Plain-English definitions for the AI, data and engineering terms that appear in strategy conversations, product roadmaps and production systems.
A controlled experiment in which users or cases are randomly assigned to two or more versions of something, such as a web page, message or process, and the outcomes are compared. Randomisation lets differences be attributed to the change rather than to other factors.
The number of a model’s parameters that are used to process each token. In a standard dense model every parameter is active. In a mixture-of-experts model only a fraction is, so the active count can be far smaller than the total.
The mechanisms that let an AI agent retain and recall information beyond a single model call, such as the current task’s history, facts learned about a user or records of previous outcomes. Memory is usually implemented in external storage rather than inside the model.
An open protocol that lets AI agents built on different platforms discover each other’s capabilities, exchange messages and coordinate tasks. It complements the Model Context Protocol, which connects an agent to tools and data rather than to other agents.
AI designed to pursue a goal through a sequence of actions rather than produce a single response. Agentic systems can plan, use tools and adapt to results, which makes permissions, monitoring and clear stopping conditions essential.
A business process in which one or more AI agents carry out steps such as gathering information, deciding what to do next and calling tools, within boundaries defined by ordinary software. The workflow, not the model, defines what the agent may do and when a person must step in.
A software component that uses an AI model to decide what action to take within a defined environment. A useful agent needs bounded objectives, approved tools, access controls and a reliable way to recover when an action fails.
Advisory work that helps an organisation decide where and how to use artificial intelligence. It typically covers opportunity assessment, use-case prioritisation, data readiness, risk and the operating model. Some firms stop at recommendations; others also build and run the systems they recommend.
An AI assistant embedded in a tool or workflow that suggests, drafts or prepares work for a person who remains in control of the outcome. The person reviews and decides; the copilot reduces effort on the steps around that decision.
A structured method for measuring whether an AI system performs acceptably on representative tasks. Evals should test accuracy, safety, consistency, cost and operational outcomes before deployment and as the system changes.
The roles, policies and controls used to decide how AI may be selected, built and operated. Effective governance assigns accountability, classifies risk and requires evidence proportionate to the consequences of each use case.
Technical and procedural controls that constrain an AI system’s inputs, outputs or actions. Guardrails can reduce known risks, but they do not replace evaluation, secure system design or accountable human oversight.
An output that appears plausible but is unsupported, incorrect or invented. Hallucination risk is managed through grounding, constrained tasks, source evidence, evaluation and review rather than prompting alone.
A maintained register of the AI systems, models and AI-enabled services an organisation builds, buys or uses, with each entry recording purpose, owner, data used, risk level and status. It is the foundation for any AI governance programme.
The ability to inspect how an AI system behaves in production using traces, quality measures, costs and user feedback. It helps teams find failures, investigate unexpected outputs and understand whether performance is changing.
Structured adversarial testing in which people deliberately try to make an AI system fail, misbehave or leak information. It looks for harmful outputs, security weaknesses, policy breaches and unsafe tool use before real users or attackers find them.
A defined set of instructions that turns an input into an output. Algorithms can be simple business rules or complex statistical processes; what matters operationally is whether their purpose, inputs and limitations are understood.
A systematic tendency for an automated system to produce unfair or distorted outcomes. Bias can enter through historical data, labels, design choices or how a model is deployed, so testing must cover affected groups and real operating conditions.
The discipline of transforming raw data into clean, tested, documented models that analysts and business users can rely on. It applies software engineering practices, such as version control, testing and code review, to analytical data.
Identifying data points, events or patterns that differ significantly from what is normally expected. It is used to spot fraud, equipment faults, data errors, security incidents and unusual operational behaviour.
Processing personal data so that individuals are no longer identifiable by any means reasonably likely to be used, taking account of other data that might be available. Effectively anonymised data falls outside UK GDPR, but the threshold is high.
A defined interface that lets one piece of software request data or actions from another, according to agreed rules about inputs, outputs and permissions. APIs are how modern applications, data platforms and AI tools connect.
A broad term for computer systems that perform tasks associated with human intelligence, such as interpreting language, recognising patterns or supporting decisions. In production, AI is one component of a wider system that also needs reliable data, controls, interfaces and accountable people.
A chronological, tamper-evident record of who did what, when and with which data or system, including actions taken by automated processes and AI components. It lets an organisation reconstruct how a record changed or a decision was reached.
