THE ENTERPRISE AI FIELD GUIDE
Understand the Language.Connect the Concepts. Find a concept Learning level All learning levels Essential Practitioner Advanced
All Concepts 258 concepts
AI Foundations · ● Essential Artificial intelligence Systems performing tasks associated with human intelligence, such as prediction, perception or language processing.
Machine learning Foundations · ● Essential ML Learning patterns from data rather than specifying every rule manually.
Deep learning Foundations · ◇ Practitioner DL Machine learning using neural networks with multiple layers.
Neural network Foundations · ◇ Practitioner A model made of connected computational units whose weights are learned from data.
Generative AI Foundations · ● Essential GenAI Models that generate new text, images, audio, video or other outputs from learned patterns.
Predictive AI Foundations · ● Essential Models that estimate outcomes, classes or numerical values from input data.
Foundation model Foundations · ● Essential A broadly trained model that can be adapted to multiple downstream tasks.
LLM Foundations · ● Essential Large language model A large model trained to process and generate language; some also handle other modalities.
SLM Foundations · ◇ Practitioner Small language model A comparatively smaller language model; there is no universal parameter cutoff.
NLP Foundations · ◇ Practitioner Natural language processing Methods for analysing, understanding and generating human language.
AGI Foundations · △ Advanced Artificial general intelligence A proposed form of broadly capable general intelligence; definitions and measurement remain contested.
ASI Foundations · △ Advanced Artificial superintelligence A hypothetical intelligence exceeding human capabilities across a broad range of tasks.
Model vs application Foundations · ● Essential The model supplies learned capabilities; an application combines it with interfaces, data, tools and controls.
Supervised learning Foundations · ● Essential Learning from examples that pair inputs with target labels or values, so a model can predict outcomes for new inputs. Training performance alone does not establish how well it generalises.
Unsupervised learning Foundations · ● Essential Learning patterns or structure from data without supplied target labels, for example grouping similar records into clusters. The resulting patterns still need interpretation and validation.
Transformer Inside the model · ◇ Practitioner A neural-network architecture using attention to process relationships in sequences.
Attention Inside the model · △ Advanced A mechanism that weights which parts of an input are relevant to a representation.
Parameters Inside the model · ◇ Practitioner Learned numerical values that shape model behaviour, including weights and biases. See Weights; parameter count alone does not determine quality.
Weights Inside the model · ◇ Practitioner The learned parameters that scale connections or transformations inside a model. Weights are a type of parameter; see Parameters.
Token Inside the model · ● Essential A unit used to encode input or output, such as part of a word; token is not synonymous with word.
Tokenisation / Tokenizer Inside the model · ● Essential Tokenization Tokenisation is the process of converting content into the token units a model uses; a tokenizer implements that conversion. Token counts vary by model, language and content, affecting context usage and token-based costs. A token is not necessarily a word.
Embedding Inside the model · ◇ Practitioner A numerical representation that captures useful features or relationships of content.
Context window Inside the model · ● Essential The maximum context a model can process in an invocation; it is not persistent memory.
MoE Inside the model · △ Advanced Mixture of experts An architecture routing inputs to subsets of specialist model components.
Dense model Inside the model · △ Advanced A model architecture where the main parameter blocks participate for each token, unlike sparse expert routing.
Open weights Inside the model · ● Essential Model weights are available under a licence; source code, training data and unrestricted use may not be.
Open-source AI Inside the model · ◇ Practitioner AI released under terms and with materials meeting an applicable open-source definition; inspect the actual licence.
Model card Inside the model · ● Essential Documentation of a model’s intended uses, evaluations, limitations and other relevant characteristics.
Knowledge cutoff Inside the model · ● Essential The documented limit of a model version’s training knowledge. It does not guarantee complete or correct knowledge before that date. Retrieval, browsing or supplied context can provide newer information without updating the model’s training cutoff; consult the model documentation rather than its self-description.
Input vs output tokens Inside the model · ● Essential Input tokens represent content supplied to a model; output tokens are generated during its response. Rates and accounting can differ, including for cached input and reasoning tokens. Check provider usage reports rather than estimating cost from visible answer length alone.
Pretraining Training and adaptation · ◇ Practitioner Initial broad training that builds a model’s general capabilities.
Training vs inference Training and adaptation · ● Essential Training learns or updates parameters; inference uses a trained model to produce results.
Fine-tuning Training and adaptation · ● Essential Additional training that changes model weights for a desired task, behaviour or domain.
SFT Training and adaptation · ◇ Practitioner Supervised fine-tuning Training on examples of desired input-output behaviour.
RL Training and adaptation · △ Advanced Reinforcement learning Learning a policy through reward signals from actions or generated outputs.
RLHF Training and adaptation · ◇ Practitioner Reinforcement learning from human feedback Using human preference feedback to help guide model training, often through a learned reward model.
RLAIF Training and adaptation · △ Advanced Reinforcement learning from AI feedback Using AI-generated feedback as a reward or preference signal.
DPO Training and adaptation · △ Advanced Direct preference optimization Training directly on preferred and dispreferred responses without a separate reinforcement-learning optimisation loop.
PEFT Training and adaptation · △ Advanced Parameter-efficient fine-tuning Adapting a model by training a limited subset or additional parameters.
