ITIL 4 CAI key terms

The capability of a system to perform tasks that normally require human judgement, such as recognising patterns, interpreting language, or recommending actions, by learning from data rather than following only pre-coded rules.

A branch of AI where algorithms build models from training data to make predictions or classifications, improving with exposure to examples rather than through explicit instructions for every case.

AI models that produce new content such as text, images, code, or summaries in response to prompts, based on patterns learned from large datasets. Outputs must be verified because they can be plausible but wrong.

A generative AI model trained on very large text corpora to predict and produce language, used for drafting, summarising, classifying, and conversational assistance in service management workflows.

The mathematical structure produced by training that maps inputs to outputs, such as a classifier that routes tickets or a model that forecasts incident volume. It is an artefact that must be versioned and maintained.

The historical examples used to build an AI model. Its quality, completeness, and representativeness directly determine how reliable and fair the resulting model's outputs will be.

The act of running a trained model on new input to produce an output, such as a predicted category or generated answer. In production, inference is what users and workflows actually consume.

Applying AI capabilities within service management activities such as incident triage, knowledge creation, request fulfilment, and demand forecasting, to augment human decisions rather than replace accountability.

A design distinction: augmentation gives people AI-generated suggestions they can accept or reject, while automation lets the system act without per-instance human approval. Risk tolerance determines which is appropriate.

The system of decision rights, policies, roles, and controls that directs how AI is selected, built, deployed, monitored, and retired, ensuring it stays aligned with organisational objectives and obligations.

A documented set of rules stating where AI may be used, which data it may touch, what approvals are required, and what behaviour is prohibited, providing a clear baseline for staff and suppliers.

A maintained record of identified AI-related risks, each with an owner, likelihood and impact assessment, mitigation actions, and review dates, used to track exposure over the AI lifecycle.

The possibility of loss or harm from incorrect, outdated, misapplied, or poorly understood model outputs, including decisions made on the basis of unreliable predictions or generated content.

Degradation of a model's accuracy over time because real-world conditions or data patterns change away from those seen in training, requiring monitoring, retraining, or retirement decisions.

Continuous observation of a deployed model's inputs, outputs, and performance indicators to detect drift, errors, bias, or misuse, and to trigger corrective action before harm reaches users.

Systematic unfair skew in data or model outputs that disadvantages particular groups or situations. It can arise from unrepresentative training data or from proxies that correlate with protected characteristics.

The degree to which people can understand why an AI system produced a given output, supporting challenge, audit, and appeal. It is often a precondition for trust in consequential decisions.

Openness about where and how AI is used, what data it relies on, and its known limitations, so that stakeholders are not misled and can make informed choices about relying on outputs.

The clear assignment of responsibility for AI outcomes to a named role or person, ensuring a human remains answerable even when the system acts automatically or a supplier provides the model.

A control design in which a person reviews, approves, or can override an AI output before it takes effect, appropriate where errors carry significant service, financial, or personal consequences.

A technical or procedural constraint that limits what an AI system may do, such as blocking certain outputs, restricting data access, or capping the value of transactions it can approve.

A graded scale describing how independently an AI system may act, from suggesting options for a human to acting and only informing afterwards. Higher levels demand stronger controls.

The fitness of data for its intended use, judged by accuracy, completeness, consistency, and timeliness. Poor data quality is a leading cause of unreliable AI outputs.

Protection of personal and sensitive information throughout AI use, covering lawful basis, minimisation, retention, and cross-border transfer, with particular care when data is sent to third-party models.

A confident but fabricated or incorrect output from a generative AI system, such as an invented policy clause or citation. It requires verification steps before outputs are trusted or published.

A sequence of activities that creates value for a stakeholder in which AI contributes at one or more steps, for example an incident resolution flow where AI drafts the diagnosis and knowledge article.

A management practice, such as incident management or knowledge management, in which AI tools support specific activities while the practice's purpose, roles, and controls remain defined by the organisation.

Embedding AI capabilities directly into the tools and steps people already use, so suggestions appear at the point of work rather than in a separate system that requires extra effort to consult.

The instruction or input given to a generative AI system. Clear prompts with context, constraints, and the desired format improve output usefulness and reduce the need for rework.

The step of checking AI-generated content or recommendations against authoritative sources or expert judgement before acting on them, especially where output will reach customers or be recorded as knowledge.

Using AI to draft, summarise, tag, and retrieve knowledge articles, while retaining human review for accuracy and ownership so that the knowledge base does not accumulate plausible but wrong content.

The application of AI and analytics to IT operations data such as events, logs, and metrics, to detect anomalies, correlate alerts, and support faster diagnosis of incidents.

The willingness of stakeholders to rely on AI-assisted services based on demonstrated competence, integrity, and care, built through consistent performance, honest disclosure of limitations, and remedy when things go wrong.

Principles guiding what is right and fair in AI use, including avoiding harm, respecting autonomy, treating people equitably, and considering effects on staff, customers, and wider society.

The condition in which AI-assisted outcomes do not systematically disadvantage individuals or groups, achieved through representative data, testing for disparate impact, and accessible routes to challenge decisions.

Informing customers, staff, and other stakeholders when AI is being used in a service interaction or decision, so they can judge how much to rely on the output and ask for a human alternative.

The documented boundary between permitted and prohibited AI usage by staff, such as banning confidential data in public tools, made concrete through examples and enforced through monitoring and training.

The recurring activity of measuring AI performance against objectives, learning from errors and feedback, and adjusting models, controls, or processes so that value and trust increase over time.

A structured mechanism for capturing user corrections, complaints, and outcome data on AI outputs and feeding them back into retraining, prompt refinement, or policy change, closing the improvement cycle.