AI Glossary
Understand the language behind AI decisions. Clear definitions and practical context for leaders, managers, and teams putting AI to work.
Search a term, acronym, or concept. Or explore a topic below.
Browse the glossary alphabetically. Select any entry for practical context and related terms. Use your browser's Find command to search.
No matching terms
Try a different word or acronym, or reset the filters to browse the full glossary.
A/B TestingAn experiment that compares two versions of a process or product, ideally using random assignment, to estimate which performs better against a defined measure.
In practice
For example, compare an AI-assisted service workflow with the existing process using resolution quality and total handling time.
Related terms
AccountabilityThe obligation to explain decisions, accept responsibility for outcomes, and address problems. Using AI does not transfer organizational accountability to a model or supplier.
In practice
Assign a named owner for an AI service, including authority to suspend it and responsibility for responding to complaints.
Related terms
AccuracyThe proportion of predictions that are correct. A high accuracy score can hide poor performance on rare but important cases.
In practice
A system can achieve 99% accuracy by always predicting the majority class when only 1% of cases belong to the other class.
Related terms
Activation FunctionA mathematical operation applied inside a neural network that determines a unit's output. Nonlinear activation functions allow networks to represent complex relationships.
In practice
A network can learn patterns more complex than a straight-line relationship between inputs and an outcome.
Related terms
Active LearningA training approach that selects the examples expected to be most useful for labeling, often asking people to resolve uncertain or informative cases.
In practice
An inspection team labels difficult product images first instead of labeling every available image.
Related terms
Adversarial AttackA deliberate attempt to make an AI system behave incorrectly by manipulating its inputs, data, model, or operating environment.
In practice
A safety test checks whether small changes to an image or document can cause an otherwise reliable system to fail.
Related terms
Agentic AIAn approach in which AI systems pursue objectives through multiple steps, selecting actions and using feedback within defined permissions. The degree of autonomy varies by system.
In practice
An agent may research suppliers and draft a purchase request, with a person approving the order before it is placed.
Related terms
AI AdoptionThe process of making AI a useful, sustained part of how people work, supported by appropriate skills, processes, and oversight.
In practice
Measure whether teams complete work more effectively, not only how many people have activated a software license.
Related terms
AI AgentSoftware that uses AI to pursue a goal by selecting and carrying out actions, often through connected tools. Its capabilities depend on its design, permissions, and environment.
In practice
A service agent checks order status, proposes a response, and escalates exceptions under defined rules.
Related terms
AI AssuranceThe collection and assessment of evidence about whether an AI system meets stated requirements for its intended use.
In practice
A review brings together evaluation results, control testing, documentation, and monitoring evidence before deployment.
Related terms
AI GovernanceThe responsibilities, policies, decision processes, and controls used to direct how an organization develops, obtains, deploys, and monitors AI.
In practice
Governance determines who can approve a use case, what evidence is required, and who handles problems after launch.
Related terms
AI Impact AssessmentA structured examination of how a proposed or existing AI system may affect people, organizations, and the wider environment.
In practice
Assess a recruitment tool's intended benefits, affected applicants, potential harms, alternatives, and safeguards before use.
Related terms
AI InventoryA maintained record of the AI systems and use cases an organization develops or uses, including owners, purposes, suppliers, and risk information.
In practice
Include embedded AI in purchased software as well as systems built internally.
Related terms
AI LiteracyThe knowledge and practical ability to understand, use, evaluate, and question AI appropriately for a person's responsibilities.
In practice
A manager should recognize unreliable outputs, understand approved uses, and know when to seek specialist review.
Related terms
AI Operating ModelThe way an organization arranges people, decision rights, funding, processes, and technical capabilities to deliver and govern AI.
In practice
A central team may set shared controls while business units own use cases and benefits.
Related terms
AI PolicyAn organization's documented rules for acceptable AI use, responsibilities, data handling, approvals, and escalation.
In practice
A policy can specify which tools are approved, which information may be entered, and when outputs need review.
Related terms
AI Risk ManagementThe ongoing process of identifying, assessing, addressing, and monitoring potential harm from AI throughout its life cycle.
In practice
Consider operational failures, security, privacy, unfair outcomes, and dependencies as the system and its use change.
Related terms
AI StrategyAn organization's choices about where AI should create value, which capabilities it will build or obtain, and how it will manage execution and risk.
In practice
Connect proposed AI investments to measurable business priorities and clear ownership.
Related terms
AlgorithmA defined sequence of steps for solving a problem or performing a task. Algorithms may be fixed rules or procedures used to train and run AI models.
In practice
A rule that sorts invoices by due date is an algorithm even if it does not use machine learning.
Related terms
AlignmentThe effort to make an AI system's behavior consistent with intended objectives, instructions, and human values. Alignment is a continuing challenge, not a guarantee of safe behavior.
In practice
A model may follow a user's instruction while still producing an answer that conflicts with organizational policy.
Related terms
AnnotationThe addition of labels or descriptive information to data so it can be used for training, evaluation, or analysis.
In practice
Reviewers mark whether service messages contain a complaint and identify the relevant sentence.
Related terms
Anomaly DetectionThe identification of observations or patterns that differ from what a system considers normal or expected.
In practice
Unusual equipment readings can trigger an investigation. An anomaly is a signal to examine, not proof of a fault.
Related terms
Application Programming Interface (API)A defined way for software systems to exchange information or request actions from one another.
In practice
An internal application can send a document to an AI service through an API and receive a structured result.
Related terms
Area Under the Curve (AUC)A summary measure of a curve. In classification, ROC-AUC measures how well a model ranks positive cases above negative cases across decision thresholds.
In practice
A strong ROC-AUC does not by itself establish the best operating threshold or the business cost of errors.
Related terms
Artificial General Intelligence (AGI)A term for AI with broad capabilities across many kinds of intellectual tasks. There is no universally accepted definition or test for deciding when AGI has been achieved.
In practice
Assess a product's demonstrated capabilities for your use case rather than relying on an AGI label.
Related terms
Artificial Intelligence (AI)Computer-based systems that infer how to produce outputs such as predictions, recommendations, content, or decisions from the inputs they receive.
