AI Glossary

Testable conditions that define when a feature or a model's output is complete and correct, so that 'done' can be checked rather than argued.

Acceptance criteria are the specific, observable conditions a piece of work must meet to be accepted. In software they are usually attached to a user story and written so that anyone can check them, often in a Given-When-Then form. For AI systems they translate requirements into measurable checks: the output follows a schema, answers match a set of reference ('golden') answers, sources are cited, safety rules are respected and quality stays above an agreed threshold. They underpin evaluation harnesses, regression suites and human-in-the-loop review, linking the specification to measurable quality.

An AI agent's live plan: the ordered next steps, open subtasks, blocked items and checkpoints it works through towards a goal.

An agent agenda is the working plan an AI agent keeps while it carries out a task. It lists the next actions, the subtasks still open, the items that are blocked waiting for data or approval and the steps already completed, and it is updated as tool results come in. The agenda comes from breaking the goal into steps and revising them as the agent learns more, an approach studied in research on planning before acting and on interleaving reasoning with actions. Showing the agenda in the user interface and in logs makes an agent's run inspectable and lets teams check it against the requirements and acceptance criteria.

The component or prompt pattern in an AI agent that breaks a goal into a sequence of steps before any tool is run.

An agent planner is the part of an AI agent that decides how to reach a goal: it breaks the request into steps, chooses which tools or sub-agents each step needs and proposes an order. Plans are usually conditioned on policies, specifications and retrieved context. The plan is then carried out by an executor, which calls the tools and reports results, and the planner may revise the plan as results arrive. Separating planning from execution makes agents easier to test and control: you can check a plan against acceptance criteria, permissions and budgets before anything with side effects happens.

A packaged capability for an AI agent: the instructions, allowed tools, examples and input and output formats it needs to perform one coherent task.

An agent skill bundles everything a model needs to perform a defined capability reliably and repeatedly, for example filing an expense or running a compliance check. It typically contains instructions written for the agent, the tools it may use, examples, scripts or reference files, and the expected inputs and outputs. Anthropic formalised the idea as Agent Skills: folders with a SKILL.md file that an agent discovers by name and description and loads only when a task calls for it. Skills are the unit of reuse between specifications, runbooks and orchestration layers such as Synapse.

AI systems that pursue a goal with some autonomy: they plan steps, use tools and act on systems, rather than only answering a prompt.

Agentic AI refers to systems built around a language model that does more than generate text: it decides what to do next, calls tools or APIs, reads the results and continues until a goal is met or it needs a human. The degree of autonomy varies, from fixed workflows where the model fills in steps to agents that choose their own path. The value lies in completing multi-step work across real systems; the cost is that errors can compound, so agentic systems need clear limits, monitoring and human checkpoints. Thinkia Synapse is an agentic platform for enterprise AI.

An AI architecture that uses autonomous agents to iteratively search, evaluate, and synthesise information across multiple data sources until a complex question is fully answered.

Agentic RAG moves beyond the single-pass document fetching of standard retrieval-augmented generation. Instead of just pulling data once based on a user's prompt, it empowers an AI agent to break down a complex task, formulate its own search queries, and evaluate the retrieved context. If the initial information is incomplete, the agent autonomously adjusts its approach and searches again across different enterprise systems. This iterative process allows the AI to handle multi-step reasoning and connect fragmented data points.

Software that pursues a goal by deciding its own next steps, using tools and data, and acting on systems with limited human input.

An AI agent is a piece of software that receives a goal, works out the steps to reach it and carries them out. Today most agents use a large language model as the reasoning core: the model reads the situation, decides which tool to call (a search, an API, a database query, a calculation), looks at the result and decides what to do next. The loop repeats until the goal is met, a limit is reached or a person needs to step in. What makes it an agent rather than a script is that the path is not fixed in advance: the model chooses it at run time.

A machine-readable inventory of everything an AI system is built from: models, datasets, dependencies, licences and their lineage.

An AI bill of materials (AI BOM) is an inventory of the components of an AI system: base models and their versions, adapters and fine-tunes, training and evaluation datasets, software libraries, containers, and the licences and provenance of each. It extends the idea of the software bill of materials (SBOM) to the parts that are specific to AI. With it, security, legal and risk teams can see what a system depends on, assess exposure when a model or library has a problem, and answer questions from auditors and regulators. Standards such as OWASP CycloneDX (ML-BOM) and SPDX 3.0 (AI and dataset profiles) define formats for it.

