controls	Governance log, approval records, audit trail, model inventory, incident register	Model remains controlled, explainable, and aligned to value	Leaving governance to technical teams only

Core PMI-CPMAI Distinctions

Distinction	Know this for the exam
AI project vs software project	AI delivery depends on data quality and experimental model performance, not only requirements and coding. Scope, schedule, and acceptance criteria may need progressive elaboration.
Model output vs business outcome	A statistically strong model is not automatically valuable. Link model metrics to measurable business decisions, cost, risk, adoption, or customer outcomes.
Training vs inference	Training builds or tunes the model using data. Inference applies the trained model to new inputs in production or test environments.
Algorithm vs model	An algorithm is the method. A model is the trained artifact produced by applying an algorithm to data.
Parameter vs hyperparameter	Parameters are learned during training. Hyperparameters are set before or during experimentation, such as learning rate or tree depth.
Feature vs label	Features are input variables. Labels are target outputs used for supervised learning.
Validation vs test set	Validation supports model selection/tuning. Test data should provide an unbiased final estimate before deployment.
Model drift vs data drift	Data drift means input data distribution changes. Model drift/concept drift means the relationship between inputs and target outcome changes.
Explainability vs interpretability	Interpretability is inherent understandability. Explainability uses techniques to explain complex models.
Automation vs augmentation	Automation replaces a task or decision. Augmentation supports a human decision maker. Human-in-the-loop needs roles, thresholds, and escalation paths.
Proof of concept vs pilot vs production	POC proves feasibility. Pilot validates in limited real conditions. Production requires monitoring, controls, support, and ownership.

AI Use-Case Pattern Reference

PMI-CPMAI candidates should be able to classify AI opportunities and connect the pattern to data, risks, and success metrics.

AI pattern	Use when the problem is…	Example outcomes	Useful metrics	Key risks
Predictive analytics and decision support	Forecasting, scoring, ranking, or recommending a decision	Churn risk, demand forecast, loan risk, maintenance prediction	RMSE, MAE, AUC, precision/recall, business lift	Bias, false positives/negatives, overreliance
Recognition	Identifying objects, people, speech, text, images, or signals	Image classification, speech-to-text, document extraction	Accuracy, F1, word error rate, detection rate	Poor data quality, edge cases, demographic bias
Conversational and human interaction	Interacting with users in natural language or multimodal channels	Chatbot, service assistant, voice interface	Task completion, containment, satisfaction, hallucination rate	Safety, hallucination, escalation failure
Hyperpersonalization	Tailoring content, offers, recommendations, or experiences	Product recommendation, dynamic pricing support, next-best action	Conversion, lift, engagement, retention	Privacy, filter bubbles, unfair targeting
Pattern and anomaly detection	Finding unusual behavior, clusters, outliers, or hidden structures	Fraud signal, network anomaly, process deviation	Detection rate, false alarm rate, precision, investigation yield	Alert fatigue, shifting baselines
Autonomous systems	Acting with limited human intervention in dynamic environments	Robotics, autonomous routing, adaptive control	Safety incidents, task success, latency, reliability	Safety, accountability, fail-safe design
Goal-driven systems	Optimizing actions toward a goal under constraints	Resource optimization, game agents, planning systems	Reward, constraint violations, optimization value	Misaligned reward, unintended behavior

Role and Responsibility Matrix

Role	Primary accountability	PMI-CPMAI exam signal
Sponsor	Owns business value, funding, strategic alignment, executive decisions	Escalate unresolved business tradeoffs here, not to the data scientist
Project manager / AI project manager	Integrates scope, schedule, risk, stakeholders, governance, and delivery lifecycle	Coordinates uncertainty, decisions, dependencies, and transparency
Product owner / business owner	Prioritizes use cases, accepts outcomes, clarifies decision workflows	Needed when requirements or value criteria are ambiguous
Data owner / data steward	Authorizes access, defines meaning, quality expectations, and usage constraints	Critical for permissions, lineage, quality, and retention issues
Subject matter expert	Validates business rules, labels, edge cases, and operational usefulness	Reduces misunderstood context and labeling errors
Data engineer	Builds data pipelines, transformations, storage, and reproducibility	Needed when data is fragmented, stale, or not production-ready
Data scientist / ML engineer	Explores data, engineers features, trains, tunes, and evaluates models	Does not own business value alone
MLOps / platform engineer	Deploys, monitors, automates, versions, and supports models in production	Critical when moving from experiment to reliable service
Security / privacy / compliance	Reviews controls for sensitive data, access, audit, model risk, and obligations	Involve early, not after model completion
Risk / governance board	Reviews responsible AI, model approval, exceptions, and ongoing controls	Used for high-impact, regulated, sensitive, or material decisions
Change manager / adoption lead	Manages user readiness, communications, training, and resistance	Especially important when AI changes work processes
Model owner	Accountable for model performance and control after go-live	Prevents orphaned models after project closure