In practice
AI can forecast demand, analyze images, or draft text. Different tasks require different methods and controls.
Related terms
Association Rule LearningA method for identifying items or events that frequently occur together in a dataset.
In practice
A retailer discovers that certain products are often bought together. The association does not show that one purchase causes the other.
Related terms
AttentionA mechanism that lets a model weight the relevance of different input elements when calculating a representation or output.
In practice
In text processing, attention helps a model use surrounding words to interpret a phrase.
Related terms
Audit TrailA record of relevant events, changes, approvals, and actions that helps reconstruct what happened in a system.
In practice
Record the model version, information retrieved, actions taken, and human approvals needed to investigate an incident.
Related terms
AugmentationThe use of AI to extend human capability while people retain an active role in the work and relevant decisions.
In practice
A manager uses AI to compare options, then checks assumptions and makes the decision.
Related terms
AutoencoderA neural network trained to reconstruct its input, commonly through a compressed internal representation.
In practice
A model may learn a compact representation of equipment signals for later analysis or anomaly detection.
Related terms
Automatic Speech Recognition (ASR)Technology that converts spoken audio into written text.
In practice
Meeting transcription should be checked for names, specialized vocabulary, and unclear audio before important actions are taken.
Related terms
AutomationThe use of technology to carry out tasks with reduced manual effort. Automation may use fixed rules, AI, or a combination of both.
In practice
Automatically moving approved invoices between systems can be useful without requiring an AI model.
Related terms
Automation BiasThe tendency to give excessive weight to a system's output, including when other evidence suggests it is wrong.
In practice
A reviewer accepts an AI-generated explanation without checking the underlying records because it sounds authoritative.
Related terms
BackpropagationA method for calculating how a neural network's parameters contributed to its error, enabling an optimizer to update those parameters.
In practice
It is a core part of training many neural networks, including systems for language and image analysis.
Related terms
BaselineA reference level of performance used to judge whether a proposed change delivers improvement.
In practice
Compare a new AI workflow with the current process and a simple alternative using the same workload and quality criteria.
Related terms
BatchA group of examples processed together. During training, a batch is commonly used to calculate one parameter update.
In practice
Batch size affects training resource use and behavior; batch processing can also group production requests for efficiency.
Related terms
Bayes' TheoremA mathematical rule for updating the probability of a hypothesis when new evidence becomes available.
In practice
A forecast can be updated as new sales data arrives, while accounting for what was believed before the new evidence.
Related terms
BenchmarkA defined task, dataset, or measurement procedure used to compare systems consistently.
In practice
A public benchmark can help shortlist models, but the final choice needs evaluation on your own tasks and operating conditions.
Related terms
BiasA systematic tendency that influences a system's outputs. In AI, bias can refer to statistical error, learning assumptions, or outcomes that disadvantage people.
In practice
Specify which meaning is being discussed. A biased dataset and a neural network's bias parameter are different concepts.
Related terms
Bias-Variance TradeoffThe tension between a model being too restrictive to capture real patterns and too sensitive to the particular data used to train it.
In practice
A more complex model may fit historical data better while performing worse on new cases.
Related terms
BoostingAn ensemble approach that builds models sequentially, with later models addressing weaknesses in the combined result.
In practice
Boosted decision trees are a common option for predicting outcomes from structured business records.
Related terms
Bounding BoxA rectangle used to identify the location and extent of an object in an image or video frame.
In practice
An annotation team draws boxes around damaged areas of products to prepare data for an inspection model.
Related terms
CalibrationThe degree to which predicted probabilities match observed outcomes over comparable cases.
In practice
Among cases assigned a 70% probability, approximately 70% should have the event occur in a well-calibrated system.
Related terms
Causal InferenceMethods for estimating how changing one factor affects another, using experiments or assumptions that support causal conclusions.
In practice
An observed association between discounts and sales does not by itself establish how much a new discount will increase sales.
Related terms
Central Processing Unit (CPU)A general-purpose processor that executes a computer's instructions and coordinates many computing tasks.
In practice
AI applications use CPUs alongside other hardware; the best configuration depends on the workload and cost constraints.
Related terms
Change ManagementThe planned work of helping people and teams adopt changes to responsibilities, processes, skills, and ways of working.
In practice
An AI rollout includes training, revised procedures, feedback channels, and support for employees whose tasks change.
Related terms
ChatbotSoftware that interacts with users through conversation. A chatbot may use scripted rules, AI models, or both.
In practice
A conversational interface alone does not establish whether the system can take actions or access current business information.
Related terms
ChunkingDividing content into smaller pieces for processing, retrieval, or inclusion in a model's context.
In practice
A knowledge assistant retrieves relevant sections of a manual instead of sending the entire document for every question.
Related terms
ClassificationA prediction task that assigns an input to one or more predefined categories.
In practice
A model routes incoming service requests into billing, technical support, or account administration categories.
Related terms
ClusteringGrouping data points by similarities without requiring a predefined label for each group.
In practice
A business explores patterns in customer behavior, then checks whether the resulting groups are useful and meaningful.
Related terms
Cold StartThe difficulty of making useful recommendations or predictions when little relevant history exists for a new user, item, or system.
In practice
A new customer may initially receive recommendations based on product information or stated preferences instead of purchase history.
Related terms
Collaborative FilteringA recommendation approach that uses patterns of interaction across users and items to suggest what someone may prefer.
In practice
A service recommends content based on similar viewing patterns, subject to privacy and suitability controls.
Related terms
Computer VisionThe field of AI concerned with extracting useful information from images and video.
In practice
A manufacturer uses images to identify surface defects, then measures performance under real lighting and production conditions.
Related terms
Concept DriftA change in the relationship between model inputs and the outcome being predicted.
In practice
The factors associated with customer cancellations change after a competitor introduces a new service.
Related terms
Confidence IntervalAn interval calculated using a method designed to contain the true population value at a stated rate across repeated sampling.
In practice
A 95% confidence method would capture the true value in about 95% of repeated studies. This differs from a model's confidence score.
Related terms
Confidence ScoreA number a model or system reports to indicate support for an output. Its interpretation depends on how it was produced and calibrated.