LoRA Training and adaptation · △ Advanced Low-rank adaptation A PEFT technique that learns small low-rank weight updates.
Distillation Training and adaptation · ◇ Practitioner Training a student model to reproduce useful behaviour from a teacher model.
Quantisation Training and adaptation · ◇ Practitioner Representing weights or activations at lower numerical precision to reduce resource requirements.
Synthetic data Training and adaptation · ◇ Practitioner Artificially generated training or evaluation data; quality and representativeness need checks.
Overfitting Training and adaptation · ◇ Practitioner Learning training-specific patterns so strongly that performance on new data suffers.
Data contamination Training and adaptation · ◇ Practitioner Training data overlaps evaluation content, potentially inflating reported performance.
Model collapse Training and adaptation · △ Advanced Degradation that can occur when successive models learn heavily from generated data and lose aspects of the original data distribution. Synthetic data does not inevitably cause collapse; selection and mixing matter.
Training-data licensing Training and adaptation · ◇ Practitioner Permissions and contractual conditions governing use of material for model training. Public availability is not blanket permission, and a dataset licence is separate from a model licence or output rights. Applicable legal exceptions and restrictions depend on jurisdiction and use.
Self-supervised learning Training and adaptation · ◇ Practitioner Learning using targets derived from the data itself, such as predicting a hidden or next token. It reduces dependence on manually labelled examples but still requires suitable data, objectives and evaluation.
Catastrophic forgetting Training and adaptation · ◇ Practitioner Loss of previously learned capabilities when a model is trained on new tasks or data. Evaluate both the new task and retained capabilities after adaptation; this differs from information falling outside a conversation context window.
Prompt Generation and reasoning · ● Essential Instructions and input provided to a model.
System instructions Generation and reasoning · ● Essential High-level directions used by an application to guide model behaviour; precedence depends on the system.
Completion Generation and reasoning · ◇ Practitioner Generated output responding to an input, historically often a continuation of text.
Temperature Generation and reasoning · ◇ Practitioner A sampling setting that changes output variability; low values do not guarantee correctness.
Top-p Generation and reasoning · △ Advanced Nucleus sampling Sampling from a set of candidate tokens covering a specified cumulative probability.
Hallucination Generation and reasoning · ● Essential An output that invents or misstates information, sometimes with convincing detail.
Structured output Generation and reasoning · ● Essential Output constrained to a defined format or schema; format compliance does not ensure factual accuracy.
Reasoning model Generation and reasoning · ● Essential A model designed or trained for extended problem-solving; it can still make reasoning errors.
Chain of thought Generation and reasoning · ◇ Practitioner CoT Intermediate reasoning expressed as steps; visible steps are not a guaranteed faithful account of internal computation.
Test-time compute Generation and reasoning · △ Advanced Inference-time compute Additional computation during use, such as more reasoning or candidate evaluation.
Self-consistency Generation and reasoning · △ Advanced Comparing multiple generated reasoning paths or answers to select an agreed result.
Multimodal Generation and reasoning · ● Essential Processing or generating more than one modality, such as text, images and audio.
VLM Generation and reasoning · ◇ Practitioner Vision-language model A model combining visual information with language understanding or generation.
OCR Generation and reasoning · ◇ Practitioner Optical character recognition Extracting machine-readable text from images or scanned documents.
ASR / TTS Generation and reasoning · ◇ Practitioner Automatic speech recognition / text-to-speech Converting speech into text / converting text into spoken audio.
Diffusion model Generation and reasoning · △ Advanced A generative model learning to reverse a noise process, widely used for image generation.
Sycophancy Generation and reasoning · ● Essential A model’s tendency to agree with or flatter a user instead of giving an accurate, independent response, including adopting the user’s mistaken beliefs.
Top-k sampling Generation and reasoning · ◇ Practitioner A generation method that samples the next token from the k highest-probability candidates. It limits the candidate set but does not guarantee correct answers. This is different from retrieval top-k, which limits retrieved results.
Constrained decoding Generation and reasoning · ◇ Practitioner Restricting allowable token choices during generation to satisfy specified rules, such as a grammar or output schema. It is one way to produce structured output; valid structure does not establish factual correctness or safe actions.
Prompt engineering Prompts and context · ● Essential Designing and testing instructions and examples to improve model responses.
Context engineering Prompts and context · ● Essential Designing what information reaches a model: instructions, retrieved content, tools, history and state.
Zero-shot Prompts and context · ◇ Practitioner Requesting a task without showing worked examples in the prompt.
Few-shot Prompts and context · ◇ Practitioner Providing a small number of examples to guide the model during an invocation.
Prompt template Prompts and context · ◇ Practitioner A reusable instruction structure with fields populated for a task.
Prompt chaining Prompts and context · ◇ Practitioner Passing outputs through a sequence of model calls with different responsibilities.
Context compaction Prompts and context · ◇ Practitioner Summarising or restructuring context to fit space limits while retaining important information.
Context rot Prompts and context · △ Advanced An informal term for degraded use of relevant information in long or cluttered context.
Prompt caching Prompts and context · ◇ Practitioner Reusing computation for repeated prompt prefixes where supported; distinct from caching final answers.
In-context learning Prompts and context · ◇ Practitioner Adapting response behaviour from examples in context without updating model weights.