Decisions made by technological means without meaningful human involvement. Data protection law places specific safeguards on solely automated decisions that have legal or similarly significant effects on individuals, such as credit, employment or benefit decisions.
Data whose volume, speed or variety exceeds what conventional tools can handle efficiently. The phrase does not imply value: large datasets only become useful when they are trustworthy, accessible and connected to a clear decision or workflow.
A technique that detects and records changes made to a database, such as inserts, updates and deletes, and passes them to other systems as they happen. It keeps analytical platforms and downstream services in step with operational systems without repeated full extracts.
A software interface that lets people interact with a service through written or spoken conversation. Modern chatbots often use language models, but useful deployment still requires bounded knowledge, clear escalation routes and access controls.
Splitting documents into smaller passages before they are embedded and indexed for retrieval. How documents are chunked strongly affects whether a retrieval system finds the right evidence.
Automated practices for building, testing and releasing software. Continuous integration checks every change against automated tests; continuous delivery keeps software in a releasable state and automates deployment.
Evidence that changes a commercial decision, rather than reporting that only describes what happened. Good insight connects trustworthy data to a specific action involving customers, pricing, operations, investment or growth.
Methods that enable computers to interpret images and video. Typical applications include inspection, classification and detection, but performance depends heavily on representative image data and how uncertain cases are reviewed.
The amount of information a language model can consider during one interaction. A larger context window can accommodate more material, but it does not guarantee that every detail will be recalled or used correctly.
A visual display of key measures and trends, usually refreshed automatically, that lets people monitor performance and spot issues at a glance. A good dashboard is designed around specific decisions rather than all available data.
Recorded facts, measurements, observations or representations that can be processed and analysed. Data becomes operationally valuable when its meaning, ownership, quality and permitted uses are clear.
The overall design of how an organisation’s data is collected, stored, integrated, governed and made available, including the platforms, data models, flows and standards involved. It sets the structure within which individual pipelines and products are built.
A searchable inventory of an organisation’s data assets, describing what each dataset contains, where it lives, who owns it and how it may be used. A catalogue turns scattered tables, files and reports into something people can find, assess and request access to.
The practice of labelling data according to its sensitivity, regulatory status or business criticality, so that handling rules can be applied consistently. Typical labels distinguish public, internal, confidential and restricted information, along with categories such as personal or special category data.
An agreed, machine-checkable specification of the data a producer will supply to its consumers, covering schema, meaning, quality thresholds and how changes are communicated. It makes the interface between systems explicit instead of relying on assumptions.
A change in the statistical profile of the data a model receives compared with the data it was trained on, such as a new customer mix or a changed input format. Data drift can degrade model performance even when the model itself is unchanged.
The discipline of building and operating the systems that collect, move, transform and store data so it can be used reliably for analytics, applications and AI. It covers pipelines, platforms, data models, quality controls and operational monitoring.
An architecture approach that uses metadata, integration services and automation to provide consistent access to data spread across many systems and locations. It emphasises connecting data where it lives rather than centralising everything.
The decisions, roles and controls that determine who can use which data and for what purpose. Effective governance makes ownership, quality, access, retention and accountability part of everyday operations rather than a one-off compliance exercise.
Adding the correct answer or category to raw examples, such as tagging an email as a complaint or marking defects in an image, so they can be used to train or evaluate a model. Labelling can be done by experts, trained annotators, existing process outcomes or AI-assisted tools.
A storage repository that holds large volumes of raw data in its original formats, structured and unstructured, until it is needed. Data lakes are typically built on low-cost object storage.
A data platform architecture that combines the low-cost, flexible storage of a data lake with warehouse-style features such as transactions, schema enforcement and fast analytical queries. It aims to serve reporting, data science and AI from one platform.
A record of where data originated, how it changed and which systems or reports use it. Lineage helps teams assess trust, investigate errors and understand the impact of changing a source or transformation.
An organisation’s ability to collect, govern, connect and use data reliably. Maturity is demonstrated by repeatable decisions and operating practices, not by the number of tools purchased or dashboards produced.
An organisational and architectural approach in which business domains own and publish their data as products, supported by a shared self-service platform and common governance standards. It decentralises data ownership while keeping interoperability.
The principle that an organisation should collect and process only the personal data that is adequate, relevant and limited to what is necessary for a stated purpose. It is one of the core data protection principles in UK GDPR.