In practice
An AI statement that it is '95% confident' is not automatically a validated 95% chance of being correct.
Related terms
Confirmation BiasThe tendency to seek or favor evidence that supports an existing belief while discounting conflicting evidence.
In practice
A project team highlights successful demonstrations but overlooks failures that challenge its preferred investment decision.
Related terms
Context WindowThe amount of information a model can process together in one interaction, commonly measured in tokens. Input and output limits depend on the model and service.
In practice
A larger window can hold more material, but does not guarantee that the model will use every detail accurately.
Related terms
Convolutional Neural Network (CNN)A neural network architecture that uses filters to detect patterns across local regions of an input, particularly images.
In practice
An image model can combine local features such as edges into patterns useful for identifying a product defect.
Related terms
CopilotA common product term for an AI assistant that supports a person inside a task or workflow. Its actual autonomy and permissions vary by implementation.
In practice
An assistant suggests a report draft while the employee reviews the facts and authorizes publication.
Related terms
Cross-ValidationA family of methods that repeatedly split data into training and validation portions to assess how a model performs beyond its training examples.
In practice
Use splits that respect time or shared customers when random splitting would let related information cross between groups.
Related terms
Data AugmentationCreating additional training examples by transforming existing data while preserving relevant meaning or labels.
In practice
An inspection dataset may include rotated or differently lit product images to help a model handle realistic variation.
Related terms
Data DriftA change in the characteristics or distribution of the data a system receives compared with a reference period.
In practice
A model built for one customer mix receives increasing traffic from another region and needs its performance reassessed.
Related terms
Data GovernanceThe responsibilities, rules, and processes used to manage data quality, access, use, security, and life-cycle decisions.
In practice
An AI project needs clarity about who owns its data and who can approve changes or new uses.
Related terms
Data LeakageIn model development, the use of information that would not legitimately be available when making a real prediction. In security, the term can also mean unauthorized disclosure of data.
In practice
Including future payment status in a model that predicts late payments creates misleading evaluation results.
Related terms
Data LineageA record of how data moves and changes as it passes through systems and processing steps.
In practice
If a dashboard or model uses an incorrect field, lineage helps identify the source and affected downstream outputs.
Related terms
Data MinimizationLimiting the data collected, retained, and used to what is needed for a defined purpose.
In practice
An AI assistant answering policy questions may need policy documents without needing employees' individual personnel files.
Related terms
Data PoisoningThe intentional manipulation of data used to train, adapt, or support an AI system in order to influence its behavior.
In practice
An attacker introduces misleading examples into a training pipeline or false material into a retrieval knowledge base.
Related terms
Data PrivacyThe appropriate handling of information about people, including how it is collected, used, shared, protected, and retained.
In practice
Check what an AI supplier does with submitted information and whether the intended use is compatible with applicable obligations.
Related terms
Data ProvenanceInformation about where data came from, who created or supplied it, and the circumstances of its collection.
In practice
Before using a dataset, establish its origin, collection context, and any restrictions that affect the proposed use.
Related terms
Data QualityHow well data meets the needs of its intended use, including accuracy, completeness, consistency, timeliness, and relevance.
In practice
A complete customer table can still be unsuitable for forecasting if the definitions changed halfway through the period.
Related terms
Data ReadinessThe extent to which the data needed for a use case is accessible, suitable, governed, and usable in the intended workflow.
In practice
A promising use case may be delayed by missing labels, restricted access, or inconsistent identifiers across systems.
Related terms
Data SovereigntyThe relationship between data and the laws or governing authority that apply to it. It is related to, but broader than, where data is physically stored.
In practice
Review storage location, processing, remote access, and supplier arrangements instead of treating local hosting as a complete answer.
Related terms
DatasetA collection of data assembled for a purpose such as analysis, training, or evaluation. It may contain records, text, images, audio, or other formats.
In practice
Document what the collection represents, what it excludes, and how it will be used.
Related terms
Decision RightsThe assignment of authority to make, approve, challenge, or stop particular decisions.
In practice
Specify who can approve a pilot, release an AI system, accept residual risk, and authorize material changes.
Related terms
Decision TreeA model that makes predictions by following a series of branching conditions on input values.
In practice
A tree may split service requests by product type and issue characteristics to predict the appropriate support route.
Related terms
Deep LearningMachine learning that uses neural networks with multiple layers to learn representations and patterns from data.
In practice
It supports many language, image, and speech systems, but is not necessary for every business prediction task.
Related terms
DeepfakeAI-generated or manipulated media that convincingly portrays someone saying or doing something they did not actually say or do.
In practice
An unexpected payment instruction in a familiar-sounding voice should be verified through an established independent channel.
Related terms
Differential PrivacyA mathematical approach that limits how much an analysis can reveal about whether a particular individual's data was included.
In practice
An organization can add carefully controlled randomness when releasing statistics, balancing privacy protection with usefulness.
Related terms
Diffusion ModelA generative model that learns to reverse a gradual noising process to produce new samples.
In practice
Many image-generation systems use diffusion methods to create images from prompts or other inputs.
Related terms
Dimensionality ReductionReducing the number of variables used to represent data while retaining information useful for a task.
In practice
A team can simplify a dataset for visualization or modeling, while checking what information the reduction discards.
Related terms
DistillationTraining one model to reproduce useful behavior or information from another model, often to obtain a smaller or more efficient system.
In practice
A compact model may be trained for a narrow service task using outputs from a larger model, then evaluated independently.
Related terms
Edge AIRunning AI processing on or near the device that collects the data, rather than relying entirely on a remote service.
In practice
A factory camera analyzes images locally when low latency or unreliable connectivity makes cloud-only processing unsuitable.
Related terms
EmbeddingA numerical representation of an item, such as text or an image, designed so useful relationships can be captured in a vector space.
In practice
A search system compares embeddings to retrieve documents with related meanings, even when their wording differs.
Related terms
Ensemble MethodsApproaches that combine multiple models to produce a prediction or decision.
In practice
Combining several models may improve performance or stability, but adds complexity and still requires evaluation.