Business Case and Use-Case Selection

High-Yield Evaluation Criteria

Criterion	Ask	Good evidence	Weak evidence
Business value	What decision, process, or outcome improves?	Quantified benefit, cost avoidance, risk reduction, service improvement	“We need AI” or “competitors use AI”
Feasibility	Can the model be built, integrated, and supported?	Available skills, architecture, data access, realistic performance target	Assumption that technology alone solves the problem
Data readiness	Is relevant, representative, permitted data available?	Data profiling, lineage, owner approval, quality assessment	Spreadsheet sample with no provenance
Risk profile	What could harm customers, employees, operations, or the organization?	Risk register, impact analysis, governance path	Risk deferred until deployment
Adoption	Will users trust and use the output correctly?	Human workflow design, explanation, training, feedback loop	“Users will adopt it if it is accurate”
Measurability	Can success be measured objectively?	Baseline, target metric, measurement plan	Vague success definition

Practical Prioritization Formula

Use scoring to compare candidate AI use cases. The exam may not require a specific formula, but it often rewards risk-aware prioritization.

\[ \text{Priority Score} = \frac{\text{Business Value} + \text{Strategic Alignment} + \text{Learning Value}} {\text{Complexity} + \text{Data Risk} + \text{Operational Risk}} \]

Interpretation: higher value and learning increase priority; high complexity and risk reduce priority unless justified by strategic importance.

Data Management Quick Reference

Data concept	What to verify	Common control
Data provenance	Where data came from and how it was collected	Lineage documentation, source approval
Data ownership	Who can authorize use	Data owner/steward approval
Data quality	Completeness, accuracy, validity, consistency, uniqueness, timeliness	Profiling, cleansing rules, quality thresholds
Representativeness	Whether data reflects the target population and future use	Sampling review, bias analysis
Ground truth	Reliable target labels or outcomes	Labeling guide, SME review, inter-rater checks
Labeling quality	Consistency and correctness of annotations	Label audit, adjudication process
Sensitive data	Personal, confidential, regulated, or proprietary data	Minimization, masking, access control
Feature engineering	Transforming raw data into useful predictors	Feature documentation, reproducible pipeline
Data leakage	Information in training that would not be available at prediction time	Time-aware split, feature review
Data versioning	Ability to reproduce results	Dataset versions, pipeline hash, experiment tracking
Synthetic data	Artificially generated data used to supplement or protect real data	Validation against target distribution, risk review
Retention	How long data and outputs are kept	Retention schedule and deletion process

Data Split Decision Table

Scenario	Prefer	Why
Randomly distributed independent records	Random train/validation/test split	Simple and usually sufficient
Time-series or forecasting	Time-based split	Prevents future data from leaking into training
Rare positive cases	Stratified split	Preserves class balance in each set
Same customer/patient/device appears multiple times	Grouped split	Prevents the same entity from appearing in both train and test
Small dataset	Cross-validation, with caution	Improves estimate stability but does not solve poor data quality
Production data differs from historical data	Back-testing and pilot monitoring	Tests real-world stability

Model Approach Selection

Approach	Best fit	Examples	Watch for
Supervised learning	Labeled examples exist and target outcome is known	Classification, regression	Label quality, leakage, imbalance
Unsupervised learning	No labels; goal is grouping, anomaly, or structure discovery	Clustering, anomaly detection, topic discovery	Harder validation, business interpretation
Semi-supervised learning	Limited labels with larger unlabeled data	Document classification, image labeling	Label propagation errors
Reinforcement learning	Agent learns actions through rewards and penalties	Optimization, robotics, simulations	Safety, reward misalignment, simulation gap
Deep learning	Large complex data such as images, speech, language, sequences	Computer vision, NLP, generative models	Compute cost, explainability, data volume
Generative AI / foundation model	Creating or transforming text, code, images, audio, or multimodal content	Summarization, drafting, assistants	Hallucination, prompt injection, IP, evaluation difficulty
Rules-based automation	Logic is stable, explicit, and deterministic	Eligibility rules, routing logic	Not adaptive; may be better than AI for simple rules
Human-in-the-loop AI	Decision impact is high or context is complex	Medical support, fraud investigation, hiring support	Role clarity, escalation, accountability