The policies, processes and controls an organisation uses to make sure its AI is safe, lawful, ethical and worth what it costs.

AI governance is the set of policies, roles, processes and technical controls that an organisation puts in place to use AI responsibly. It decides who may build or buy AI, which uses are allowed, how risks are assessed, how systems are tested and monitored, and who is accountable when something goes wrong. It covers security, privacy, regulatory compliance, cost control, audit trails and the measurement of return. Good governance does not slow AI down; it gives teams clear rules so they can move faster with less risk.

Systems, products and processes designed from the start with AI as a core part, rather than with AI added afterwards.

AI-native describes systems, products, processes or organisations designed with AI as a first-class component from the start, rather than bolting AI onto something built without it. In an AI-native product, the main way the user gets value depends on AI; take the AI away and the product no longer makes sense. In an AI-native process, the work is redesigned around what AI can do, with people focusing on judgement, exceptions and oversight. The term has no single formal definition and is used in software, telecoms and consulting with slightly different meanings. Thinkia specialises in AI-native consulting and platform development.

A defined interface that lets one software system request data or actions from another; it is how applications call AI models and connect them to business systems.

An application programming interface (API) is a contract between two pieces of software: it states which requests one system can send, in what format, and what it will get back. Most modern APIs work over the web, with requests and responses in JSON, and are often described in a standard format such as the OpenAPI Specification. In AI, APIs are how applications call large language models (from providers such as OpenAI or Anthropic), generate embeddings and use other AI services. They are also how AI agents reach the organisation's own systems, such as a CRM, an ERP or a ticketing tool, to read data and take actions.

A formal, machine-readable description of an API's endpoints, data schemas, authentication and errors, such as an OpenAPI document.

An API specification is a formal description of how an API works: which endpoints or operations it offers, what each request must contain, what each response returns, how authentication works and which errors can occur. Written in a standard, machine-readable format such as OpenAPI, often with JSON Schema for the data, it lets clients and servers agree without ambiguity. The same document drives documentation, mock servers, generated client code and contract tests. In AI systems it also matters for tool calling, because the model must produce calls that match a strict schema.

Having software carry out tasks without a person doing each step, increasingly with AI to handle decisions and unstructured inputs.

Automation is the execution of tasks or processes by software or machines without a person doing each step. Traditional automation follows fixed rules: if this, then that. With AI, automation can read documents, understand requests, classify cases and make bounded decisions, so it reaches work that used to need human judgement. Hyperautomation extends this from single tasks to whole business processes, combining workflows, RPA, AI models and people.

Systems that sense their environment, decide and act on their own in real time, without a person supervising every step.

An autonomous system perceives its environment, decides what to do and acts, without a person approving each decision. The idea covers physical systems, such as self-driving vehicles, drones and warehouse robots, and software systems, such as AI agents that handle cases or adjust operations on their own. Autonomy is a matter of degree: most real systems act alone within defined limits and hand over to a person when a situation falls outside them. The more a system is allowed to do alone, the more it matters to define its boundaries, monitor it and keep a way for people to step in.

A Transformer language model from Google, pre-trained to read context in both directions; widely used for search, classification and embeddings.

BERT (Bidirectional Encoder Representations from Transformers) is a language model introduced by Google researchers in 2018. It is pre-trained on large amounts of unlabelled text by hiding some words and learning to predict them from the words on both sides, which gives it a strong grasp of context. A pre-trained BERT can then be fine-tuned with one extra output layer for tasks such as question answering, classification or entity recognition. It reads and represents text rather than generating it, and BERT-style encoders remain the basis of many semantic search and embedding systems.

Systematic distortion in the data, design or use of an AI system that produces unfair, discriminatory or inaccurate results for some people or cases.

Bias in AI is a systematic error, not a random one: the system performs worse, or decides differently, for certain groups or situations in a way that is consistent and often hard to see. It can come from training data that under-represents some people, from labels that reflect past human decisions, from design choices about what to optimise, or from how people use and trust the output. Bias cannot be removed entirely, but it can be measured, reduced and monitored. Responsible AI practice treats bias detection and mitigation as part of the life cycle, from data collection to operation.