Related terms
EntropyA mathematical measure of uncertainty in a probability distribution.
In practice
A classifier that spreads probability across many possible categories has higher entropy than one that concentrates it on a single category.
Related terms
EpochOne complete pass through the training dataset during model training.
In practice
More epochs mean more opportunities to learn, but excessive training can make a model fit the training data too closely.
Related terms
EU AI ActAn EU legal framework that regulates AI using differentiated requirements and prohibitions. Responsibilities depend on matters including the system's use and the organization's role.
In practice
A business needs to establish which rules apply to its particular activities, including when obtaining AI from another supplier.
Related terms
EvaluationThe systematic assessment of an AI system against defined criteria, using evidence appropriate to its intended use.
In practice
Assess the complete workflow for quality, safety, cost, and usability, including difficult cases and relevant user groups.
Related terms
ExplainabilityThe ability to provide useful reasons or accounts for a system's outputs or behavior. An explanation must be assessed for accuracy and suitability for its audience.
In practice
A plausible narrative generated after a decision is not necessarily a faithful explanation of how the system produced it.
Related terms
F1 ScoreThe harmonic mean of precision and recall: 2 × precision × recall ÷ (precision + recall), where defined.
In practice
It summarizes a balance between two metrics, but does not directly represent business costs or account for true negatives.
Related terms
FairnessThe assessment and treatment of unjustified differences in how an AI system affects people or groups. Different fairness measures can conflict.
In practice
Select and justify relevant criteria for the use case, then examine performance and impacts across affected groups.
Related terms
False NegativeA case incorrectly classified as negative when it is actually positive.
In practice
A defect-detection system marks a damaged item as acceptable and the item proceeds to shipping.
Related terms
False PositiveA case incorrectly classified as positive when it is actually negative.
In practice
A legitimate customer transaction is incorrectly flagged as suspicious and requires unnecessary review.
Related terms
FeatureAn input characteristic used by a model to make a prediction.
In practice
For a demand model, features might include recent sales, season, price, and delivery lead time.
Related terms
Feature EngineeringDesigning or transforming model inputs to make useful patterns easier to learn.
In practice
A team converts individual purchase records into measures such as days since the last purchase.
Related terms
Feature LearningLearning useful representations from data instead of specifying every feature manually.
In practice
A neural network learns image patterns from examples rather than relying only on features designed by an engineer.
Related terms
Feature SelectionChoosing which available input variables to include in a model.
In practice
Remove inputs that add little value or would not be available at prediction time, then reassess performance.
Related terms
Federated LearningTraining a model across multiple devices or organizations while keeping raw training data at those locations and exchanging model-related updates.
In practice
It can reduce centralized data collection, but privacy and security still require additional safeguards.
Related terms
Few-Shot PromptingGiving a model a small number of examples in the prompt to demonstrate the desired task, format, or style.
In practice
Provide a few correctly categorized support messages before asking the model to categorize a new one.
Related terms
Fine-TuningAdditional training that changes a pretrained model's parameters using data selected for a particular task or behavior.
In practice
A team tunes a model to produce a consistent response format. This does not automatically give it current knowledge or eliminate errors.
Related terms
Foundation ModelA model trained on broad data that can be adapted or used for a range of downstream tasks.
In practice
Evaluate the chosen model and its application together; broad capability does not establish suitability for a specific business use.
Related terms
Garbage In, Garbage OutThe principle that poor-quality inputs can produce unreliable outputs even when the underlying processing works as designed.
In practice
A forecasting model cannot reliably correct for sales records that systematically omit a major channel.
Related terms
General Data Protection Regulation (GDPR)An EU regulation governing the processing of personal data. Its application to an AI activity depends on the data, processing, and circumstances involved.
In practice
Using AI does not remove obligations concerning personal data, including those relevant to the organization's specific processing activity.
Related terms
GeneralizationA model's ability to perform usefully on new examples beyond the data used to train it.
In practice
Good training results matter only if the system also performs well on the kinds of cases it will encounter in operation.
Related terms
Generative Adversarial Network (GAN)A model setup in which a generator learns to create samples while a discriminator learns to distinguish generated samples from reference data.
In practice
A GAN can produce synthetic images, but generated material still needs quality and suitability checks.
Related terms
Generative AIAI that produces content such as text, images, audio, video, or code using patterns learned from data.
In practice
A model can draft a customer response, but the organization must still check its accuracy and suitability before use.
Related terms
Gradient DescentAn optimization method that repeatedly adjusts parameters in a direction intended to reduce a model's loss.
In practice
The learning rate controls the size of the updates and influences whether training progresses effectively.
Related terms
Graphics Processing Unit (GPU)A processor designed to execute many operations in parallel, making it useful for many AI training and inference workloads.
In practice
GPU requirements can materially affect infrastructure cost, availability, and deployment options.
Related terms
Ground TruthReference information treated as correct for a particular training or evaluation task. It may come from measurement or human judgment and can itself contain errors.
In practice
Check disagreement and ambiguity in reference labels before interpreting the model's score as an objective measure of quality.
Related terms
GroundingConnecting a model's response to supplied evidence, retrieved information, or other relevant external context.
In practice
An assistant bases an answer on an approved policy document and provides a reference that the user can check.
Related terms
GuardrailsControls intended to constrain an AI system's inputs, outputs, or actions within defined boundaries.
In practice
Combine access controls, content checks, approval requirements, and monitoring. A written prompt alone is not a complete control system.
Related terms
HallucinationAn AI-generated statement or output that is unsupported, incorrect, or inconsistent with the relevant evidence, even when presented convincingly.
In practice
A model invents a policy clause or cites a source that does not support its answer.
Related terms
Human OversightThe arrangements that give people the information, competence, authority, and opportunity to supervise or challenge an AI system.
In practice
A reviewer needs enough time and evidence to assess an output, plus the ability to reject or stop it.
Related terms
Human-in-the-Loop (HITL)A design in which a person directly contributes to specified stages of an AI process, such as labeling data, checking outputs, or approving an action.
In practice
A person approves a consequential transaction before the system executes it. The review must allow meaningful intervention.