Designing how data is structured, related and defined so that it accurately represents the business and supports the questions people need to answer. Models range from conceptual diagrams of entities to the physical tables in a database.
The ability to monitor the health of data as it moves through pipelines, detecting problems such as late arrivals, unexpected volume changes, schema changes and distribution shifts. It applies the monitoring discipline of software operations to data itself.
An automated sequence of steps that moves data from sources to destinations and transforms it along the way, for example from an operational system into a warehouse and on to a report or model. Pipelines can run on a schedule or continuously.
A reusable, managed data asset designed to meet a defined user need. Like a software product, it should have an owner, documented meaning, quality expectations and a process for change.
A structured assessment of how a planned processing activity could affect individuals’ privacy and what measures will reduce those risks. Under UK GDPR, a DPIA is required where processing is likely to result in a high risk to individuals.
The degree to which data is fit for a specific use, usually judged across dimensions such as accuracy, completeness, consistency, timeliness, validity and uniqueness. Quality is always relative to a purpose: data that is adequate for a monthly trend may be unfit for an automated decision.
The rules that define how long each type of data is kept, where it is stored during that period and how it is deleted or archived afterwards. Retention periods are set by legal obligations, contractual commitments and genuine business need.
The practice of using statistics, computation and domain knowledge to learn from data. Its outputs may include analysis, experiments or predictive models, but they create value only when connected to a real decision and measured in use.
A person accountable for the day-to-day meaning, quality and appropriate use of a defined set of data. Stewards sit close to the business process that creates the data and turn governance policy into practical decisions.
A central database designed for analysis, holding cleaned and integrated data from multiple source systems in structures optimised for querying. It is the traditional foundation for reporting, dashboards and analytical models.
An engineering approach that improves system performance by treating data quality, structure and feedback as first-class design concerns. It shifts attention from repeatedly changing models to building dependable data pipelines, labels, evaluation sets and monitoring.
A structured collection of related data used for analysis, reporting or model development. A dataset should have a defined source, scope, owner and quality standard so users know what conclusions it can support.
A branch of machine learning based on neural networks with many processing layers. It powers advances in language, vision and speech, but often requires substantial data and computing resources and can be difficult to interpret.
Synthetic or manipulated media designed to make a person appear to say or do something they did not. Deepfakes create fraud, misinformation and identity risks that require verification processes as well as detection technology.
A digital representation of a physical asset, process or environment that is updated using real-world data. Organisations use digital twins to monitor conditions, test scenarios and improve operations without disrupting the real system.
A data modelling technique for analytics that organises data into fact tables, which record measurable events such as sales or shipments, and dimension tables, which describe context such as customer, product, date and location. The typical result is a star schema.
Analytics, such as charts, metrics and insights, built directly into an application or workflow rather than presented in a separate reporting tool. It can serve internal users or be offered to customers as part of a digital product.
A numerical representation that captures selected meaning or similarity in data such as text, images or products. Embeddings support semantic search, recommendations and retrieval by placing related items near one another mathematically.
The transformation of readable information into a protected form that can only be accessed with the correct key. Encryption protects data in transit and at rest, but it must be paired with sound identity, access and key-management practices.
Identifying and linking records that refer to the same real-world entity, such as a person, company, product or property, even when names, formats and identifiers differ. It is also called record linkage or deduplication.
A data integration pattern that extracts data from source systems, transforms it into the required structure and loads it into a destination such as a warehouse. In the related ELT pattern, raw data is loaded first and transformed inside the destination platform.
The European Union’s regulation on artificial intelligence, Regulation (EU) 2024/1689, which sets obligations according to the risk an AI system poses. It prohibits certain practices, imposes substantial requirements on high-risk systems, adds transparency duties for some others and regulates general-purpose AI models.
A software design in which systems communicate by publishing and reacting to events, such as ‘order placed’ or ‘claim submitted’, rather than calling each other directly. It lets services respond to changes as they happen and evolve independently.
Methods that help people understand why an AI system produced a particular result. The required explanation depends on the audience and consequence, ranging from source evidence to model-level analysis and decision records.
Transforming raw data into input variables, called features, that help a model learn useful patterns. Examples include days since last purchase, rolling averages or ratios between related measures.
A system for defining, storing and serving model features consistently for both training and live predictions. It lets teams reuse features across models and ensures production uses the same logic as development.