Instruction hierarchy Prompts and context · ◇ Practitioner Rules for resolving instruction priority; untrusted retrieved text should not override trusted instructions.
Context budget Prompts and context · ◇ Practitioner Allocation of available tokens across instructions, sources, history, tools and output.
Prompt delimiters Prompts and context · ● Essential Markers such as XML tags or clear section boundaries that separate instructions, source material and examples. They can improve interpretation, but are not an enforced security boundary against prompt injection.
Context selection Prompts and context · ◇ Practitioner Choosing which instructions, history, retrieved content and tool results to include for a task. Prioritise relevance, permissions and freshness within the context budget; including more material can add cost and distraction.
RAG Knowledge and retrieval · ● Essential Retrieval-augmented generation Retrieving relevant external information and providing it to a model to support generation; it does not retrain the model.
Grounding Knowledge and retrieval · ● Essential Anchoring an output in supplied sources or evidence; source quality still matters.
Chunking Knowledge and retrieval · ◇ Practitioner Splitting content into retrievable sections with enough meaning and context.
Vector database Knowledge and retrieval · ◇ Practitioner A store supporting search over vector representations, often used with embeddings.
Semantic search Knowledge and retrieval · ● Essential Retrieval based on meaning or learned similarity rather than exact word matches alone.
Keyword search Knowledge and retrieval · ◇ Practitioner Retrieval based on words and lexical matching, useful for exact names and identifiers.
Hybrid search Knowledge and retrieval · ◇ Practitioner Combining lexical and semantic retrieval signals.
Reranking Knowledge and retrieval · ◇ Practitioner Reordering retrieved candidates with an additional relevance model or scoring method.
Metadata filtering Knowledge and retrieval · ◇ Practitioner Restricting retrieval using fields such as date, region, document type or access rules.
Knowledge graph Knowledge and retrieval · ◇ Practitioner A representation of entities and their relationships.
GraphRAG Knowledge and retrieval · ◇ Practitioner RAG approaches using graph structures or graph-derived summaries to improve retrieval and synthesis.
Agentic RAG Knowledge and retrieval · ◇ Practitioner Retrieval where an agent selects searches, sources or follow-up steps dynamically.
Provenance Knowledge and retrieval · ● Essential The origin and processing history of data or an answer’s supporting evidence.
Permission-aware retrieval Knowledge and retrieval · ● Essential Enforcing user access rights while retrieving content, including indexes and downstream results.
Retrieval evaluation Knowledge and retrieval · ◇ Practitioner Measuring whether retrieval finds relevant evidence before evaluating the generated answer.
Retrieval freshness Knowledge and retrieval · ◇ Practitioner How current the indexed sources and retrieved evidence are relative to the task. Updates, deletions, changed permissions and stale caches must be reflected in retrieval.
Query rewriting Knowledge and retrieval · ◇ Practitioner Rephrasing or expanding a search request to improve retrieval, for example resolving a reference or generating alternative wording. Preserve the user’s intent and evaluate whether changes actually improve results.
Query decomposition Knowledge and retrieval · ◇ Practitioner Breaking a complex question into smaller retrieval questions, then combining the evidence needed for the answer. It can improve coverage but adds retrieval work, latency and opportunities for error.
Retrieval top-k Knowledge and retrieval · ◇ Practitioner The number of highest-ranked candidates requested from a retrieval stage. Increasing k can improve coverage but also add irrelevant content and processing cost. Ranking, filtering and reranking affect what ultimately reaches the model.
AI agent Agents and orchestration · ● Essential A system using a model to choose actions and tools toward a goal within configured boundaries.
Agentic AI Agents and orchestration · ● Essential AI systems with goal-directed action and varying autonomy; industry usage is not uniform.
Copilot Agents and orchestration · ● Essential A product pattern where AI assists a person; actual autonomy varies by implementation.
Workflow Agents and orchestration · ● Essential A predefined sequence or graph of tasks, possibly containing model calls and conditional branches.
Autonomy Agents and orchestration · ● Essential The degree of action a system can take without intervention; define it per task and risk.
Planning Agents and orchestration · ◇ Practitioner Forming and revising steps to achieve a goal.
Tool calling Agents and orchestration · ● Essential Function calling A model requests an operation through a defined tool interface; execution is handled by the application or provider-hosted service.
Orchestration Agents and orchestration · ◇ Practitioner Coordinating models, tools, state and task execution.
ReAct Agents and orchestration · △ Advanced Reasoning and acting A pattern interleaving model reasoning with actions and observations.
Multi-agent system Agents and orchestration · ◇ Practitioner Multiple agents collaborating or specialising; more agents do not automatically improve performance.
Handoff Agents and orchestration · ◇ Practitioner Transferring task responsibility and relevant context to another agent or person.
Agent memory Agents and orchestration · ● Essential Information retained for later use, such as preferences or task history. Personal and shared memory have different audiences: define who can read, reuse, change or delete it, and check retention and confidentiality boundaries. Permission-aware retrieval does not automatically establish memory privacy. Microsoft’s Frontier agent-account preview uses shared memory even in one-to-one chats; consult its specific safeguards and limitations.
Agent state Agents and orchestration · ◇ Practitioner Current task data, intermediate results and execution status.