Related terms
Human-on-the-Loop (HOTL)A design in which people supervise an operating system and can intervene, rather than approving every individual action.
In practice
A supervisor monitors exceptions and can pause automation when performance or risk exceeds a defined limit.
Related terms
HyperparameterA setting that controls a model's structure or learning process, rather than a value learned directly as a model parameter.
In practice
Examples include the learning rate, tree depth, and batch size. Tuning decisions should use validation evidence.
Related terms
Incident ResponseThe coordinated process for detecting, containing, investigating, and recovering from an incident involving an AI system.
In practice
A response plan identifies who can stop the system, preserve evidence, assess affected parties, and approve a safe return to service.
Related terms
InferenceUsing a trained model to generate an output from an input. Inference is distinct from the training process that establishes or changes the model's parameters.
In practice
Each request to summarize a document runs inference; it does not necessarily train or update the model.
Related terms
Information RetrievalFinding and ranking information relevant to a query from a collection of documents or other records.
In practice
An enterprise search system retrieves candidate documents before a person or model uses them to answer a question.
Related terms
InterpretabilityThe extent to which a person can understand a model's operation or the relationship between its inputs and outputs.
In practice
A small decision tree may be easier to inspect directly than a large neural network, depending on the task.
Related terms
ISO/IEC 42001An international standard specifying requirements for an artificial intelligence management system within an organization.
In practice
It addresses organizational management processes. Certification of a management system is not proof that every AI output is safe or correct.
Related terms
JailbreakAn attempt to get a model to bypass its intended behavioral restrictions, often by manipulating the request or conversation.
In practice
Authorized testing checks whether the model can be induced to violate a relevant policy or reveal restricted material.
Related terms
Knowledge GraphA representation of entities and their relationships that makes connected information usable by software.
In practice
An organization links suppliers, contracts, products, and locations to support search and analysis.
Related terms
Large Language Model (LLM)A model trained on large amounts of language data to recognize and generate patterns in text. Some models also process other modalities.
In practice
An LLM can draft or summarize text, but fluent language is not evidence that its claims are correct.
Related terms
LatencyThe time between a request or event and the relevant system response. Different measures may track the first output or the completed result.
In practice
A workflow can feel slow even when the model generates quickly if retrieval, tool calls, or approvals add delays.
Related terms
LayerA group of operations or units within a neural network that transforms information before passing it to another part of the network.
In practice
Hidden layers build intermediate representations between the inputs and the final output.
Related terms
Learning RateA training setting that controls the size of parameter updates during optimization.
In practice
An unsuitable learning rate can slow learning or make training unstable.
Related terms
Learning to RankUsing machine learning to order results according to their predicted relevance or usefulness.
In practice
A search system ranks help articles so the most relevant answers appear first.
Related terms
Least PrivilegeGiving a person or system only the access and authority needed for its approved purpose.
In practice
An assistant that summarizes invoices may need read access, without permission to change supplier details or release payments.
Related terms
Logistic RegressionA statistical model commonly used for classification that maps a weighted combination of inputs to a probability, often using a logistic function.
In practice
Despite its name, it is frequently used for yes-or-no prediction tasks rather than predicting a continuous quantity.
Related terms
Long Short-Term Memory (LSTM)A type of recurrent neural network with mechanisms for retaining and updating information across a sequence.
In practice
It has been used for tasks such as analyzing time series or processing sequential language data.
Related terms
Loss FunctionA mathematical measure of how far a model's outputs depart from its training objective.
In practice
The training loss is not automatically the same as the business cost of an incorrect output.
Related terms
Machine Learning (ML)Methods that learn patterns from data to improve performance on a task, rather than specifying every decision through hand-written rules.
In practice
A model learns from past examples to estimate demand, categorize documents, or detect unusual behavior.
Related terms
Machine Learning Operations (MLOps)The practices used to develop, deploy, monitor, maintain, and govern machine learning systems reliably over time. LLMOps applies related practices to language-model applications.
In practice
Maintain versioned models, data pipelines, evaluations, release procedures, and monitoring rather than treating launch as the end of the project.
Related terms
Machine TranslationThe automatic conversion of content from one language to another.
In practice
For important materials, review terminology, meaning, and cultural context before publishing a machine-generated translation.
Related terms
ModelA mathematical or computational representation used to produce outputs from inputs. In machine learning, its behavior is shaped through training.
In practice
A model is one component of a larger system that may also include data sources, tools, interfaces, and human review.
Related terms
Model CardDocumentation describing a model's intended use, evaluation, characteristics, and known limitations.
In practice
Use model documentation as one input to due diligence, then test the application under your own operating conditions.
Related terms
Model Context Protocol (MCP)An open protocol for connecting AI applications with external tools and data sources through a consistent interface.
In practice
A connector can make a document repository available to an assistant. Authentication and permissions still need to be configured appropriately.
Related terms
MonitoringThe ongoing measurement of system behavior, performance, costs, and risks during operation.
In practice
Track failure patterns, user outcomes, response times, and unexpected changes, with defined thresholds for action.
Related terms
Monte Carlo SimulationA method that repeatedly samples from specified probability distributions to explore possible outcomes.
In practice
A team explores a range of project costs and benefits instead of relying on a single forecast.
Related terms
Multimodal AIAI that processes or generates more than one type of information, such as text, images, audio, or video.
In practice
An assistant analyzes an equipment photograph alongside a written maintenance note.
Related terms
Multitask LearningTraining a model on multiple related tasks so shared representations can support more than one objective.
In practice
One model may learn to identify the topic and sentiment of service messages using shared language features.
Related terms
Naive BayesA family of classifiers based on Bayes' theorem that assumes features are conditionally independent given the class.
In practice
It can provide a simple starting point for classifying text, even when the independence assumption is imperfect.
Related terms
Named Entity Recognition (NER)Identifying and categorizing references to entities in text, such as people, organizations, locations, or dates.
In practice
A document workflow extracts supplier names and contract dates for review and indexing.
Related terms
Narrow AIAI designed or used for a limited range of tasks, rather than general competence across all intellectual activities.
In practice
A highly capable inspection model does not automatically have the ability to plan a business strategy.