Further training a pre-existing model on selected examples to change its behaviour for a domain or task. Fine-tuning can improve consistency, but it requires suitable data, evaluation and ongoing maintenance.
A delivery model in which senior software, data and AI engineers work directly inside a client’s operation and systems alongside the people who do the work. Rather than building from a static specification or stopping at a slide deck of recommendations, forward-deployed engineers scope the problem with operators, ship working software in short cycles and stay until the workflow changes in production.
A large model trained on broad data that can be adapted to many downstream tasks. Language and vision models are common examples; organisations typically configure, ground or fine-tune them rather than train one from scratch.
A machine-learning architecture in which one network generates examples and another evaluates whether they resemble real data. GANs can create realistic images and synthetic data, although newer generative methods are now used for many of the same tasks.
AI systems that create new text, images, audio, video or code from learned patterns. Business use requires more than content generation: outputs need grounding, evaluation, permissions and human review appropriate to their consequences.
A curated set of representative inputs with agreed correct or acceptable outputs, used as a fixed benchmark to evaluate an AI system. It is the reference against which changes to models, prompts or retrieval are tested.
The verified, correct answer for an example, used to train a model or judge its outputs. Ground truth might be a confirmed outcome, such as whether a transaction was fraudulent, or an expert’s agreed judgement.
Anchoring a model’s output in specific, approved sources of information, such as retrieved documents or database records, rather than in its general training. A grounded answer can be traced back to evidence the organisation trusts.
A system design in which people review, correct or approve parts of an automated process. Human involvement is most useful when responsibilities, escalation thresholds and the evidence available to reviewers are explicitly designed.
A retrieval approach that combines keyword search with semantic search and merges the results. It captures both exact matches, such as names and codes, and conceptual matches expressed in different words.
The process of using a trained model to produce a prediction or generated output from new input. In production, inference design affects response time, cost, privacy, reliability and where a model can run.
Defining and managing cloud and platform infrastructure, such as networks, databases, permissions and compute, through version-controlled configuration files rather than manual changes in a console. The same definitions can create identical environments repeatedly.
The use of AI to classify documents, extract information from them and validate the results, turning forms, letters, invoices and reports into structured data. Modern IDP combines optical character recognition, machine learning and language models.
The international standard for an artificial intelligence management system, published in 2023. It specifies requirements for establishing, operating and continually improving the policies, processes and controls an organisation uses to develop, provide or use AI responsibly.
A lightweight text format for representing and exchanging structured data. JSON is widely used by APIs, applications and AI tools because both people and software can read it consistently.
A structured representation of entities and the relationships between them. Knowledge graphs help organisations connect fragmented information, preserve context and support search, analysis and AI systems with explicit relationships.
A model trained on large collections of text to predict and generate language. LLMs can support search, drafting, extraction and workflow automation, but they can also produce plausible errors and must not be treated as an authoritative database.
Updating or replacing ageing software and infrastructure so that it is secure, maintainable and able to support new requirements, such as integration, analytics and AI. Approaches range from rehosting and refactoring to incremental replacement.
Using a language model to assess the outputs of another AI system against defined criteria, such as factual support, relevance, tone or policy compliance. It makes it practical to evaluate large volumes of generated text.
The practices used to deploy, evaluate, monitor and maintain applications built around language models. It extends conventional software operations to cover prompts, model versions, retrieval, quality, safety and usage cost.
A branch of AI in which systems learn patterns from examples rather than relying only on hand-written rules. A production ML system includes the data, model, software, monitoring and operating process needed to keep those patterns useful over time.
The processes and technology used to create and maintain a single, trusted version of core business entities such as customers, products, suppliers, assets or locations. MDM reconciles duplicate and conflicting records held across different systems.
A way of organising a data platform into layers of increasing quality, commonly called bronze for raw data, silver for cleaned and conformed data, and gold for business-ready tables. Each layer has clear rules about what it contains and who uses it.
Data that describes other data, such as its structure, meaning, source, owner, sensitivity, freshness and lineage. Metadata is what allows people and software to understand whether a dataset can be trusted and how it may be used.
A model architecture in which parts of the network are split into many smaller sub-networks, called experts, and a routing step sends each token to only a few of them. This gives the model a large total capacity while using only part of it for each token.
The engineering practices used to deploy, monitor and update machine-learning systems reliably. MLOps connects data, model development and software operations so changes can be tested, traced and reversed.