Build, Buy, or Adapt Decision

Option	Choose when	Advantages	Risks
Build custom model	Need differentiated capability, control, special data, or integration	Tailored performance and ownership	Higher cost, skills, support burden
Buy vendor AI solution	Common use case, faster time to value, vendor support acceptable	Speed, packaged features	Lock-in, opaque model, data sharing risk
Adapt/fine-tune foundation model	Need language or multimodal capability with domain adaptation	Strong baseline, faster than building from scratch	Evaluation, hallucination, privacy, model updates
Use API/prompting only	Low customization, rapid experimentation, moderate risk	Fastest start	Less control, variable behavior
Do not use AI	Rules or reporting solve the problem adequately	Lower risk and complexity	May not capture complex patterns

Evaluation Metrics Reference

Classification Metrics

Metric	Plain formula	Use when	Trap
Accuracy	Correct predictions / all predictions	Classes are balanced and errors have similar cost	Misleading with imbalanced data
Precision	TP / (TP + FP)	False positives are costly	May miss many actual positives
Recall / sensitivity	TP / (TP + FN)	False negatives are costly	May create too many false alerts
Specificity	TN / (TN + FP)	Correctly identifying negatives matters	Often ignored in screening use cases
F1 score	2 × precision × recall / (precision + recall)	Need balance between precision and recall	Hides business cost differences
AUC-ROC	Ranking quality across thresholds	Comparing classifiers broadly	Can be misleading with severe imbalance
Confusion matrix	TP, FP, TN, FN counts	Explaining error types to stakeholders	Do not stop at one aggregate score

Regression, Forecasting, and Ranking Metrics

Metric	Use when	Lower or higher is better	Trap
MAE	Average absolute error is understandable	Lower	Treats all errors linearly
RMSE	Larger errors should be penalized more	Lower	Sensitive to outliers
MAPE	Percentage error is meaningful	Lower	Fails or distorts near zero values
R-squared	Explaining variance in target	Higher	Does not prove business usefulness
Lift / gain	Ranking improves targeting vs baseline	Higher	Needs baseline comparison
Business KPI	Model affects revenue, cost, cycle time, risk, service	Depends	Must be measured after deployment

Generative AI Evaluation

Evaluation area	Ask	Example evidence
Factuality	Are outputs correct and grounded?	Human review, retrieval grounding tests, citation checks
Relevance	Does output answer the user’s task?	Rubric scoring, task completion
Safety	Does it avoid harmful, prohibited, or sensitive output?	Red-team tests, policy checks
Hallucination	Does it invent unsupported facts?	Grounded evaluation, refusal tests
Toxicity and bias	Are outputs discriminatory or harmful?	Bias test suites, reviewer audits
Robustness	Does it resist prompt injection and adversarial inputs?	Attack simulations, guardrail tests
Consistency	Are similar prompts handled consistently?	Regression test set
Cost and latency	Can it meet operational constraints?	Token cost, response time, throughput tests

Risk, Governance, and Responsible AI Controls

Risk exposure is often summarized as:

\[ \text{Risk Exposure} = \text{Probability} \times \text{Impact} \]

For AI projects, include both project delivery risk and model behavior risk.

Risk	Warning signs	PM action	Evidence/artifact
Bias or unfair impact	Unequal errors across groups, nonrepresentative data	Perform bias assessment, involve SMEs/governance, adjust data/model/process	Fairness report, mitigation log
Privacy misuse	Sensitive data used without clear purpose or approval	Apply minimization, masking, access controls, privacy review	Data use approval, privacy assessment
Security risk	Exposed model endpoints, weak access, prompt injection	Involve security early, threat model, test controls	Security review, penetration/red-team results
Hallucination	Model generates plausible unsupported content	Ground outputs, add human review, define refusal/escalation	Evaluation report, guardrail design
Explainability gap	Stakeholders cannot understand or challenge decisions	Select interpretable model or explainability technique; document limitations	Model card, explanation method
Model drift	Production performance degrades over time	Monitor data/performance, set retraining triggers	Drift dashboard, retraining plan
Data drift	Incoming data differs from training data	Monitor input distributions and quality	Data monitoring report
Automation bias	Users over-trust model output	Train users, design confidence indicators, require review for high impact	Adoption plan, UI decision controls
Accountability gap	No owner after deployment	Assign model owner and support process before go-live	RACI, operations handbook
Third-party/vendor risk	Black-box model, unclear data use, dependency risk	Due diligence, contract review, exit plan	Vendor assessment, SLA/support terms
IP/copyright risk	Unclear rights for training data or generated content	Legal/IP review and content policy	Usage policy, approval record
Safety risk	AI action may cause physical, financial, or human harm	Fail-safe design, simulation, staged rollout, human override	Safety case, rollback plan

