Understand the language behind the decision.

Stored information an agent can reuse across steps or sessions.

Agent memory may include conversation history, summaries, retrieved facts, user preferences or records of prior actions.

A system that plans and takes actions using models, tools and state.

An AI agent iteratively interprets a goal, chooses actions, uses tools, observes results and continues until it completes or stops a task.

Making AI behaviour consistent with intended goals and constraints.

Alignment research and practice seek to ensure models and agents pursue specified objectives without harmful or unintended behaviour.

A controlled record of models, AI systems, owners, uses and risk status.

A model inventory captures each system's purpose, components, owner, provider, data, materiality, validation and lifecycle state.

Systematic differences in model behaviour that may create unfair outcomes.

AI bias can arise from data, labels, objectives, model design, deployment context or feedback loops and affect groups differently.

Using models to support eligibility, affordability, pricing or credit limits.

AI credit decisioning applies predictive or generative systems within origination, underwriting, account management or collections.

Using AI to detect, investigate or explain potential financial crime.

AML AI may prioritise alerts, link entities, analyse adverse media, summarise cases or support suspicious-activity investigations.

Coordinating models, prompts, tools, data, memory and control flow.

Orchestration software manages the sequence, branching, retries, state and policies that turn AI components into a working application.

Assessing whether model outcomes treat relevant groups appropriately.

Algorithmic fairness uses legal, ethical and statistical criteria to evaluate differences in errors, decisions or benefits across populations.

Computer systems performing tasks that normally require human intelligence.

Artificial intelligence is the broad field of building systems that perceive, predict, generate, reason, recommend or act toward defined goals.

A mechanism that weights which parts of an input matter for each output.

Attention lets a model dynamically combine information from different token positions when creating contextual representations.

A durable record of inputs, evidence, decisions, actions and versions.

An AI audit trail records what the system received, retrieved, generated and changed, together with model and configuration identifiers.

The proportion of time a service can successfully respond.

Availability measures whether the complete AI service and its dependencies are accessible and functioning when required.

A standardised test used to compare systems consistently.

A benchmark combines a defined task or dataset, instructions, scoring method and test conditions so multiple models or systems can be compared on the same basis.

Test material leaking into model training or optimisation.

Contamination occurs when a model or development team has seen benchmark examples or close variants before the final evaluation.

Loss of previously learned capability when a model is trained on new data.

Catastrophic forgetting occurs when adaptation shifts model parameters in ways that degrade performance on earlier tasks or knowledge.

Intermediate reasoning steps produced or used while solving a task.

Chain-of-thought methods encourage a model to decompose a problem into steps before providing its final answer.

Splitting documents into smaller units for indexing and retrieval.

Chunking divides source material by length, structure or meaning before creating embeddings or passing context to a model.

Whether a cited source genuinely supports the associated claim.

Citation correctness checks that references exist, can be opened and provide evidence for the precise statement they are attached to.

Risk that a firm's behaviour produces poor customer or market outcomes.

Conduct risk covers product design, advice, communications, sales, servicing, complaints and market behaviour that may cause harm.

A range expressing uncertainty around an estimated score.

A confidence interval estimates the range within which the underlying performance measure is likely to fall given sampling uncertainty.

Training and steering models using an explicit set of written principles.

Constitutional AI asks models to critique and revise outputs against a defined constitution and can use those judgements as training feedback.

The UK expectation that firms deliver good outcomes for retail customers.

The FCA Consumer Duty requires firms to act to deliver good outcomes across products, value, understanding and customer support.

The maximum information a model can process in one interaction.

The context window includes instructions, retrieved documents, conversation history, tool results and the generated response.