CAIM flashcards

32 cards. Read the term and the definition here. The deck at the top of this page is optional.

AI Steering Committee

Cross-functional body that sets AI priorities, approves high-risk use cases, and resolves escalations. It owns the AI inventory and policy exceptions, reporting periodically to the board or its risk committee.

Model Owner Accountability

Named individual accountable for a model's intended use, performance, controls, and retirement. Ownership sits in the business first line, not in data science or internal audit.

AI Use Case Inventory

Register of every AI system in production or development, recording purpose, owner, data used, risk tier, and review date. It is the foundation for governance reporting and regulatory responses.

Build Versus Buy Rule

Buy commodity capabilities where differentiation is low; build when proprietary data or workflow creates advantage; partner when domain or regulatory expertise is missing internally.

Total Cost of Ownership Formula

Sum of data preparation, compute, integration, licensing, labeling, monitoring, retraining, and support costs over the asset life. It is the denominator in ROI calculations.

Benefit Attribution Baseline

Pre-initiative measurement of the target metric plus a defined comparison window. Without it, observed improvements cannot be credibly linked to the AI solution.

Payback Period

Upfront investment divided by annual net benefit. It shows how quickly cash is recovered and is most useful alongside NPV when benefits ramp over several years.

Adoption-Adjusted Value

Theoretical benefit multiplied by realistic adoption rate and productivity ramp. A thirty percent saving at half adoption yields roughly fifteen percent realized value, not thirty.

Training-Serving Skew

Divergence between feature computation during training and during live inference. It silently degrades predictions and is a primary reason to use a shared feature store.

Data Lineage Requirement

Documented trace from each training and inference dataset back to source systems, transformations, and version. It enables reproducibility, audit response, and impact analysis after upstream changes.

Data Classification Tiers

Labels such as public, internal, confidential, and restricted that determine encryption, access control, retention, and whether data may be used for model training.

Feature Store

Platform component that computes features once and serves them consistently to both training pipelines and real-time inference. It reduces duplication and guards against training-serving skew.

Shadow Deployment

New model runs alongside the incumbent without affecting decisions, and outputs are compared. It limits blast radius while validating behavior on live data before any user impact.

Model Card

Short document stating intended use, training data, evaluation metrics by segment, known limitations, and approval record. It supports review, audit, and safe reuse of a model.

Definition of Done for AI

Release criterion requiring monitoring dashboards, alert thresholds, retraining triggers, rollback path, and an accountable on-call owner, not merely a passing offline metric.

Champion-Challenger Testing

Production pattern where a candidate model is scored against the current champion on live traffic, and promotion occurs only when evidence shows superior business or model performance.

Fairness Metric Impossibility

Demographic parity, equalized odds, and predictive parity generally cannot all hold at once except in degenerate cases. Teams must choose the metric matching the decision's harm profile.

Meaningful Human Oversight

Reviewer has authority, competence, and time to override or stop the system. Nominal review without override power or context does not satisfy oversight obligations.

High-Risk AI Obligations

Typical requirements include a risk management system, data governance, technical documentation, logging, human oversight, accuracy standards, and post-market monitoring throughout the lifecycle.

AI Incident Response

Model-specific extension of incident response: detect harmful output or drift, contain by rollback or disablement, notify affected parties, remediate root cause, and log in the risk register.

AI Center of Excellence

Central team providing platforms, standards, and scarce expertise while business units own use cases. It suits organizations with many early use cases and limited technical depth.

Hub-and-Spoke AI Model

Central hub sets standards, platforms, and governance while embedded spokes own domain use cases. It balances reuse and control with business proximity as the portfolio matures.

AI Literacy Tiers

Three levels: general awareness for all staff, role-specific competency for tool users, and deep technical skill for builders. Each tier needs separate content and assessment.

Resistance Diagnosis

Classify pushback as skill gap, incentive conflict, or trust deficit. Each cause has a different remedy: training, metric redesign, or transparency and pilot evidence.

Covariate Drift

Shift in the distribution of input features relative to training data. It may degrade performance and is detectable without labels, making it a useful early warning signal.

Concept Drift

Change in the relationship between inputs and the target outcome, so the learned mapping becomes stale. It requires labeled data to confirm and usually triggers retraining.

Label Drift

Shift in the distribution of the target variable itself, such as a changed base rate. It alters calibration and threshold choices even when input features look stable.

Cost Per Inference

Total serving spend divided by request volume, tracked separately for real-time and batch paths. It exposes unit-economics erosion as models or traffic grow.

Retraining Trigger

Predefined condition that initiates model refresh: scheduled cadence, performance below threshold, detected drift, or sufficient new labeled data. Each trigger needs an owner and approval path.

Feedback Loop Bias

When model outputs influence future training data, past patterns get reinforced. Human overrides and corrections must be captured carefully to avoid entrenching existing bias.

Alert Fatigue Control

Pair every monitoring alert with a statistical threshold, a business threshold, and a runbook action. Alerts without actionable owners are removed or reclassified as dashboards.

Three Lines Model

Business owns and manages AI risk, independent risk and compliance set limits and challenge, and internal audit provides assurance. The structure separates risk-taking from risk oversight.