Agent harness Agents and orchestration · ◇ Practitioner The software runtime around an AI model that coordinates model calls, tool use, context, task state and execution controls. A harness supports agent operation but does not guarantee safe or reliable outcomes. Examples of SDKs and frameworks providing these capabilities include Microsoft Agent Framework, Google ADK, Claude Agent SDK and OpenAI Agents SDK; their scope and controls differ.
Agent skills Agents and orchestration · ◇ Practitioner Reusable instructions and resources for specialised tasks; packaging varies across systems.
Human-in-the-loop Agents and orchestration · ● Essential HITL A person participates in decisions, reviews or approvals at defined points.
Stop conditions Agents and orchestration · ◇ Practitioner Rules ending execution, such as success, timeout, cost budget or maximum steps.
Microsoft Copilot Landscape Agents and orchestration · ● Essential Microsoft AI experiences by purpose Microsoft uses the Copilot name across productivity, development, security, cloud, data and business platforms. Group these experiences by the work they support. Copilot experiences can include agents; Copilot Studio is a platform for building and managing agents.
Agent loop Agents and orchestration · ◇ Practitioner A repeated cycle in which an agent processes context, chooses an action, receives results and decides whether to continue. Bound the loop with permissions, budgets and stop conditions; repetition alone does not ensure progress.
Agent delegation Agents and orchestration · ◇ Practitioner Assigning a bounded task to another agent with defined inputs, permissions and expected results. An orchestrator may retain overall control and review the result. A handoff can instead transfer control; terminology varies across frameworks.
API Connections and actions · ● Essential Application programming interface A defined interface through which software requests data or operations.
SDK Connections and actions · ◇ Practitioner Software development kit Libraries and utilities for building against a platform or interface.
MCP Connections and actions · ● Essential Model Context Protocol An open protocol connecting AI applications to external tools, resources and prompts; it does not itself grant trust or permissions.
MCP host / client / server Connections and actions · ◇ Practitioner The application hosting AI / its connection component / the provider exposing capabilities.
MCP resources / tools Connections and actions · ◇ Practitioner Resources expose context or data; tools expose callable operations.
A2A Connections and actions · ◇ Practitioner Agent2Agent protocol An open protocol supporting communication and interoperability between agents.
Agent discovery Connections and actions · ◇ Practitioner Finding agents and learning their capabilities and interaction requirements.
Connector Connections and actions · ● Essential An integration linking an application with an external system; not every connector uses MCP.
Computer use Connections and actions · ● Essential AI interaction with graphical interfaces, often using screenshots, clicks and typing.
Sandbox Connections and actions · ● Essential An isolated execution environment limiting impact; isolation strength depends on implementation.
OAuth Connections and actions · ◇ Practitioner A framework for delegated authorisation, often used to give scoped access without sharing passwords.
Least privilege Connections and actions · ● Essential Granting only the access needed for a particular task and duration.
Agent identity Connections and actions · ◇ Practitioner An identifiable principal or credential context for agent access, audit and accountability. Distinguish an agent identity from a dedicated account with resources such as a mailbox. Microsoft’s Copilot Studio agent-account feature (Frontier preview) has a separately managed Entra Agent ID. Assign owners, inspect the identity and permissions each tool uses, review access and revoke associated connections and resources at retirement.
Idempotency Connections and actions · ◇ Practitioner An operation has the same intended effect when repeated as when performed once. Idempotency controls help make retries safe, such as preventing the same refund from being issued twice.
Microsoft Graph Connections and actions · ◇ Practitioner Microsoft’s API and service platform for authorised access to data and capabilities across Microsoft 365 and other Microsoft cloud services. It supports integrations and work-context experiences; it is not an AI model or the same concept as a knowledge graph.
Authentication / authorisation Connections and actions · ● Essential Authentication verifies an identity; authorisation determines which resources and actions it may access. An authenticated agent is not automatically permitted to read every document or execute every tool. See least privilege and agent identity.
Non-human identity / workload identity Connections and actions · ◇ Practitioner NHI Non-human identities represent software or devices rather than people. Workload identities represent applications, services and other software workloads and are a subset of non-human identities. Agent identities need appropriate ownership, permissions and lifecycle controls.
Webhook Connections and actions · ◇ Practitioner An event-triggered HTTP notification sent to a configured endpoint. It can start an automation or agent workflow. Verify the sender, validate the payload and handle retries safely; receiving an event does not itself authorise an action.
Responsible AI Responsible AI and security · ● Essential Practices for developing and operating AI with attention to fairness, safety, privacy, transparency and accountability.
AI governance Responsible AI and security · ● Essential Decision rights, policies, ownership and oversight across the AI lifecycle.
AI ethics Responsible AI and security · ● Essential Consideration of moral consequences, values and impacts on people and society.
Trustworthy AI Responsible AI and security · ● Essential AI assessed against qualities such as reliability, safety, security and accountability; not a blanket certification.
Fairness / bias Responsible AI and security · ● Essential Evaluating unjust differences or systematic skew; fairness criteria can conflict and depend on context.
Explainability Responsible AI and security · ● Essential Making outputs or behaviour understandable to relevant audiences; generated explanations require scrutiny.
Transparency Responsible AI and security · ● Essential Disclosing relevant information about AI use, capabilities, data and limitations.
