PMI CPMAI key terms

A structured first-pass review that ranks candidate AI initiatives by business value, data readiness, feasibility, and risk before any budget or build commitment is made.

A documented justification for an AI initiative covering expected benefit, total cost of ownership, delivery assumptions, key risks, and the baseline metric the project must improve.

A short authorizing document that names the AI project's objective, sponsor, scope boundaries, success measures, major constraints, and decision rights before work begins.

The pre-project measurement of current performance plus an external comparison point, used to prove that an AI solution actually improved outcomes rather than merely shipped.

A design pattern in which a person reviews, approves, or overrides model output at defined decision points, preserving accountability for consequential or ambiguous cases.

The tendency of people to accept machine recommendations uncritically, which can erode the safeguard value of human review and must be countered with training and challenge prompts.

The end-to-end sequence of phases an AI initiative passes through, from problem framing and data preparation to model development, deployment, monitoring, and eventual retirement.

A widely used six-phase process model for data mining and analytics projects: business understanding, data understanding, data preparation, modeling, evaluation, and deployment.

A delivery approach that ships AI capability in short increments, using feedback from each increment to refine data, features, and requirements rather than committing to a single fixed specification.

A time-boxed, narrowly scoped build that tests whether a technical approach can work at all, explicitly not intended as production-ready software or evidence of business value.

A limited production release to a restricted user group or region, used to validate operational behavior, adoption, and risk controls before wider rollout.

The set of engineering practices and tooling that automates building, testing, deploying, monitoring, and retraining machine learning models so they remain reliable in production.

An evaluation of whether available data is sufficient in volume, quality, coverage, labeling, and access rights to support the intended AI use case.

The measurable properties used to judge fitness of data, typically including accuracy, completeness, consistency, timeliness, validity, and uniqueness.

A recorded trail of where data originated, how it was transformed, and where it flows, enabling impact analysis, audit, and debugging of model inputs.

Documentation of the origin, ownership, and collection circumstances of a dataset, including consent status and any restrictions on its permitted uses.

A governance principle holding that data collected for one stated purpose should not be repurposed for unrelated uses, such as training a model, without fresh justification and consent.

Rules specifying how long data, model inputs, and model outputs are kept, how they are secured, and when and how they must be deleted.

Systematic distortion in the data used to fit a model, such as historical discrimination, sampling gaps, or labeling error, that can produce unfair or inaccurate predictions.

A quantified measure of how model outcomes are distributed across groups, for example demographic parity or equalized odds, chosen deliberately because the definitions conflict.

Deliberate interventions before, during, or after model training, such as resampling, reweighting, constraint-based training, or post-processing thresholds, to reduce disparate outcomes.

Classification of AI systems into risk levels based on potential harm, autonomy, and affected populations, which then determines required documentation, testing, and approval depth.

A structured record of a model's intended use, training data, performance results, known limitations, and owners, often called a model card or system card.

Degradation of model performance over time as real-world conditions or input distributions change, requiring detection through monitoring and remediation through retraining.

Ongoing tracking of production model accuracy, stability, and error patterns against thresholds, with defined escalation paths when results fall outside acceptable bounds.

The principle that named people and organizations remain answerable for AI-driven decisions, including the ability to explain, contest, and remediate adverse outcomes.

The degree to which a person can understand why a model produced a given output, whether through inherently interpretable models or post-hoc explanation techniques.

The structured work of preparing people for AI-enabled process change, addressing skills, incentives, role shifts, and trust so that adoption survives after go-live.

An assessment of whether an organization has the data infrastructure, skills, governance, and leadership commitment needed to absorb AI capability into daily operations.

A cross-functional body that sets AI policy, reviews high-risk use cases, approves exceptions, and owns the escalation path for ethical and compliance concerns.

An organization's stated commitments for AI conduct, commonly covering fairness, transparency, privacy, safety, human oversight, and sustainability, translated into reviewable practices.

A predefined plan for detecting, containing, communicating, and learning from harmful AI behavior in production, including rollback and notification responsibilities.

Artificially generated records that mimic the statistical properties of real data, used to fill coverage gaps or protect privacy, with its own validity risks.

A shared repository that computes, stores, and serves model input variables consistently for both training and live inference, reducing training-serving skew.

A failure mode where the data transformations used in model training differ from those applied in production, silently degrading prediction quality after deployment.

A living log of identified AI risks with owners, likelihood and impact ratings, mitigating controls, and review dates, maintained across the project lifecycle.