Data sets so large, fast-moving or varied that traditional databases and tools cannot store or process them efficiently; a key input for AI.

Big data refers to data sets whose size, speed or variety exceeds what traditional databases and tools can handle efficiently. It includes transaction records, logs, sensor readings, clickstreams, documents, images and audio. Working with it needs distributed storage and processing, such as data lakes, lakehouses and cloud data platforms. For AI, big data is the raw material: it is used to train and evaluate models and to feed AI applications with current information. Volume alone does not create value; quality, governance and a clear use do.

Software that automates conversations or repetitive tasks, such as a chatbot that answers in text or a voicebot that speaks.

A bot is software that carries out conversational or repetitive tasks automatically. A chatbot talks in text on a website, an app or a messaging channel; a voicebot talks by voice, for example on a phone line. Older bots followed fixed rules and decision trees; modern bots usually rely on large language models (LLMs) to understand free text and answer in natural language, often with access to company knowledge and systems. In customer service and internal support they handle frequent requests and pass complex cases to a person.

A way of prompting or training a language model to write out intermediate reasoning steps before giving its final answer.

Chain-of-thought (CoT) is a technique in which a large language model produces intermediate reasoning steps before its answer, instead of answering directly. It was introduced as a prompting method: showing the model a few examples that include worked reasoning improved its performance on arithmetic, commonsense and symbolic reasoning tasks. Today many models are trained to reason step by step on their own. CoT is used in agent planning and tool selection, and the written steps help people review how an answer was reached, although they are not a guaranteed record of the model's internal process.

A conversational assistant that talks with users in text, driven either by predefined rules or by a large language model.

A chatbot is software that holds a conversation with users through text, and sometimes voice. Rule-based chatbots recognise an intent from the user's message and reply with predefined answers or flows. LLM-based chatbots use a large language model to understand open-ended questions and generate natural replies, often grounded in company documents through retrieval. They are used in customer service, sales and internal tools, wherever people need answers or simple tasks done without waiting for a person.

AI capabilities offered as managed cloud services, such as Azure OpenAI, Amazon Bedrock or Google Vertex AI, so companies use models without running the infrastructure.

Cloud AI refers to AI models, tools and computing power that a provider runs in its data centres and offers as a managed service. Examples include Azure OpenAI in Microsoft Azure, Amazon Bedrock in AWS and Google Vertex AI in Google Cloud. Companies call these services through APIs and pay for what they use, instead of buying GPUs and building and maintaining their own infrastructure. Most cloud AI platforms also bundle the surrounding pieces: model catalogues, fine-tuning, vector search, monitoring, security controls and regional data residency.

The field of AI that lets machines interpret images and video: recognise objects, read text, detect defects or follow movement.

Computer vision is the branch of AI that enables machines to extract meaning from images and video. Models trained on large sets of labelled images learn to classify what they see, locate and count objects, segment regions, read text and track movement over time. Deep learning, and convolutional networks and vision transformers in particular, made these tasks reliable enough for production. Common uses include quality control, document reading, medical imaging, retail analytics and safety monitoring.

A training approach in which a model learns to follow a written set of principles, a constitution, by critiquing and revising its own answers.

Constitutional AI is a method, introduced by Anthropic researchers in 2022, for training AI assistants to be helpful, honest and harmless using a short list of written principles, the constitution, as the main human input. In a first phase, the model critiques and revises its own answers against those principles and is fine-tuned on the revised versions. In a second phase, the model compares pairs of answers according to the principles, and those AI judgments train a preference model used as the reward signal, an approach called reinforcement learning from AI feedback (RLAIF). The result is a model that can engage with difficult requests and explain its objections, with far fewer human labels for harmful content. Companies apply similar ideas in guardrails and policy layers.

The practice of deciding what information a model sees at each step (instructions, retrieved documents, tool results, memory) and fitting it into the context window.

Context engineering is the discipline of designing everything that goes into a model's context window, not only the prompt. It covers which documents are retrieved, which tool outputs are kept, how earlier conversation is summarised, what long-term memory is recalled and how all of it is ordered within a limited token budget. The idea is that a model can only reason over what it is given, so the quality of an answer depends as much on the context as on the model. It is the practical backbone of reliable RAG systems and agents, especially in long tasks where context accumulates quickly.