Accountability Responsible AI and security · ● Essential Assigning responsibility and mechanisms for review, challenge and remediation.
AI safety Responsible AI and security · ● Essential Reducing harmful behaviour and consequences during development and use.
Alignment Responsible AI and security · ◇ Practitioner Seeking model behaviour consistent with intended goals, instructions and values.
Guardrails Responsible AI and security · ● Essential Controls that constrain, detect or respond to undesirable inputs, actions and outputs.
Prompt injection Responsible AI and security · ● Essential Malicious or untrusted content attempting to redirect model behaviour, including through retrieved documents.
Jailbreak Responsible AI and security · ◇ Practitioner An attempt to bypass a model’s behavioural restrictions.
Data exfiltration Responsible AI and security · ● Essential Unauthorised transfer of information out of its permitted environment.
PII Responsible AI and security · ● Essential Personally identifiable information Information that identifies or can be linked to a person; applicable legal definitions vary.
DLP Responsible AI and security · ◇ Practitioner Data loss prevention Controls to detect and restrict inappropriate disclosure or movement of sensitive data.
Data residency Responsible AI and security · ◇ Practitioner Where data is stored or processed; check each service and data flow.
Red teaming Responsible AI and security · ◇ Practitioner Adversarial testing to discover weaknesses and harmful behaviours.
AI impact assessment Responsible AI and security · ◇ Practitioner A structured review of intended use, affected people, risks and mitigation measures.
Shadow AI Responsible AI and security · ● Essential AI used outside approved organisational visibility or controls.
NIST AI RMF Responsible AI and security · ◇ Practitioner AI Risk Management Framework A voluntary framework organised around Govern, Map, Measure and Manage.
Content credentials Responsible AI and security · ◇ Practitioner Provenance information about media creation and edits; provenance does not prove every claim is true.
EU AI Act Responsible AI and security · ◇ Practitioner Regulation (EU) 2024/1689 EU legislation setting risk-based requirements for AI systems and obligations for general-purpose AI models. Applicability depends on the actor, use and scope; check current official guidance and implementation dates.
EN 18286 Responsible AI and security · ◇ Practitioner AI quality management for EU AI Act regulatory purposes A European standard for quality management systems supporting providers of high-risk AI systems under the EU AI Act. EN 18286:2026 supports presumption of conformity with the Article 17 requirements it covers; it does not establish compliance with the entire Act or replace ISO/IEC 42001.
ISO/IEC 27001 Responsible AI and security · ◇ Practitioner Information security management system requirements Requirements for establishing, maintaining and continually improving an information security management system (ISMS). Organisations can seek certification for a defined management-system scope; certification does not guarantee that every product or AI output is secure.
ISO/IEC 27002 Responsible AI and security · ◇ Practitioner Information security control guidance Practical guidance on information security controls, including access management, encryption, supplier security and incident response. Supports the controls used within an information security management system; it is guidance rather than a standalone certification standard.
ISO/IEC 23894 Responsible AI and security · ◇ Practitioner AI risk management guidance Guidance for organisations developing or using AI to incorporate AI-related risk management into their activities. It can be adapted to the organisation and use case. It addresses broader AI risk management rather than only cybersecurity threats.
ISO/IEC 42001 Responsible AI and security · ◇ Practitioner AI management system standard An international standard specifying requirements for an organisational AI management system. It concerns management processes and continual improvement, not a guarantee that every model or output is safe.
ISO/IEC 27090 Responsible AI and security · ◇ Practitioner AI security guidance Guidance for assessing and reducing cybersecurity risks to AI systems across their lifecycle, including data poisoning and model theft. Unlike ISO/IEC 42001’s AI management-system requirements, it focuses on AI security threats. It does not certify that a system is secure.
IEEE 2863 Responsible AI and security · ◇ Practitioner Organisational AI governance recommended practice IEEE 2863:2026 provides governance principles, implementation processes and case studies for organisations developing or using AI. It helps translate responsible AI principles into organisational practice, including addressing trade-offs; it is a recommended practice, not legislation.
Tool poisoning Responsible AI and security · ◇ Practitioner An indirect prompt-injection attack in which malicious instructions are embedded in tool descriptions or metadata to redirect an agent’s behaviour. Reviewing tools requires checking what they describe as well as what they execute.
Data poisoning Responsible AI and security · ◇ Practitioner Manipulating training data or a knowledge source to influence model behaviour or answers. Training-data poisoning and retrieval-source poisoning act at different stages.
Data minimisation Responsible AI and security · ● Essential Limiting personal data collection, processing and retention to what is necessary for a defined purpose. Applied to AI, this includes prompts, retrieved content, logs and memory.
Oversharing Responsible AI and security · ● Essential Information made accessible to more people or systems than the intended business purpose requires. AI retrieval can make existing excessive access easier to discover; it does not mean that the AI bypasses permissions.
Sensitivity labels Responsible AI and security · ◇ Practitioner Organisation-defined classifications that can apply handling and protection settings to supported content or containers. Protection depends on configuration and platform support; a label is not the same as access permissions, DLP or a retention policy.
Synthetic media Responsible AI and security · ● Essential Images, audio, video or other media generated or substantially altered using AI. Uses can be legitimate or misleading; consider consent, disclosure, provenance and rights for the intended use.