A short, structured document describing a model’s intended use, training data, evaluation results, known limitations and ethical considerations. It gives reviewers and users the context they need to decide whether the model is appropriate for a task.
An open protocol for connecting AI applications to tools and external sources of context through a standard interface. MCP can reduce bespoke integration work, while the application remains responsible for authentication, permissions and safe tool use.
A decline or change in model behaviour as real-world data and conditions move away from those used during development. Monitoring should detect drift early enough for teams to investigate, retrain, adjust or withdraw the model.
A central record of trained model versions with their training data, parameters, evaluation results, approval status and deployment history. It is the system of record for which model is running where, and why.
The framework for identifying, validating, monitoring and controlling the risk that a model is wrong, misused or no longer fit for purpose. It covers statistical models, machine-learning systems and, increasingly, generative AI components.
The infrastructure and software that make a trained model available to applications, typically through an API or batch job, with the performance, scaling and security production use requires. It is where a model becomes part of an operational system.
A system in which multiple autonomous software agents coordinate or compete to complete tasks. Multi-agent designs can separate specialist responsibilities, but they also increase complexity around state, permissions, failure handling and evaluation.
AI that can work across more than one type of information, such as text, images, audio or video. Multimodal systems can support richer workflows but require evaluation across each input type and their combinations.
Methods that enable computers to analyse, interpret and generate human language. NLP is used for tasks such as classification, extraction, search, summarisation and conversational interfaces.
A machine-learning model made of connected processing units that learn relationships from data. Neural networks are effective for complex patterns but usually need careful evaluation because their internal reasoning is not straightforward to inspect.
A voluntary framework published by the US National Institute of Standards and Technology to help organisations identify and manage risks from AI systems. It is organised around four functions: Govern, Map, Measure and Manage.
Software whose source code is available under a licence that permits specified forms of use, inspection and modification. Open source can improve flexibility and transparency, but teams still need to assess licence obligations, maintenance, security and operational ownership.
A model whose trained weights are published for anyone to download and run. The licence attached to the weights decides what you may do with them, and the training data and code are often not released, so an open-weights model is not necessarily open source.
When a model learns the noise and specific quirks of its training data rather than patterns that generalise, so it performs well in development and poorly on new data. It is one of the most common reasons models disappoint after deployment.
Information that can identify a living individual, either on its own or when combined with other data. In the UK and EU the legal term is personal data, which is broader than many PII definitions and includes identifiers such as online IDs and location data.
Analytics that recommends what action to take, typically by combining predictions with optimisation, simulation or business rules. It goes beyond describing what happened or forecasting what will happen.
A complete operational system in which AI performs a defined role alongside data pipelines, software, controls, monitoring and people. Production readiness means it can deliver dependable value under real conditions, not merely demonstrate technical potential.
The design and testing of instructions and context supplied to a generative model. Good prompts improve behaviour, but dependable applications also need evaluation, structured inputs, source grounding and software controls.
An attack in which text supplied to a language model, directly by a user or indirectly through a document, web page or email, contains instructions that override the application’s intended behaviour. It is one of the most significant security risks for applications built on language models.
Replacing direct identifiers such as names or account numbers with artificial keys, while keeping the information needed to re-identify individuals separately and under additional controls. Under UK GDPR, pseudonymised data is still personal data.
A technique that reduces the numerical precision of a model so it uses less memory and computing power. Quantisation can make models cheaper and faster to run, but the compressed version still needs evaluation against the original quality standard.
Analysis of data within seconds or minutes of it being generated, so that people or systems can respond while events are still unfolding. It relies on streaming or near-real-time data pipelines.
A language model trained to work through intermediate steps before producing an answer, typically spending more computation at inference time on harder problems. Reasoning models tend to perform better on multi-step analysis, planning and coding tasks.
A machine-learning approach in which an agent learns actions through rewards and penalties received from an environment. It is useful for sequential decisions, but the reward design must reflect the real objective or the system may optimise the wrong behaviour.
A second retrieval step in which a more precise model reorders an initial set of search results by relevance to the query. It improves the quality of the few passages that are actually passed to a language model.
The practice of designing and operating AI in ways that are lawful, fair, secure and accountable. It turns broad principles into decisions about data, testing, transparency, oversight and what the system must not do.