The maximum amount of text, measured in tokens, that a language model can take into account in a single request, including its own reply.

The context window is the maximum amount of input and output, measured in tokens, that a large language model can process in one request. Everything the model sees counts against it: system instructions, conversation history, retrieved documents, tool results and the answer it writes. Larger windows allow longer documents and conversations, but they raise cost and latency, and models do not always use information in the middle of a long input as well as information at the start or end. Choosing what goes into the window is therefore as important as its size.

AI that holds natural, context-aware conversations with people by text or voice, and can act on their requests.

Conversational AI is the set of technologies that let a machine hold a conversation with a person in natural language, by text or by voice. It combines language understanding, dialogue management that keeps track of context, generation of the reply and, for voice, speech recognition and speech synthesis. Large language models have made these systems far more flexible than rule-based chatbots, and connecting them to company systems lets them not only answer but also act, for example by checking an order or booking an appointment. It powers virtual assistants, contact centre automation and digital humans.

An AI assistant built into an application or workflow that helps a person with tasks such as writing, coding or analysis, while the person stays in control.

A copilot is an AI assistant embedded in the tools people already use, such as a code editor, an office suite or a business application. It suggests text, code or actions, answers questions about the user's content and automates small steps, but the person decides what to accept. Well-known examples include GitHub Copilot for software development and Microsoft Copilot in Office applications. The term has become generic for in-context AI assistance, as opposed to agents that carry out work on their own.

A collection of documents or texts used to train a model or to serve as the knowledge base an AI system retrieves from, as in RAG.

A corpus is a structured collection of documents gathered for a purpose. In AI it plays two roles: as training data, the text a language model learns from, and as a knowledge base, the documents a retrieval-augmented generation (RAG) system searches to ground its answers. In an enterprise, the corpus is usually the set of policies, manuals, contracts, tickets or product documents an assistant is allowed to use. Its quality, coverage, freshness and structure directly affect how well the system retrieves and how accurate its answers are.

Customer relationship management: the strategy and the software used to manage customer data and interactions; AI adds prediction, automation and assistants.

CRM stands for customer relationship management. The term covers both a business strategy for managing relationships with customers and prospects across their lifecycle, and the software that supports it by storing contacts, accounts, opportunities, cases and every interaction. A CRM gives sales, marketing and service teams a shared view of each customer. AI extends it with predictions (which leads to prioritise, which customers may leave), automation of routine tasks and assistants that summarise accounts or draft messages. The quality of the AI depends on the quality of the data in the CRM.

Customer experience is the sum of a customer's interactions with a company and how they perceive them; AI helps personalise and automate it.

Customer experience (CX) covers every interaction between a customer and a company, from the first advert to purchase, use, support and renewal, and the perception those interactions leave. It is shaped by products, channels, people and systems, not by a single department. AI changes CX in three main ways: it personalises recommendations and messages, automates support with chatbots and voicebots, and routes each request to the right channel or person. The goal is consistent, faster service at scale, without losing the human contact where it matters.

The ownership, quality rules, lineage and policies that decide who can use which data, and how, so it is fit for analytics and AI.

Data governance is the set of roles, decision rights, policies and controls that determine who owns which data, who may use it and for what, how it is classified, retained and protected, and how its quality is assured. It connects privacy, security, master data and, increasingly, model risk. Good governance is a prerequisite for compliant AI, reliable features and explainable results, because a model can only be as trustworthy as the data it learns from or retrieves. It is an ongoing function, not a one-off project.

A central repository that stores structured and unstructured data in its raw form, as a source for analytics, machine learning and AI.

A data lake is a centralised repository that stores large amounts of structured, semi-structured and unstructured data in its original form. Unlike a data warehouse, it does not require a schema to be defined before data is loaded; structure is applied when the data is read. This makes it a flexible source for analytics, machine learning and AI applications that need varied data, such as documents, logs, images or event streams. Without cataloguing, access control and quality rules, a data lake quickly becomes hard to use.

An automated sequence of steps that collects data from its sources, cleans and transforms it, and delivers it to the models, applications or analytics that use it.