Related terms
Natural Language Processing (NLP)The field concerned with computational methods for processing, analyzing, and generating human language.
In practice
Applications include document classification, translation, information extraction, and conversation.
Related terms
Neural NetworkA model made of connected computational units whose adjustable parameters transform inputs into outputs.
In practice
Neural networks range from small prediction models to large systems for language, images, and other complex data.
Related terms
NeuronA computational unit in a neural network that transforms incoming values into an output, commonly using weights, a bias, and an activation function.
In practice
The term is inspired by biology, but an artificial neuron is a mathematical operation rather than a biological brain cell.
Related terms
NIST AI Risk Management FrameworkA voluntary framework for managing AI risks, organized around the functions Govern, Map, Measure, and Manage.
In practice
Organizations can use it to structure responsibilities, understand context, assess risk, and choose responses throughout the AI life cycle.
Related terms
ObservabilityThe ability to understand a system's internal behavior from evidence such as logs, metrics, and traces.
In practice
For an AI workflow, trace which documents were retrieved, which tools were called, and where an error or delay occurred.
Related terms
Open-Weight ModelA model whose trained weights are made available under specified terms. This alone does not mean its training data, code, or permitted uses are unrestricted.
In practice
Review the license and available materials before assuming that a downloadable model is fully open source or suitable for commercial use.
Related terms
Optical Character Recognition (OCR)Technology that converts text in images or scanned documents into machine-readable text.
In practice
An invoice workflow extracts printed text, then checks whether key fields such as amounts and account numbers were read correctly.
Related terms
OrchestrationCoordinating models, tools, data retrieval, and other steps so they work together as a complete process.
In practice
A document assistant retrieves sources, generates a draft, runs checks, and routes exceptions for review.
Related terms
OverfittingLearning details of the training data that do not generalize well to new examples.
In practice
A model performs impressively on familiar records but makes many errors when used on genuinely new cases.
Related terms
ParameterA numerical value within a model that helps determine its outputs. Many parameters are learned during training.
In practice
Model size is often described by parameter count, but a larger count alone does not establish better performance for your task.
Related terms
Pattern RecognitionIdentifying regularities or structures in data that can support tasks such as classification, detection, or prediction.
In practice
A system identifies recurring combinations of sensor readings associated with equipment wear.
Related terms
Personal InformationInformation that identifies a person or can reasonably be linked to them, depending on the applicable legal definition and context.
In practice
Names, identifiers, and combinations of otherwise ordinary details can create privacy obligations when linked to individuals.
Related terms
PilotA limited deployment used to test whether an approach works in realistic conditions before a broader rollout. A proof of concept may address a narrower feasibility question.
In practice
Define success, cost, risk, and exit criteria before starting so the pilot can support a clear scale, revise, or stop decision.
Related terms
PoolingAn operation that summarizes nearby values in a representation, often reducing its size in a convolutional network.
In practice
Max pooling keeps the largest value in a region; average pooling uses the region's average.
Related terms
PrecisionAmong the cases a model labels positive, the proportion that are actually positive: true positives ÷ all predicted positives.
In practice
If 80 of 100 flagged invoices contain the target error, precision is 80%.
Related terms
PredictionAn output that estimates a category, value, event, or other unknown result from available inputs. It need not concern the future.
In practice
Predicting a document's category and forecasting next month's demand are both prediction tasks.
Related terms
PreprocessingPreparing data for analysis or modeling through steps such as cleaning, transformation, normalization, or handling missing values.
In practice
Apply transformations consistently and avoid using test data to make training-stage decisions.
Related terms
Pretrained ModelA model that has already undergone training and can be used directly or adapted for another purpose.
In practice
Starting with a pretrained model can reduce development work, but suitability and licensing still need assessment.
Related terms
Principal Component Analysis (PCA)A linear method that transforms data into uncorrelated components ordered by the variation they capture.
In practice
It can compress numerical data, but the components with the most variation are not necessarily those most useful for a business prediction.
Related terms
PromptThe instructions and other input provided to a generative model to guide its response.
In practice
A useful request states the task, relevant context, expected format, and any constraints that matter to the result.
Related terms
Prompt EngineeringDesigning and testing the instructions, examples, and context given to a model to improve performance on a task.
In practice
Test a revised prompt on representative cases instead of judging it from a single successful response.
Related terms
Prompt InjectionAn attack that supplies instructions through inputs or external content in an attempt to redirect an AI application's behavior.
In practice
A retrieved document contains text that tries to override the assistant's task or cause an unauthorized action.
Related terms
QuantizationRepresenting model values with lower numerical precision to reduce storage or computation requirements.
In practice
A smaller memory footprint can make local deployment more practical, but the effect on quality must be tested.
Related terms
Random ForestAn ensemble of decision trees trained with randomness in data sampling and feature selection, whose outputs are combined.
In practice
It is a common method for classification and numerical prediction using structured records.
Related terms
RecallAmong all actual positive cases, the proportion correctly identified as positive: true positives ÷ all actual positives.
In practice
If a system finds 80 of 100 invoices containing the target error, recall is 80%.
Related terms
Recommender SystemA system that ranks or suggests items based on their expected relevance to a user or context.
In practice
Recommendations may support content discovery or product selection, with controls for unsuitable or misleading suggestions.
Related terms
Recurrent Neural Network (RNN)A neural network that carries information from earlier steps into later steps when processing a sequence.
In practice
It can model relationships across ordered observations such as sensor readings or words.
Related terms
Red TeamingA structured effort to discover weaknesses by deliberately testing how a system might fail, be misused, or respond to adversarial conditions.
In practice
Authorized testers probe a customer assistant for misleading answers, inappropriate disclosures, and actions outside its approved scope.
Related terms
RegressionIn machine learning, a prediction task that estimates a numerical value. In statistics, regression also refers more broadly to modeling relationships between variables.
In practice
A model estimates delivery time in hours. Linear regression is one method; logistic regression is commonly used for classification.
Related terms
RegularizationMethods that constrain learning or model complexity to reduce overfitting and improve performance on new data.
