Study guide
Oversee Model Governance in Production
PMI Certified Professional in Managing AI PMI-CPMAI
Track your progress as you read
Finished sections get a check mark.
1 Chapter 1: PMI-CPMAI Exam Foundations & the CPMAI Methodology
The PMI-CPMAI Exam: Format, Eligibility, Fees & Scoring The CPMAI Methodology: Six Phases and Data-Centric Iteration CPMAI vs Waterfall, Agile, and CRISP-DM
2 Chapter 2: AI and Machine Learning Fundamentals for CPMAI
The Seven Patterns of AI Supervised, Unsupervised & Reinforcement Learning Why AI Projects Fail: Value, Data Readiness & Right-Sizing
3 Chapter 3: Domain I — Support Responsible and Trustworthy AI Efforts (15%)
Oversee the Privacy and Security Plan Manage AI/ML Transparency in Data and Algorithm Selection Conduct Bias Checks on Model, Data, and Algorithm Monitor Regulatory and Policy Compliance Manage Accountability Documentation and the Audit Trail
4 Chapter 4: Domain II — Identify Business Needs and Solutions, Part 1 (26%)
Identify the Problem to Be Solved Evaluate Initial AI Feasibility Conduct Security, Safety, and Ethics Risk Assessments Develop the AI Project Scope Statement Determine Return on Investment
5 Chapter 5: Domain II — Identify Business Needs and Solutions, Part 2
Manage Adoption and Integration Risks Draft the AI Solution Define Success Criteria and KPIs Support Business Case Creation Identify Project Resources: People, Hardware, and Contractors
6 Chapter 6: Domain III — Identify Data Needs, Part 1 (26%)
Define the Required Data Identify Data Subject-Matter Experts Identify Data Sources and Locations Coordinate the AI Workspace and Infrastructure Gather the Required Data
7 Chapter 7: Domain III — Identify Data Needs, Part 2
Check Data Privacy, Compliance, and Access Oversee Data Evaluation Determine If the Data Meets Solution Needs: The Phase 2 Go/No-Go Convey Data Understanding to Leadership
8 Chapter 8: Domain IV — Manage AI Model Development and Evaluation (16%)
Oversee AI/ML Model Technique and Algorithm Selection Oversee Model QA/QC and Configuration Management Manage AI/ML Model Training Manage Data Transformation and Data Preparation Verify Data Quality for the Data Preparation Go/No-Go Verify the Model Is Ready for Operationalization
9 Chapter 9: Domain V — Operationalize AI Solution, Part 1 (17%)
Manage Creation of the AI Solution Deployment Plan Manage the AI Solution Deployment Oversee Model Governance in Production Oversee AI Solution Metrics and Performance Monitoring
10 Chapter 10: Domain V — Operationalize AI Solution, Part 2
Prepare the Final Report and Lessons Learned Manage the AI Solution Transition Plan Oversee the AI Solution Contingency Plan
Oversee Model Governance in Production
Key Takeaways Explain This Topic AI
Every production model needs a registry entry and a named model owner, with a defined lifecycle running from registration through active service, revision and retirement.
A retrained model is a change requiring version control, approval and testing, not a routine refresh that can be pushed to production unreviewed.
Data drift means the input distribution shifted; concept drift means the relationship between inputs and outcome changed, which is why a model can degrade with no code change at all.
Retraining can be scheduled or trigger-based, but either way the training data must be checked for feedback loop contamination caused by the model's own decisions.
When accuracy declines and no code has changed, the exam expects you to investigate drift in the inputs and in the outcome relationship before rebuilding the model.
Last updated: August 2026
The Model Is an Asset with a Lifecycle
Ask an organization that has been running models in production for two years how many are currently serving live traffic, who owns each one, and what each was approved to do. The answer is very often a guess, and closing that gap is what ECO Domain V, Task 3 exists to do. The task asks you to govern each deployed model as a living asset with a named owner and a controlled lifecycle rather than as a finished deliverable — necessary because a model's accuracy moves while its code sits untouched, so nobody is prompted to look. Every production model runs through four stages:
Registration. The model is entered into the model inventory with its version, purpose, owner, risk tier, training data reference, approved use, documented limitations, and the approvals that let it into production.
Active service. The model runs under monitoring, with defined performance expectations and a review cadence proportionate to its risk tier.
Revision. Retraining, re-thresholding, feature changes and scope changes all pass through change control.
Retirement. The model is withdrawn, its endpoint decommissioned, and its records retained for the required period. Undocumented zombie models still serving traffic are a genuine audit finding.
