ITIL 4 CAI flashcards

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

AI in service management

Applying machine learning, natural language processing, generative AI and rule-based automation to design, deliver, support and improve services, augmenting human roles rather than replacing accountability for outcomes.

Classification model

A model that assigns inputs to one of a fixed set of labels, such as incident category or priority. Useful for triage and routing because outputs are bounded and checkable.

Generative AI

Models that produce new text, code or images from patterns in training data. Outputs are probabilistic and must be verified before entering records, knowledge bases or customer communication.

Agent copilot

An AI assistant embedded in the service desk tool that drafts responses, summarizes ticket history and suggests resolutions while the human agent retains decision authority and sends the output.

Utility and warranty for AI

Utility is fit for purpose, what the service does; warranty is fit for use, how reliably. For AI services warranty must also cover output accuracy, security and consistency.

Four dimensions of service management

Organizations and people, information and technology, partners and suppliers, and value streams and processes. AI adoption affects all four, so change in one dimension requires adjustment in the others.

Training data dependency

A model's quality is bounded by the representativeness and accuracy of its training data; historical bias, gaps and mislabeled records propagate directly into predictions and generated content.

AI use-case suitability rule

Prefer AI for high-volume, pattern-based tasks tolerant of probabilistic error. Keep deterministic, regulated or low-volume high-stakes decisions under explicit human control.

AI governance

The decision rights and controls that determine who approves an AI use case, who owns the model in production, who accepts residual risk and who can suspend the system.

Accountable model owner

A named human role responsible for a specific AI system's performance, compliance and incidents. Accountability cannot be delegated to the model or to a collective team.

Human in the loop

A control requiring human review or approval before an AI output takes effect. Mandatory where outputs affect safety, legal rights, employment, credit or essential service access.

Model risk tiering

Classifying AI use cases by impact and reversibility into low, medium and high tiers, with higher tiers requiring stronger validation, more frequent monitoring and senior approval.

Data drift

A change in the distribution of input data compared with training data, which degrades model accuracy over time. Detected by monitoring input statistics against a stored baseline.

Concept drift

A change in the relationship between inputs and correct outputs, so previously accurate predictions become wrong even when input distributions look stable. Requires retraining or rule updates.

Hallucination

A confident but factually incorrect or fabricated model output. Mitigated by retrieval grounding, citation requirements, confidence thresholds and mandatory human verification before use.

Prompt injection

A security attack where hidden or malicious instructions in input content cause a model to ignore its intended constraints, leak data or take unintended actions. Requires input filtering and least privilege.

AI use-case inventory

A register of every AI system in use, recording purpose, owner, risk tier, data sources, approval date and status. The baseline artifact for governance, audit and retirement decisions.

Residual risk acceptance

Formal sign-off by an accountable role that remaining risk after controls is tolerable for a defined use case, recorded with date and review trigger.

Third-party model risk

Risk from externally supplied or foundation models: unclear training data, silent model updates, deprecation, and limited audit access. Addressed through contract terms and change notification clauses.

Service value chain

Six activities: plan, improve, engage, design and transition, obtain/build, and deliver and support. AI can support each activity while practice owners keep accountability.

AI-assisted incident management

Using AI to classify incidents, match similar past tickets, summarize long threads and suggest resolutions, while humans confirm severity, diagnosis and closure.

AI-assisted problem management

Applying pattern detection and log correlation across incident history to surface candidate root causes; causal validation and known-error acceptance remain human decisions.

AI-assisted change enablement

Using historical change-failure data to score risk and draft change records, while the change authority remains accountable for authorization and scheduling decisions.

Chatbot escalation rule

A defined condition, such as low confidence, repeated failure or user request, that triggers handoff to a human agent with full conversation context transferred.

Knowledge article review gate

A mandatory human review step before AI-drafted knowledge articles are published, preventing incorrect content from propagating into future resolutions and model training.

AI suggestion acceptance rate

The share of AI recommendations accepted by agents or users without modification. A high rate can indicate good quality or uncritical over-reliance, so pair it with override and error rates.

AI override rate

The proportion of AI suggestions changed or rejected by humans. A rising rate signals degrading model quality, changed context or insufficient trust and training.

Explainability

The ability to state why a specific AI output occurred in terms a user or auditor can understand. Required strength increases with the impact of the decision.

Controllability

The property that humans can intervene in, override, pause or stop an AI system. Depends on documented intervention points, kill switches and clear authority to use them.

Disaggregated bias testing

Evaluating model performance separately across affected groups and subgroups rather than only in aggregate, because overall accuracy can conceal failure concentrated in one group.

Feedback loop risk

Degradation caused by training a model on its own outputs, amplifying errors and narrowing variety. Mitigated by keeping human-generated data in the training pipeline.

AI retirement criterion

A pre-agreed performance floor or trigger, such as sustained accuracy drop or unresolved incident, below which a model is suspended or withdrawn. Defined before deployment.