In practice
Penalties on large weights or stopping training early can help, but their value should be checked on validation data.
Related terms
Reinforcement Learning (RL)Learning how to choose actions through interaction with an environment and feedback expressed as rewards.
In practice
A system may learn a scheduling policy in simulation. A poorly designed reward can encourage behavior that misses the actual business goal.
Related terms
Reinforcement Learning from Human Feedback (RLHF)A training approach that uses human feedback to inform a reward signal for reinforcement learning, helping shape model behavior.
In practice
Human preference feedback can improve responses, but preferred answers are not necessarily accurate, fair, or suitable for every user.
Related terms
ReproducibilityThe ability to obtain consistent or comparable results when a process is repeated with documented data, code, settings, and conditions.
In practice
Version prompts, model settings, evaluation data, and tools so performance changes can be investigated.
Related terms
Residual RiskThe risk that remains after controls and other treatments have been applied.
In practice
A named decision-maker should understand the remaining exposure before approving a deployment or material change.
Related terms
Responsible AIThe practice of developing and using AI with appropriate consideration for people, safety, fairness, privacy, transparency, and accountability.
In practice
Translate principles into assigned responsibilities, design choices, operating controls, and evidence of how the system performs.
Related terms
RetrainingTraining a model again, often with updated data, to maintain or change its performance.
In practice
Retraining should be followed by evaluation and controlled release. It should not be assumed to fix every performance issue.
Related terms
Retrieval-Augmented Generation (RAG)An approach that retrieves relevant information and supplies it to a generative model as context for producing a response.
In practice
An assistant answers from approved company documents. Retrieval can improve grounding, but poor sources or retrieval can still produce wrong answers.
Related terms
Return on Investment (ROI)A measure comparing net benefit with investment cost over a stated period, often calculated as (benefits minus costs) ÷ costs.
In practice
Include implementation, oversight, rework, and ongoing operating costs. Time saved becomes financial value only when it is usefully converted.
Related terms
Risk AppetiteThe amount and types of risk an organization is willing to pursue or retain in meeting its objectives.
In practice
Translate high-level statements into practical limits, approval requirements, and escalation criteria for particular AI uses.
Related terms
Robotic Process Automation (RPA)Software that automates repetitive interactions with applications, often by following explicit rules.
In practice
A bot copies approved data between legacy systems. AI may be added to handle less structured inputs.
Related terms
RollbackReturning a system to a previous working state or switching to a defined fallback when a change or failure makes continued operation unacceptable.
In practice
Reverting software does not necessarily undo messages sent, payments made, or other external actions; those need separate recovery steps.
Related terms
Self-Supervised LearningLearning from targets generated from the structure of the data itself, rather than relying on separate human-created labels for every example.
In practice
A language model can learn by predicting missing or subsequent parts of text.
Related terms
Semantic SearchSearching for information by meaning and context rather than relying only on exact word matches.
In practice
A query about 'taking time off' may retrieve the annual leave policy even if the wording differs.
Related terms
Semi-Supervised LearningTraining approaches that use both labeled and unlabeled data to improve a model.
In practice
A small set of reviewed support messages is combined with a much larger collection of unlabeled messages.
Related terms
Sentiment AnalysisIdentifying attitudes or emotional tone expressed in content, often using categories such as positive, negative, or neutral.
In practice
Service teams can monitor broad feedback patterns, while accounting for sarcasm, mixed opinions, and cultural context.
Related terms
Shadow AIAI tools or uses operating outside an organization's established visibility, approval, or control processes.
In practice
Employees may upload business information into personal AI accounts when approved tools do not meet their needs.
Related terms
SimulationA representation of a process or environment used to explore behavior under specified conditions.
In practice
Test alternative operational decisions in a simulated setting while checking whether its assumptions reflect the real environment.
Related terms
Small Language Model (SLM)A language model with a smaller resource footprint than larger alternatives. The term has no universal size threshold.
In practice
A smaller model may suit a narrow task or local deployment when it meets required quality, cost, and response-time targets.
Related terms
Stage GateA planned decision point where evidence is reviewed before work proceeds to the next phase.
In practice
Before moving from pilot to production, review performance, benefits, ownership, controls, and operational readiness.
Related terms
Statistical DistributionA description of the values a variable can take and how frequently or how probably they occur.
In practice
Check whether production inputs differ from the distributions represented in training and evaluation data.
Related terms
Structured DataData organized according to a defined structure, such as rows and columns with specified fields. Structured data does not have to be labeled for machine learning.
In practice
An order table with dates, quantities, and product codes is structured data.
Related terms
Supervised LearningLearning a mapping from inputs to target outputs using examples that contain both.
In practice
Historical messages paired with their correct service categories can train a routing model.
Related terms
Support Vector Machine (SVM)A learning method that finds a decision boundary with a large margin between classes; related versions also support regression.
In practice
It can be useful for some structured or text tasks, with suitability depending on data size, features, and tuning.
Related terms
Synthetic DataData created through generation or simulation rather than directly collected as real observations.
In practice
It may support testing or training, but is not automatically realistic, unbiased, or free of privacy risks.
Related terms
TemperatureA generation setting that adjusts the distribution used when sampling outputs. Higher values typically allow more variation; effects depend on the model and implementation.
In practice
A lower temperature may make responses more consistent, but does not guarantee factual accuracy or identical results.
Related terms
Test DataData reserved to assess performance after development choices have been made, with separation from training and tuning.
In practice
Repeatedly adjusting the system based on the same test results can weaken the independence of the final assessment.
Related terms
ThresholdA cutoff used to convert a score into a decision, such as whether to flag a case for review.
In practice
Changing a fraud-alert threshold changes the balance between missed problems and unnecessary investigations.
Related terms
ThroughputThe amount of work a system completes in a given time, such as requests or documents processed per minute.
In practice
A service must handle peak demand without unacceptable delays or a decline in output quality.
Related terms
Time SeriesObservations recorded in time order, often at regular intervals.
In practice
Demand forecasting should account for trends, seasonality, and the fact that future information must not leak into earlier predictions.
Related terms
TokenA unit used by a model to represent and process information. For text, a token may be a word, part of a word, punctuation, or another element.
