AI Transformation Leader flashcards

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

Value-driver mapping

Linking each AI initiative to a specific business lever - revenue growth, cost reduction, risk reduction, or experience - with a quantified baseline before investment approval.

Use-case prioritization matrix

Scoring candidate AI use cases on business value, feasibility, and data readiness, then sequencing by value-to-effort ratio and time-to-first-value.

Build-buy-partner rule

Buy commodity capability, build where proprietary data creates advantage, and partner when speed or regulated-domain expertise outweighs internal control benefits.

AI maturity stages

Progression from ad hoc pilots to systematic reusable platforms, then scaled portfolio governance, and finally AI embedded in the operating model and P&L.

North-star and guardrail metrics

One outcome metric tied to enterprise strategy, paired with two or three guardrail metrics that prevent local optimization from harming overall performance.

Three lines of defense

AI governance structure where developers own controls, risk and compliance set policy and challenge, and internal audit independently verifies effectiveness.

AI risk tiering

Classifying AI systems by impact on individuals and reversibility: rights-critical systems require human sign-off, low-impact reversible ones get light-touch review.

Model inventory fields

Minimum record per model: purpose, owner, data lineage, risk tier, validation evidence, monitoring metrics, last review date, and retirement plan.

Human-in-the-loop threshold

Rule that human review is mandatory for decisions affecting legal rights, safety, credit, employment, or health, while low-impact reversible actions may be automated.

Fairness metric impossibility

Demographic parity, equalized odds, and predictive parity generally cannot all hold at once, so the metric must be chosen to match the decision context.

AI incident response

Predefined severity levels, rollback or kill-switch paths, notification obligations, and root-cause reviews that feed findings back into the risk register.

Data quality gate dimensions

Completeness, accuracy, consistency, timeliness, uniqueness, and validity, each with thresholds set per use case before model training begins.

Data readiness tiers

Four stages from raw siloed data to cataloged and governed, then feature-ready and versioned, and finally real-time streaming; most enterprises sit at the first two.

Lineage and provenance

Traceability of every model input back to source system, transformation steps, and access approvals, which is a prerequisite for auditability and regulatory evidence.

Data contracts

Agreements between data producers and consumers specifying schema, service levels, and quality thresholds so upstream changes do not silently degrade production models.

MLOps reference stack

Combined tooling for data versioning, feature stores, experiment tracking, model registries, CI/CD for models, and production monitoring and observability.

Compute selection rule

Use GPUs for training and latency-sensitive inference, CPU or batch for periodic scoring, and managed services when demand is variable to avoid idle-capacity cost.

Baseline-first rule

Establish a simple rule-based or current-process benchmark before building complex models; a model must beat that baseline on the agreed metric to proceed.

Data leakage in evaluation

Contamination of test results when training data includes information unavailable at prediction time; prevented by time-based splits and strict separation of held-out sets.

Precision versus recall tradeoff

Choose precision when false positives are costly, recall when false negatives are costly, and cost-weighted or F1 scoring when both error types carry material cost.

Drift types

Data drift changes input distributions, concept drift changes the input-output relationship, and label drift changes target distributions; each implies a different retraining response.

Explicit retraining triggers

Predefined conditions such as performance dropping below threshold, drift statistics exceeding limits, a new data regime, or a scheduled cadence, rather than ad hoc retraining decisions.

RAG quality dependency

In retrieval-augmented generation, answer quality is bounded by retrieval quality, so grounding citations, prompt-injection defenses, and hallucination evaluation sets are required.

Adoption diffusion curve

Innovators and early adopters precede the early and late majority and laggards; targeting champions first and using peer proof accelerates spread.

AI skills taxonomy

Three categories: build skills such as ML engineering and data, use skills such as prompting and critical evaluation, and govern skills such as risk and compliance.

Change readiness assessment

Evaluation of leadership alignment, incentive compatibility, capability gaps, and process disruption, with remediation prioritized on the weakest dimension first.

Sources of AI resistance

Perceived job threat, loss of autonomy, and distrust of model outputs; mitigated through role redesign and clear communication of decision authority.

Value realization prerequisites

A baseline, target, measurement method, and attribution rule defined before launch; without them, claimed AI benefits cannot be verified or defended.

Attribution method strength

A/B or holdout testing is strongest, pre-post comparison is weaker, and expert estimate is weakest; the method and its confidence should always be stated.

AI total cost of ownership

Includes data preparation, integration, inference compute, monitoring, retraining, governance, and change management, not just the initial model build cost.

AI unit economics

Comparing compute cost per inference or transaction against value generated per transaction to confirm that scaling remains margin-positive.

EU AI Act risk tiers

Four tiers: prohibited unacceptable-risk practices, high-risk systems needing conformity assessment and human oversight, limited-risk with transparency duties, and minimal-risk with no added obligations.

GDPR automated decision rights

Decisions with legal or significant effects require a lawful basis, meaningful information about the logic used, and a right to obtain human intervention.