Deepfakes Responsible AI and security · ● Essential AI-generated or manipulated media that realistically depicts a person saying or doing something they did not. It may be authorised or deceptive; verify sensitive requests through an independent trusted channel rather than relying on apparent likeness or a detector alone.
Copyright and AI Responsible AI and security · ◇ Practitioner Copyright questions can arise in source material, training and generated outputs. Rights depend on the applicable law, human contribution, use and terms; AI generation does not automatically establish ownership or freedom to reuse content. Official references may be jurisdiction-specific.
GRC Responsible AI and security · ◇ Practitioner Governance, Risk and Compliance An organisational approach connecting accountability and decision-making, risk assessment and treatment, and compliance with applicable laws, contracts and policies. GRC combines people, processes and evidence; it is not simply a software tool.
AI system inventory Responsible AI and security · ◇ Practitioner A maintained record of AI systems, their purposes, owners, dependencies and lifecycle status. It supports oversight and decommissioning; a list of model names alone does not capture the applications and agents using them.
Risk register Responsible AI and security · ◇ Practitioner A structured record of identified risks, their potential consequences, accountable owners, planned treatments and review status. For AI, record risks in the deployment context rather than assuming a model-level assessment covers every use.
Risk appetite / risk tolerance Responsible AI and security · ◇ Practitioner Risk appetite expresses the types and amount of risk an organisation is willing to take in pursuing its objectives. Risk tolerance translates that direction into specific acceptable boundaries. Neither overrides applicable legal obligations.
Residual risk Responsible AI and security · ◇ Practitioner Risk that remains after controls or other risk treatments are applied. It should be reviewed against agreed boundaries and assigned to an accountable owner; implementing a control does not mean the remaining risk is automatically acceptable.
Third-party risk management Responsible AI and security · ◇ Practitioner TPRM Assessing and managing risks arising from suppliers and external services throughout their lifecycle. For AI, this includes model and data providers, contractual safeguards, service changes, dependencies and exit arrangements.
Data retention and deletion Responsible AI and security · ◇ Practitioner Rules and mechanisms governing how long data is kept and when it is removed. For AI, check prompts, outputs, memory and logs separately. Deleting a visible conversation does not necessarily remove retained copies; holds and service-specific policies may apply.
Zero Trust Responsible AI and security · ● Essential A security approach that explicitly verifies access, limits privileges and assumes compromise is possible. Apply it to people, devices and AI workloads; being inside a network does not by itself establish trust.
IRM Responsible AI and security · ◇ Practitioner Insider Risk Management Identifying, assessing and addressing potential risks arising from people with legitimate access, including inadvertent and malicious activity. Use proportionate controls, privacy safeguards and authorised investigation processes; a risk indicator is not proof of wrongdoing.
Lines of defence Responsible AI and security · ◇ Practitioner LoD · Three Lines Model A way to distinguish governance roles: first-line management delivers services and manages risks; second-line roles provide expertise, support, monitoring and challenge; internal audit provides independent third-line assurance. The governing body oversees the organisation. The IIA now calls this the Three Lines Model; it is not three layers of technical security.
Risk treatment Responsible AI and security · ◇ Practitioner Choosing and implementing a response to an identified risk, such as avoiding an activity, reducing exposure, sharing risk or accepting it within authorised boundaries. Record ownership and residual risk; transferring a service does not automatically transfer accountability.
Control effectiveness Responsible AI and security · ◇ Practitioner The extent to which a safeguard achieves its intended objective in practice. Assess its design and operation using evidence, testing and review. A configured policy or a completed checklist does not by itself demonstrate that the control works.
Evals Evaluation and assurance · ● Essential Evaluations Repeatable tests of whether an AI system meets specific quality, safety and task requirements.
Benchmark Evaluation and assurance · ● Essential A standardised test set or task suite used for comparison; it may not predict your production outcomes.
Golden dataset Evaluation and assurance · ◇ Practitioner A curated reference set of test cases with reviewed expected answers or criteria.
Groundedness / faithfulness Evaluation and assurance · ◇ Practitioner How well an answer is supported by or consistent with supplied evidence; definitions vary by evaluator.
Precision / recall Evaluation and assurance · ◇ Practitioner Precision measures the relevant share of selected items; recall measures the retrieved share of all relevant items.
Task success Evaluation and assurance · ● Essential Whether the complete user goal was achieved within constraints.
Tool accuracy Evaluation and assurance · ◇ Practitioner Whether the agent selected the appropriate tool and supplied correct arguments.
LLM-as-judge Evaluation and assurance · ◇ Practitioner Using a model to score outputs; calibrate against human review and check judge bias.
Human evaluation Evaluation and assurance · ● Essential People review outputs using defined criteria; agreement and reviewer expertise matter.
Regression testing Evaluation and assurance · ◇ Practitioner Checking that changes do not break previously acceptable behaviour.
Adversarial evaluation Evaluation and assurance · ◇ Practitioner Testing deliberately difficult or hostile inputs, permissions and tool interactions.
Calibration Evaluation and assurance · △ Advanced Whether expressed or estimated confidence corresponds to observed correctness.
A/B testing Evaluation and assurance · ◇ Practitioner Comparing alternative systems or configurations on controlled groups or traffic.
Data / concept drift Evaluation and assurance · ◇ Practitioner Changes in input distributions or in relationships between inputs and outcomes.