A data pipeline is an automated workflow that ingests data from sources such as databases, files, applications or APIs, transforms it and moves it to a destination such as a data warehouse, a feature store, a vector index or an application. Each step is defined in code or configuration and runs on a schedule or whenever new data arrives. Pipelines replace manual extracts and spreadsheets with a repeatable process that can be monitored and audited. For AI they are critical: a model or RAG system is only as current and reliable as the data that reaches it.

A form of machine learning that uses neural networks with many layers to learn patterns directly from raw data such as text, images or sound.

Deep learning is a branch of machine learning based on artificial neural networks with many layers. Each layer transforms its input into a slightly more abstract representation, so the network learns useful features by itself instead of relying on features designed by hand. Training adjusts millions or billions of parameters using large datasets and the backpropagation algorithm. Deep learning underpins computer vision, speech recognition, language models and most of today's state-of-the-art AI.

A lifelike animated avatar, driven by conversational AI, that talks with people by voice and expression on behalf of a brand or service.

A digital human is a realistic animated character, usually in 3D, that holds a spoken conversation with a person. Behind the face sits conversational AI: speech recognition to hear the user, a language model and company knowledge to decide the answer, speech synthesis to say it and animation that syncs lips, gaze and expression. The aim is a more natural, human-like interaction than a text chatbot, for example at a kiosk, in an app or on a website. Thinkia offers a Digital Humans product built on this idea.

A live virtual model of a physical asset, facility or process, kept up to date with real-time data so teams can monitor, predict and test changes.

A digital twin is a virtual representation of something physical, such as a machine, a building, a production line or a supply chain, that is continuously fed with data from its real counterpart. Sensors, control systems and business applications send telemetry to the twin, so its state mirrors what is happening now. Teams use it to monitor operations, predict failures and simulate decisions before applying them in the real world. AI adds forecasting, anomaly detection and scenario planning on top of the twin.

Adapting a general-purpose AI model so it performs well in a specialised field, such as law, healthcare or insurance, with its own language and rules.

Domain adaptation is the work of making a general-purpose model perform well in a specialised domain, such as legal documents, medical records or insurance claims. General models know a lot, but they may miss the vocabulary, formats and rules of a particular field or company. The options range from giving the model domain context at query time with retrieval-augmented generation (RAG), to fine-tuning it on domain examples, to continuing its pre-training on large amounts of domain text. Which one fits depends on whether the gap is knowledge, behaviour or language.

A period when a system, machine or service is unavailable. Predictive AI helps reduce unplanned downtime by forecasting failures before they happen.

Downtime is the time during which a system, machine or service cannot do its job. It can be planned, such as scheduled maintenance, or unplanned, caused by failures, errors or incidents. Unplanned downtime is usually the costly kind, because it stops production or service without warning. Predictive AI uses sensor, log and operational data to spot the patterns that precede a failure, so teams can intervene before it happens. This is a core use case in manufacturing, energy, logistics and other critical systems, and the idea behind zero-downtime operations.

Running AI models on devices or local servers close to where data is produced, to cut latency, keep data local and work without constant connectivity.

Edge AI means running AI inference on or near the device that produces the data, such as a camera, a sensor, a phone, a vehicle or a server in a factory, instead of sending everything to a central cloud. It reduces latency, saves bandwidth, keeps sensitive data local and lets systems keep working when the connection is poor. Models are usually trained in the cloud and then compressed and deployed to the edge. Federated learning extends the idea to training, by learning from data that stays on the devices.

A list of numbers (a vector) that represents the meaning of a piece of text, an image or audio, so that similar meanings sit close together.

An embedding is a dense numerical vector produced by a model to represent the meaning of a word, sentence, document, image or sound. The model is trained so that items with similar meaning end up close to each other in that vector space, even when they share no words. Comparing vectors, usually with cosine similarity, lets software find related content, group it or detect duplicates. Embeddings are the basis of semantic search and of the retrieval step in RAG systems.

AI built and run for large organisations: governed, secure, scalable and integrated with existing systems such as ERP, CRM and data platforms.

Enterprise AI is AI designed, deployed and operated inside large organisations, as opposed to consumer tools used by individuals. It has to work with the company's own data and systems, such as ERP, CRM, document repositories and identity, and respect their permissions. It must meet requirements that consumer apps can ignore: security, privacy, regulatory compliance, auditability, availability and cost control. The hard part is rarely the model; it is integration, governance and getting the solution into daily operations. Thinkia specialises in enterprise AI strategy and implementation.