Governance Artifacts to Recognize

Artifact	Purpose	Created/updated when
AI use-case canvas	Captures business problem, users, data, value, and risks	Initiation and refinement
Data inventory	Lists sources, owners, sensitivity, quality, and access status	Data understanding
Data sheet / dataset documentation	Documents dataset origin, contents, limitations, and intended use	Before model development and reuse
Model card	Summarizes model purpose, training data, performance, limitations, ethical considerations	Before approval and deployment
Experiment log	Tracks runs, parameters, datasets, and results	During model development
Risk register	Tracks project, data, model, governance, and operational risks	Throughout lifecycle
Decision log	Records major tradeoffs and approvals	When selecting use case, data, model, threshold, deployment
Monitoring plan	Defines metrics, thresholds, alerts, owners, and retraining triggers	Before production
Change management plan	Prepares users, process changes, training, communications	Before pilot and rollout
Incident response plan	Defines what to do when AI fails, drifts, or harms	Before go-live

Project Management Decision Tables

What Should the Manager Do Next?

Situation	Best next action	Why
Stakeholders ask for “an AI solution” but cannot define the decision or value	Facilitate business understanding and define measurable outcomes	AI selection should follow problem definition
Data scientists want to start modeling before data access is approved	Resolve data ownership, permission, and governance	Prevents rework and compliance issues
Model accuracy is high but false negatives are costly	Review confusion matrix and adjust metric/threshold with business owners	Business impact depends on error type
Test performance is much worse than training performance	Investigate overfitting, data leakage, split strategy, and data representativeness	Technical score must generalize
Sponsor wants full rollout after a successful lab demo	Recommend pilot or staged deployment with monitoring and rollback	Lab success does not prove operational readiness
Users distrust model recommendations	Add explanation, training, feedback channels, and human review design	Adoption risk is a project risk
Model behavior changes after deployment	Trigger monitoring response: assess drift, impact, rollback/retrain if needed	AI requires lifecycle management
Compliance team raises concerns late	Pause affected work, assess impact, update risk and governance plan	Governance cannot be bypassed for schedule
Vendor model is opaque	Conduct vendor risk review and define transparency, audit, and performance requirements	Black-box risk must be managed
Product owner keeps changing target metric	Reconfirm business objective and update change control/backlog priority	Prevents uncontrolled experimentation

Predictive, Agile, or Hybrid Delivery

Delivery approach	Best fit	AI-specific tailoring
Predictive	Stable compliance, procurement, infrastructure, or deployment constraints	Still allow experimentation within planned stages
Agile	Uncertain model approach, evolving requirements, iterative discovery	Use spikes, experiments, model backlog, frequent stakeholder review
Hybrid	Most AI initiatives: governed phases plus iterative modeling	Stage gates for data/model/governance; agile iterations inside phases
Kanban/flow	Operational model monitoring, incident handling, labeling queues	Manage WIP, service levels, and continuous improvement
Experimental POC	Feasibility unknown	Timebox, define learning goals, avoid production promises

Change Control in AI Projects

Change type	Examples	Manage through
Business objective change	New target outcome, different user group	Sponsor/product owner approval; update business case
Data change	New source, changed definition, quality issue	Data owner approval; update lineage and tests
Feature change	Add/remove predictors	Experiment log; validation and leakage review
Model change	New algorithm, retraining, fine-tuning	Model governance, evaluation, versioning
Threshold change	Adjust fraud alert cutoff or risk score	Business owner decision; document error tradeoff
Deployment change	New integration, scaling, endpoint, user workflow	Release plan, security review, rollback
Governance exception	Use of sensitive data, reduced explainability, high-risk automation	Formal risk acceptance or rejection

MLOps and Operationalization Workflow

    flowchart LR
	    A[Business objective] --> B[Data access and profiling]
	    B --> C[Data preparation and feature pipeline]
	    C --> D[Model training and experiments]
	    D --> E[Validation and risk review]
	    E --> F{Go / no-go}
	    F -- No --> B
	    F -- Yes --> G[Deploy to pilot or production]
	    G --> H[Monitor performance, drift, cost, safety]
	    H --> I{Trigger?}
	    I -- Drift or incident --> J[Rollback, retrain, or remediate]
	    J --> E
	    I -- Stable --> H