Model risk management Evaluation and assurance · ◇ Practitioner Identifying, validating, monitoring and governing risks from model use.
Abstention Evaluation and assurance · ◇ Practitioner A system declining to answer or act when evidence, confidence or permissions are insufficient, ideally with a useful clarification or human escalation path.
Audit trail Evaluation and assurance · ◇ Practitioner Records that allow relevant actions and changes to be reconstructed, including who or what acted, when and with which result. For AI agents, capture appropriate tool actions and approvals while protecting sensitive data; diagnostic logs alone may not provide sufficient audit evidence.
AI assurance Evaluation and assurance · ◇ Practitioner Building and maintaining evidence-based confidence that an AI system meets specified requirements in its intended context. Evaluations, control testing and ongoing review contribute to assurance; a benchmark score or security assessment alone does not establish overall fitness for use.
Attestation Evaluation and assurance · ◇ Practitioner A documented statement that specified requirements or conditions have been met, based on an identified review and evidence. In governance, an owner may attest that access remains appropriate. Its scope, date, signer and evidence matter; an attestation is not automatically an independent audit or certification. Technical device attestation is a different use of the term.
Evaluation coverage Evaluation and assurance · ◇ Practitioner How well an evaluation set represents relevant tasks, users, conditions and failure modes for the intended use. High scores on a narrow set can miss important risks; revisit coverage when models, workflows or user populations change.
MLOps Production and performance · ◇ Practitioner Machine learning operations Practices for building, deploying, monitoring and maintaining ML systems.
LLMOps Production and performance · ◇ Practitioner Large language model operations Operational practices for LLM applications, including prompts, retrieval, evaluation and deployment.
AgentOps Production and performance · ◇ Practitioner Agent operations An emerging umbrella term for operating agents: traces, tools, state, quality, cost and controls.
Observability Production and performance · ● Essential Understanding system behaviour through logs, metrics, traces and other signals.
Tracing Production and performance · ◇ Practitioner Recording the sequence of model calls, retrieval, tools and other steps in a task.
Latency Production and performance · ● Essential Time to respond or complete a task; measure user-visible end-to-end time.
TTFT Production and performance · ◇ Practitioner Time to first token Delay before the first generated token reaches the user.
Throughput Production and performance · ◇ Practitioner Work completed per unit time, such as tokens or requests per second.
Rate limits Production and performance · ● Essential Restrictions on request or token volume over a period.
Streaming Production and performance · ◇ Practitioner Delivering output incrementally as it is generated.
Batch inference Production and performance · ◇ Practitioner Processing groups of requests, often with different cost and turnaround characteristics.
Model routing Production and performance · ◇ Practitioner Selecting models based on task, risk, capability, cost or latency requirements.
Fallback Production and performance · ◇ Practitioner An alternative model, workflow or human route when the primary path fails.
Semantic cache Production and performance · △ Advanced Reusing answers for meaningfully similar queries; requires freshness and permission safeguards.
GPU / accelerator Production and performance · ◇ Practitioner Hardware used to speed model training or inference.
Edge / on-device AI Production and performance · ◇ Practitioner Running AI close to the user or device, potentially reducing network reliance and data transfer.
Versioning / rollback Production and performance · ● Essential Tracking changes and restoring an earlier model, prompt, index or configuration when necessary.
Kill switch Production and performance · ● Essential A control that stops an agent or disables actions when intervention is needed.
AI gateway / LLM gateway Production and performance · ◇ Practitioner An intermediary between applications and AI providers that can centralise routing, authentication, monitoring and usage controls. Features vary; a gateway alone does not secure the whole AI system.
Business continuity planning Production and performance · ◇ Practitioner BCP / BCDR Planning how essential business activities will continue during disruption. Define critical work, accountable owners, acceptable service levels, alternative working arrangements and exercises. Disaster recovery is the technology-restoration part of the wider continuity approach.
Disaster recovery Production and performance · ◇ Practitioner DR The plans and capabilities used to restore technology services and data after a major disruption. Specify recovery priorities, dependencies, procedures and acceptance checks. A documented plan needs testing; backups alone do not establish recoverability.
Business impact analysis Production and performance · ◇ Practitioner BIA An assessment of how disruption affects business activities over time. Use it to prioritise critical processes, identify people and service dependencies, and set justified recovery requirements with business owners.
RTO / RPO Production and performance · ◇ Practitioner Recovery time objective / recovery point objective RTO sets the target time for restoring a service after disruption. RPO sets the maximum acceptable data loss measured in time. Agree scope and targets with the business, then validate achievable recovery through tests; objectives are not guarantees.
Enterprise recovery Production and performance · ◇ Practitioner Coordinated restoration of business operations across people, processes, technology, data and suppliers. Set recovery order, decision authority, communications and business acceptance, including reconciliation of interrupted work. This is a broad operating concept; organisations define its scope differently.
Cyber recovery Production and performance · ◇ Practitioner Restoring trusted operations after a cyber incident. Alongside service restoration, address the compromise, verify recovery data and systems, and validate identities and access before reconnecting. Speed alone is insufficient if the recovered environment remains compromised.