In practice
Token counts influence context limits, usage charges, and response length; they do not correspond to a fixed number of words.
Related terms
Tool UseAn AI application's use of external functions or systems to retrieve information or perform actions.
In practice
A model requests an order lookup; the application checks authorization, executes the call, and returns the result.
Related terms
Topic ModelingMethods for finding recurring themes across a collection of documents.
In practice
A team explores thousands of feedback comments to identify themes, then checks whether the labels and groupings make sense.
Related terms
Total Cost of Ownership (TCO)The full cost of obtaining, operating, maintaining, and retiring a system over a defined period.
In practice
Include integration, data work, licenses, compute, monitoring, human review, change management, and exit costs.
Related terms
TrainingThe process of adjusting a model using data or feedback to improve its performance against an objective.
In practice
Training changes model behavior through parameter updates; simply submitting a request for inference does not necessarily do so.
Related terms
Training DataThe examples used to fit a model's parameters during training.
In practice
Check relevance, representativeness, quality, permissions, and separation from evaluation data.
Related terms
Transfer LearningUsing knowledge or representations learned for one task or dataset to support another.
In practice
An image model pretrained on broad data is adapted to recognize defects in a specific product line.
Related terms
TransformerA neural network architecture built around attention mechanisms that capture relationships among elements of an input.
In practice
Transformers underpin many language models and are also used with images, audio, and other data.
Related terms
TransparencyProviding appropriate information about a system's purpose, use, capabilities, limitations, and responsibilities to relevant people.
In practice
Users may need to know when they are interacting with AI, what it can do, and how to seek human assistance.
Related terms
Turing TestA proposed test of whether a machine's conversation can be distinguished from a human's by an evaluator in a specified setting.
In practice
Appearing human in conversation does not establish truthfulness, safety, understanding, or suitability for a business role.
Related terms
UncertaintyA lack of complete knowledge about an outcome or estimate, arising from factors such as limited data, variability, or model limitations.
In practice
When the evidence is weak or the situation is unfamiliar, the workflow may need more information or human judgment.
Related terms
UnderfittingA failure to learn enough of the relevant structure in the data to perform the task well.
In practice
A model that is too simple or insufficiently trained may perform poorly on both training and new examples.
Related terms
Unstructured DataInformation that does not follow a fixed tabular schema, such as free text, images, audio, or video. It can still be organized and processed.
In practice
A collection of well-edited policy documents is unstructured data, even though it is neither raw nor disorganized.
Related terms
Unsupervised LearningLearning patterns or structure from data without a target label for each example.
In practice
A model groups similar customer records to support exploration, with people assessing the meaning and usefulness of the groups.
Related terms
Use CaseA defined situation in which a system supports a particular user, task, or decision to achieve an intended outcome.
In practice
Describe the problem, users, workflow, expected value, constraints, and success measures before selecting technology.
Related terms
Validation DataData used during development to compare options and tune a system without directly fitting the model's parameters on those examples.
In practice
Use a separate final test set or other appropriate independent assessment after choosing the model and settings.
Related terms
Value RealizationTurning expected benefits into demonstrated improvements in business outcomes after implementation.
In practice
If a tool saves time, establish whether that time improves service, increases capacity, reduces costs, or supports another measurable outcome.
Related terms
Vanishing and Exploding GradientsTraining problems in which calculated gradients become so small or large that useful parameter updates are difficult or unstable.
In practice
These issues affect model design and training reliability rather than being errors a business user can fix with a better prompt.
Related terms
VarianceA measure of how spread out values are around their mean. In model evaluation, high variance can also describe sensitivity to the particular training sample.
In practice
A model that changes substantially when trained on slightly different samples may perform inconsistently on new data.
Related terms
Vector DatabaseA system designed to store and search numerical vectors, often embeddings, according to similarity and other filters.
In practice
An enterprise assistant retrieves relevant document passages while applying access controls and metadata filters.
Related terms
Vendor Due DiligenceThe assessment of a supplier's capabilities, limitations, controls, terms, and dependencies before and during engagement.
In practice
Review evidence relevant to your use case, including data handling, evaluation, service continuity, and exit arrangements.
Related terms
Vendor Lock-InDependence on a supplier that makes moving to an alternative difficult or costly.
In practice
Consider data export, model portability, proprietary integrations, contract terms, and the work required to replace the service.
Related terms
Word Error Rate (WER)A speech-recognition metric calculated from word substitutions, deletions, and insertions relative to the number of words in a reference transcript.
In practice
A low average WER can still hide serious errors in critical names, amounts, or instructions.
Related terms
Workflow RedesignChanging the sequence of tasks, responsibilities, handoffs, and controls to improve how work gets done with AI.
In practice
Redesign the complete service process, including review and exception handling, rather than measuring only the speed of drafting.
Related terms
Zero-Shot PromptingAsking a model to perform a task without providing demonstrations of that task in the prompt. The model still relies on what it learned during training.
In practice
Request a summary by describing the desired audience and format, without including an example summary.
Related terms
Put AI knowledge
into practice
Explore AIACI's programs for the leaders setting direction and the managers responsible for execution.
About this glossary & further reading
Definitions and illustrative examples are written for practical business use. Some terms have different meanings across disciplines. Legal and standards entries summarize concepts; consult the official material for applicable requirements. Selected technical and official references are linked within entries and below.
- Google: Machine Learning Glossary (opens in a new tab)
- Google Cloud: Generative AI Glossary (opens in a new tab)
- NIST: AI Risk Management Framework (opens in a new tab)
- NIST: AI Risks and Trustworthiness (opens in a new tab)
- ISO: ISO/IEC 42001 (opens in a new tab)
- European Commission: AI Act (opens in a new tab)
- European Commission: Data Protection (opens in a new tab)
- OWASP: Prompt Injection (opens in a new tab)
- Anthropic: Building Effective Agents (opens in a new tab)
- Model Context Protocol: Introduction (opens in a new tab)
- Lewis et al.: Retrieval-Augmented Generation (opens in a new tab)
- Vaswani et al.: Attention Is All You Need (opens in a new tab)