AI designed, built and used according to ethical principles such as fairness, transparency, accountability, privacy and human oversight.

Ethical AI is artificial intelligence designed, developed and used in line with principles such as fairness, transparency, accountability, privacy and respect for human rights. It asks not only whether a system works, but whether it treats people fairly, whether its decisions can be explained and challenged, and who answers when something goes wrong. International frameworks, such as the European Commission's ethics guidelines for trustworthy AI and the OECD AI Principles, set out these values, and regulations such as the EU AI Act turn part of them into legal duties. For companies, it is also a matter of trust with customers, employees and regulators.

The European Union regulation that sets rules for AI systems according to their level of risk, from banned uses to transparency duties.

The EU AI Act (Regulation (EU) 2024/1689) is the European Union's legal framework for artificial intelligence. It classifies AI systems by risk: practices with unacceptable risk are prohibited, high-risk systems must meet strict requirements, some systems carry transparency obligations and minimal-risk uses are largely left free. It also sets rules for general-purpose AI models. It applies to providers and deployers of AI systems placed on the EU market or used in the EU, wherever they are based, and its obligations apply in stages. Thinkia helps organisations work towards compliance.

The process of measuring how well an AI model or application performs: accuracy, relevance, safety, bias, cost and other criteria.

Evaluation is the systematic measurement of how well an AI model or AI application performs against the criteria that matter for its use: accuracy, relevance, groundedness, safety, bias, robustness, latency and cost. For language models it combines automatic metrics, test sets with known answers, judgements by other models and review by people. Public benchmarks compare models in general; your own evaluations tell you whether a system works for your task and data. Solid evaluation is needed before a model goes to production and again every time the model, prompt or data changes.

A software design in which systems communicate by publishing and reacting to events, such as "order placed", instead of calling each other directly.

In an event-driven architecture, a component that detects a change (an order placed, a payment failed, a sensor reading) publishes an event, and other components that care about it react on their own. Events travel through a broker or streaming platform, so the producer does not need to know who consumes them. This decouples systems, lets several consumers act on the same event and supports near-real-time processing at scale. For AI, it is how models and agents get triggered by what happens in the business, rather than waiting for a nightly batch or a user request.

A specification precise enough to run, so that it generates or drives tests, simulations or code without manual translation.

An executable specification is a description of required behaviour written precisely enough that a machine can check it or act on it. Instead of a document that people read and then translate into tests, the specification itself runs: concrete examples bound to test code, schemas that validate data, contracts that check an API, or constraints that bound what a language model may return. It bridges human intent and automation, reduces ambiguity and keeps the specification and the system in step. In spec-driven and AI-assisted development it is what lets generated code be verified against what was actually asked for.

The ability to explain, in terms people can follow, how an AI model reached a given output or decision (also called explainable AI, XAI).

Explainability is the capacity to describe how an AI model arrives at its outputs in a way that the people affected can understand. It covers global explanations (which factors drive the model overall) and local ones (why this customer was declined). It matters for trust, for debugging models, for detecting bias and for regulatory compliance, since laws such as GDPR and the EU AI Act give people and authorities rights to meaningful information about automated decisions. Explainable AI (XAI) is the set of methods used to produce those explanations.

Turning unstructured text, such as emails, contracts or invoices, into structured data: entities, fields, relations and events.

Information extraction is the task of pulling structured data out of unstructured content. From a contract it can take the parties, dates and amounts; from an email, the customer, the product and the request; from a report, the events and how they relate. Traditionally this needed rules or models trained for each field and document type. Large language models can now extract many fields from varied documents with instructions and a few examples, often returning the result directly in a fixed format such as JSON, which makes extraction one of the most practical uses of AI in document-heavy processes.

A software engineer who works inside the client's organisation, on its systems and data, until a solution runs in production.

A forward deployed engineer (FDE) is a software engineer embedded with a client rather than working on a product from a distance. The role mixes engineering with direct contact with the people who use the system: the FDE learns the client's processes, connects the platform or model to its real data and systems, writes the code that is missing and stays until the solution works in daily operation. The model was popularised by Palantir, where these engineers are called Deltas, and is now common at AI companies, because AI is easy to demo and hard to deploy.