High availability Production and performance · ◇ Practitioner HA Designing a service to remain available through expected component failures, using redundancy and appropriate failover. It reduces interruption but does not replace disaster recovery or business continuity planning; shared dependencies can still fail.
Backup and restore Production and performance · ● Essential A backup is a retained copy or recovery point used to restore data or system state. Test restoration, access and integrity against recovery needs. Synchronisation and replication may propagate deletion or corruption; neither alone proves that a usable historical recovery point exists.
AI FinOps Economics and enterprise adoption · ◇ Practitioner Financial accountability and optimisation for AI consumption linked to business value.
Token pricing Economics and enterprise adoption · ● Essential Charges for processed or generated tokens; input, output, cached and reasoning usage may differ by service.
Cost per successful task Economics and enterprise adoption · ◇ Practitioner Total relevant cost divided by successful outcomes, including retries and human review.
TCO Economics and enterprise adoption · ◇ Practitioner Total cost of ownership Full lifecycle costs including licences, integration, infrastructure, operations, governance and people.
ROI Economics and enterprise adoption · ● Essential Return on investment Net benefit divided by investment cost, often expressed as a percentage. Define the benefits, costs and time horizon; productivity gains are not automatically cash savings.
Consumption vs licence Economics and enterprise adoption · ● Essential Usage-based charges / subscription or entitlement charges; deployments may combine both.
Budget / quota Economics and enterprise adoption · ● Essential A spending allowance / a usage or capacity allocation; an alert is not necessarily an enforced cap.
AI CoE Economics and enterprise adoption · ● Essential AI centre of excellence A coordinating function for standards, expertise, delivery and governance, with defined decision rights.
Use-case portfolio Economics and enterprise adoption · ◇ Practitioner A managed set of AI opportunities prioritised by value, feasibility and risk.
AI readiness Economics and enterprise adoption · ● Essential Preparedness across data, technology, security, skills, governance and operating processes.
AI literacy Economics and enterprise adoption · ● Essential The ability to understand, use and critically assess AI in relevant work.
Adoption vs value Economics and enterprise adoption · ● Essential Usage measures engagement; outcome measures establish whether AI improves the work.
Change management Economics and enterprise adoption · ● Essential Preparing people and processes for new ways of working and sustained adoption.
Human oversight Economics and enterprise adoption · ● Essential Designated monitoring, intervention and accountability; may extend beyond individual approval steps.
Build / buy Economics and enterprise adoption · ● Essential Choosing custom development, purchased capability or a combination based on lifecycle needs.
Vendor lock-in Economics and enterprise adoption · ● Essential Dependency that makes changing providers difficult due to interfaces, data, skills or commercial terms.
Pilot vs production Economics and enterprise adoption · ● Essential A limited trial / an operational service requiring ownership, support, security and assurance.
AI service ownership Economics and enterprise adoption · ● Essential Accountability for service outcomes, lifecycle, controls, operations, costs and continuous improvement.
Agent sprawl Economics and enterprise adoption · ◇ Practitioner Uncontrolled proliferation of agents across an organisation, creating unclear ownership, duplicate functions, access risks and unnecessary cost. An inventory and lifecycle governance help manage it.
AI sovereignty Economics and enterprise adoption · ◇ Practitioner An organisation’s or jurisdiction’s ability to control its AI infrastructure, data, models and operations. The term has varying scope and goes beyond data residency alone.
AI inference economics Economics and enterprise adoption · ◇ Practitioner AI token economics / tokenomics How the resources used to run AI translate into cost and business value. Consider model choice, input and output tokens, reasoning, caching, retries, tool calls and supporting infrastructure. Optimise cost per useful outcome while preserving required quality; token price alone is insufficient.
Cost allocation / showback / chargeback Economics and enterprise adoption · ◇ Practitioner Allocation assigns costs to teams, services or cost centres. Showback reports attributed costs to stakeholders; chargeback posts charges to their budgets or accounts. For shared AI services, define how usage and shared platform costs are apportioned.
No concepts match. Try another term or category.
Concepts that are often confused RAG / fine-tuning Retrieve knowledge at runtime / change model weights through training. MCP / A2A Connect AI applications to tools and context / enable communication between agents. Model / agent harness The model generates responses and can propose actions; the harness coordinates model calls, tools, context and execution controls. They work together but are not the same thing. Workflow / agent Predefined paths / model-directed next steps within boundaries. Context / memory Information available now / information retained for later use. Responsible AI / guardrails Lifecycle principles and accountability / specific controls. Open weights / open source Available model parameters / broader openness and licensing requirements. Grounded / correct Supported by sources / factually accurate. Sources can be wrong. Hallucination / sycophancy Hallucination invents or misstates information; sycophancy agrees with or flatters the user at the expense of accuracy. A response can do both. Prompt injection / tool poisoning / data poisoning Prompt injection attempts to redirect AI through untrusted instructions. Tool poisoning is a form of indirect prompt injection through tool descriptions or metadata. Data poisoning manipulates training data or knowledge sources to influence behaviour or answers; it need not contain instructions.
Emerging concepts and sources Context engineering, agent harnesses, agentic RAG, MCP, A2A, computer use, test-time compute and AgentOps are important evolving areas. This is a dated reference, not a live news feed. Product and protocol versions require separate checks. AgentOps and agentic AI have varying industry usage; AGI and ASI definitions remain contested.