Production Readiness Checklist

Area	Must be true before go-live
Ownership	Sponsor, product owner, model owner, support owner identified
Data	Source permissions, quality checks, lineage, and versioning established
Model	Performance, limitations, bias, explainability, and thresholds documented
Integration	Interfaces, latency, reliability, and fallback paths tested
Security	Access controls, endpoint protection, secrets management, and testing complete
Monitoring	Metrics, alerts, drift checks, cost tracking, and review cadence defined
Human process	Review, override, escalation, and feedback loops designed
Change control	Retraining, model updates, and rollback procedures defined
Auditability	Decisions, versions, approvals, and evidence retained as required by the organization
Adoption	Training, communications, and user readiness completed

Thresholds, Tradeoffs, and Human-in-the-Loop

Decision point	If you increase threshold	If you decrease threshold	Manager focus
Fraud alert threshold	Fewer alerts, higher precision, more missed fraud	More alerts, higher recall, more false alarms	Balance investigation capacity and loss tolerance
Medical screening sensitivity	Fewer false positives but more missed cases	More detected cases but more unnecessary follow-up	Align with clinical risk and governance
Content moderation threshold	Less overblocking, more harmful content may pass	More blocking, more legitimate content removed	Align with policy, appeal process, safety
Credit/risk cutoff	Fewer approvals, lower default risk	More approvals, higher default risk	Ensure fairness, explainability, and business policy
Human review trigger	Less human workload, more automation risk	More oversight, slower process	Match review level to impact and uncertainty

Quality and Acceptance Criteria

AI Quality Dimensions

Dimension	Meaning	Evidence
Technical performance	Model meets metric targets on appropriate data	Evaluation results, confidence intervals where useful
Business performance	Model improves real KPI or decision quality	Pilot results, A/B test, operational KPI
Robustness	Handles edge cases, noise, and adversarial conditions	Stress tests, red-team tests
Fairness	Performance and impact are acceptable across relevant groups	Bias analysis, mitigation actions
Explainability	Users and reviewers can understand enough to trust and challenge output	Explanations, documentation, training
Reliability	Service works consistently under expected load	Uptime, latency, error rate
Maintainability	Model and pipeline can be updated and reproduced	Versioning, automation, documentation
Compliance/governance	Meets organizational policies and approval requirements	Review records, audit trail

Definition of Done for AI Work

Work item	Done means…
Data source onboarded	Approved, documented, profiled, secured, and accessible
Labeling completed	Guide followed, quality checked, conflicts resolved
Model candidate trained	Reproducible run with dataset/model version and metrics
Model selected	Compared to baseline and alternatives; tradeoffs documented
Risk review completed	Material risks assessed, mitigations accepted or required
Pilot completed	Real users/workflow tested; feedback and metrics reviewed
Production release completed	Monitoring, support, rollback, and ownership active

Common Exam Traps

Trap	Better answer pattern
Choosing an algorithm first	Define problem, value, data, and constraints first
Equating high accuracy with success	Evaluate business impact, error costs, fairness, and adoption
Ignoring data ownership	Confirm access, permissions, lineage, and stewardship
Treating AI as a one-time deliverable	Plan monitoring, retraining, governance, and lifecycle ownership
Deferring ethics and compliance	Integrate responsible AI and risk controls early
Assuming more data is always better	Prefer relevant, representative, high-quality, permitted data
Deploying directly from POC	Use pilot, production readiness checks, rollback, and monitoring
Letting technical teams own all decisions	Business owners decide value and error tradeoffs; governance decides risk acceptance
Using a black-box model without justification	Match explainability to impact, risk, and stakeholder need
Ignoring user workflow	Design how humans receive, challenge, override, and improve AI output

Rapid Review Checklist

Before the exam, be able to answer these quickly:

What business decision or process does the AI system improve?
Which AI pattern fits the use case?
What data is needed, who owns it, and is it fit for purpose?
What could make the model unfair, unsafe, insecure, or noncompliant?
Which metric matches the business cost of errors?
What is the baseline, and how does the model improve on it?
How will stakeholders validate, accept, and adopt the solution?
What governance artifacts are needed before deployment?
Who owns the model after project closure?
How will production performance, drift, incidents, and retraining be managed?

Practical Next Step

Use this Quick Reference to identify weak areas, then move into PMI-CPMAI-style scenario practice: for each question, decide the lifecycle phase, the key risk or decision, the responsible role, and the most appropriate next project management action.

Scenario Guide

Support Responsible and Trustworthy AI Efforts