Every entry has a named model owner who is accountable for its performance and for calling the decision to retrain, restrict or retire. This is operational stewardship; it is distinct from the accountability documentation and audit trail of Domain I, Task 5, which records who was answerable and what evidence exists, while this task controls how the asset is managed over time.
Versioning and Change Control in Production
The most common governance failure is treating a retrained model as a routine refresh. It is not. A retrained model is new software with new behavior , produced from different data, and it must move through the same controls the original did: a new version identifier, a registry entry, updated documentation, validation against a held-out set, a comparison against the model currently in production, and a recorded approval before promotion.
Build the change process to answer three questions for any production change: what changed (data, features, algorithm, thresholds, or serving configuration), who approved it, and how it can be reversed. Keep an emergency path for urgent fixes, with a documented retrospective approval, so that teams do not bypass governance when something is genuinely on fire.
Monitoring Drift
Accuracy can fall while the code is untouched. That is the property of AI that distinguishes this domain from ordinary IT operations, and it is examined through the drift distinction.
Drift type What changed Detection signal Typical response
Data drift (input or covariate drift) The distribution of the inputs has shifted, while the input-to-outcome relationship still holds Production input distributions diverge from the training baseline: new categories appear, means and ranges move, missing rates climb Confirm the shift is a real population change rather than a pipeline defect, then retrain on data that represents the new population
Concept drift The relationship between the inputs and the outcome has changed, even when the inputs look entirely normal Accuracy falls once ground truth arrives although input distributions appear stable; errors cluster in a time period or a segment Retrain on data from after the change, using a shorter, more recent window; the target definition or feature set may need rework
Prediction drift (leading indicator) The distribution of the model's own outputs has moved Approval rate, flag rate or confidence distribution departs from the deployment baseline Treat as an early warning while ground truth is still pending; check input health first, then investigate the relationship
Upstream break (not drift at all) A field, unit, encoding or default in the data pipeline changed A step change rather than a gradual trend: null rates spike, values fall outside valid ranges, volumes jump Fix the pipeline; retraining on corrupted inputs would bake the fault into the model
A worked contrast makes the difference stick. A credit model trained mainly on applicants aged 35 to 50 starts receiving a wave of applications from 18 to 24 year olds: the inputs moved, but income and debt still predict default the same way, so this is data drift. During a sharp downturn, borrowers with historically strong scores begin defaulting at twice their usual rate while the applicant mix is unchanged: what counts as a safe borrower has itself changed, so this is concept drift. The responses differ, and so does the urgency.
Coordinating Retraining
Retraining strategy is a management decision, and there are two defensible forms:
Scheduled retraining on a fixed cadence, such as monthly or quarterly. It is predictable, easy to resource and easy to audit, but it can arrive too late for a fast-moving concept shift and can waste effort when nothing has changed.
Trigger-based retraining driven by monitoring thresholds: a drift index breach sustained over a defined window, an accuracy floor crossed once ground truth arrives, or a known external event such as a product launch, a pricing change or a regulatory change. It responds faster but needs carefully set thresholds, or it fires constantly.
Most mature programs run both: a baseline cadence with trigger-based escalation. Whichever you choose, agree the thresholds and the approval path before you need them.
Two traps recur. The first is the feedback loop : retraining on data the model itself shaped. A fraud model that blocks a transaction means you never learn whether it would have been fraudulent; a lending model that rejects an applicant generates no repayment history; a recommender only sees clicks on items it chose to show. Retraining on that record narrows the model onto its own past decisions and hides its errors. Mitigations are managerial as much as technical: hold out a small randomized control slice, keep sampling human review on cases the model would have handled, and deliberately capture outcomes for overridden decisions.
The second trap is promoting a retrained model because it is newer. Validate every retrained candidate against the incumbent on the same evaluation set, including performance across the segments your bias checks cover, and promote only on evidence.
Ongoing Compliance and Periodic Review
Schedule a periodic model review with a cadence set by risk tier, at which the owner presents performance since the last review, drift history, incidents and overrides, changes to the data supply, whether the documented intended use still matches actual use, and a recommendation to continue, retrain, restrict or retire. Confirm at the same time that the model still complies with the privacy, transparency and regulatory obligations set in Domain I, since laws and internal policy move independently of the model.
The exam behavior to internalize: when accuracy declines and no code has changed, investigate drift in the inputs and in the outcome relationship before rebuilding the model. Rebuilding first is expensive and frequently solves nothing.