A technique that retrieves relevant information from an approved source before asking a generative model to respond. RAG can improve relevance and provide source evidence, but retrieval quality and access permissions must also be evaluated.
Moving modelled data from a warehouse or lakehouse back into operational tools, such as a CRM, support platform or marketing system, so that people and processes can act on it. It closes the loop between analysis and action.
Software that automates repetitive, rule-based tasks by mimicking how a person uses applications, such as clicking, copying and entering data. It is often used where systems lack suitable APIs.
An access model in which permissions are granted to roles, such as claims handler or finance analyst, and people receive access by being assigned to those roles. It makes access easier to review, grant and revoke than managing permissions person by person.
An approach that lets business users explore data and build their own reports without relying on a central data team for every question. It depends on governed, well-documented data and tools suited to non-specialists.
A shared business representation that gives data consistent definitions, relationships and measures across tools. It helps prevent teams from calculating the same metric differently in dashboards, analysis and AI applications.
Search that finds results based on meaning rather than exact keyword matches, usually by comparing embeddings of the query and the content. It can find a relevant passage even when it uses different words from the question.
The use of AI tools and services by staff without the organisation’s knowledge, approval or controls. Common examples include pasting work documents into public chat assistants or enabling AI features in software without review.
An agreed, authoritative location and definition for each important piece of data or metric, so that everyone uses the same figure for the same question. It is achieved through governance and design, not by putting all data in one system.
A language model designed with fewer parameters and lower computing requirements than the largest general-purpose models. SLMs can offer lower cost, faster responses and greater deployment control for focused tasks.
Processing data continuously as events arrive, rather than in periodic batches. It supports use cases that need a response within seconds or minutes, such as fraud checks, operational alerts and live tracking.
Constraining a language model to return data in a defined format, such as JSON that matches a schema, rather than free text. It lets downstream software parse, validate and act on model outputs reliably.
A machine-learning approach in which a model learns from examples that are labelled with the correct answer, such as past claims marked as fraudulent or not. The trained model then predicts that label for new cases.
Artificially generated data designed to reproduce selected properties of real data. It can support testing, simulation and model development where real examples are scarce or sensitive, but it must be checked for realism, bias and privacy leakage.
Using a language model to translate a question in plain English into a database query, run it and return the result. It lets people who do not write SQL ask questions of structured data directly.
The unit of text a language model reads and writes, often a word or a fragment of a word. Context window limits, model pricing and response speed are all usually measured in tokens.
The ability for an AI model to request that an application execute an approved function, such as searching a system or creating a record. The surrounding software must validate arguments, enforce permissions and decide whether consequential actions require confirmation.
The examples a machine-learning model learns from during development. Its coverage, quality, labelling and provenance largely determine what the model can do and where it will fail.
Information that does not sit neatly in rows and columns, such as documents, emails, images, audio and chat logs. Most organisational knowledge is unstructured, which is why extraction, classification and retrieval are often the first production AI problems.
A machine-learning approach that finds structure in data without labelled answers. It is used for clustering, anomaly detection and exploration, but the resulting groups still need a human interpretation of what they mean operationally.
A database designed to store embeddings and find items with similar numerical representations. It is commonly used for semantic search and RAG, where metadata filtering and source permissions remain as important as similarity.
The use of software to move work through a defined process with less manual effort: routing tasks, collecting information, applying rules, requesting approvals and recording what happened. It ranges from simple rule-based steps to processes where AI classifies or drafts and people decide.
The coordination of tasks, systems and decision points within a repeatable process. In AI-enabled workflows, orchestration controls sequencing, state, retries, approvals and what happens when a model or external service fails.
A machine-learning method based on an ensemble of boosted decision trees. It is widely used for classification, forecasting and risk scoring with structured business data because it can perform strongly while remaining comparatively practical to inspect and operate.
A human-readable format commonly used for configuration, pipelines and infrastructure definitions. In production systems, YAML often encodes how services, data jobs and AI workflows are deployed, so small mistakes can change behaviour as much as application code.
A security model that assumes no user, device or service is trusted by default, even inside the organisation’s network. Access is granted through identity, least privilege and continuous verification, which is especially important when AI systems call tools or touch sensitive data.
The ability of a model to perform a task it was not explicitly trained on, using general knowledge and the instruction given at the time. Zero-shot behaviour is useful for speed, but it still needs evaluation against the organisation’s actual examples and success criteria.