A way to train a shared model across many devices or organisations, keeping raw data where it is and sharing only model updates.

Federated learning trains a machine learning model across many separate data holders, such as phones, hospitals or bank branches, without collecting their raw data in one place. Each participant trains the current model on its own data and sends back only the resulting update; a coordinating server combines the updates into a new shared model and repeats the process. The approach was introduced by Google researchers, whose federated averaging algorithm remains the standard reference. It suits regulated and privacy-sensitive settings where data cannot or should not leave its source.

Getting a model to perform a task from only a few examples, which large language models can do by reading the examples in the prompt.

Few-shot learning means getting a model to handle a task from very few labelled examples, which is useful when labelled data is scarce or expensive. With large language models it usually takes the form of in-context learning: a handful of input and output examples are placed in the prompt and the model continues the pattern, with no retraining. The idea became widely known when researchers showed that a sufficiently large language model could perform many tasks this way. It is quick to set up and easy to change, but results depend heavily on which examples are chosen.

Continuing the training of a pre-trained model on a smaller, task- or domain-specific dataset so that it behaves the way you need.

Fine-tuning takes a model that has already been pre-trained on large amounts of general data and trains it further on a smaller set of examples from your domain or task. The model's weights change, so the new behaviour is built into the model rather than written into each prompt. It is used to teach a format, a style, a vocabulary or a narrow task that prompting alone does not get right consistently. It is an alternative to prompt engineering when more control is needed, and it is often combined with retrieval rather than replacing it.

A large model trained on broad data that can be adapted, by prompting or fine-tuning, to many different tasks.

A foundation model is a large machine learning model trained on very broad data, usually with self-supervised learning, so that it picks up general patterns of language, images or code. Instead of training a new model for each task, teams adapt the same foundation model through prompting, retrieval or fine-tuning. The term was introduced by researchers at Stanford in 2021 to describe this shift. Large language models such as the GPT, Claude and Llama families are the best-known examples, and there are also foundation models for images, speech, code and other data types.

A document that describes what a system must do from the user's and the business's point of view: actors, flows, rules and expected outcomes.

A functional specification describes what a system must do, seen from the point of view of its users and the business, without going into how it is built. It covers who uses the system, the journeys and flows they follow, the business rules, the expected results and the acceptance criteria, plus product-level expectations such as performance or accessibility. It feeds product requirement documents, user stories and, increasingly, the instructions and guardrails given to AI agents, so that what gets built matches the agreed intent. Technical specifications then explain how to deliver it.

Combining information from several sources, modalities or retrieval methods (text and images, keyword and vector search) into one richer result.

Fusion is the integration of data or representations from more than one source into a single output. In multimodal AI, it means combining modalities such as text, images, audio or sensor data so a model can reason over all of them together. In search and RAG, it means merging the results of several retrieval methods, for example keyword search and vector search, or several reformulations of the same query, into one ranked list. In both cases the goal is the same: each source covers gaps in the others, which improves the accuracy and richness of the result.

AI that creates new content, such as text, images, code or audio, instead of only classifying or predicting from existing data.

Generative AI is a family of models that produce new content in response to an instruction: text, images, code, audio or video. Instead of assigning a label to an input, the model learns the patterns of very large datasets and generates an output that follows them. It powers chat assistants such as ChatGPT, image generators, code assistants and a growing number of enterprise applications. Its strength is flexibility; its weakness is that outputs can be fluent and still wrong, so it needs grounding, evaluation and human review where the stakes are high.

Generative pre-trained transformer: a type of language model that is pre-trained on large text corpora to predict the next token and then generates text.

GPT stands for generative pre-trained transformer, an approach to building language models introduced by OpenAI. The model uses the transformer architecture, is first pre-trained on a very large body of text to predict the next token, and can then be adapted to specific tasks. Because it generates text one token at a time, each conditioned on the previous ones, it is called autoregressive. GPT is also the name of OpenAI's model family behind ChatGPT, but the same paradigm underlies many other large language models.

Grounding ties a model's answers to verifiable sources, such as documents, databases or search results, instead of relying only on what it memorised.

Grounding is the practice of anchoring an AI model's answers in verifiable information supplied at the moment of the request, such as company documents, database records, API results or web search, rather than relying only on what the model learnt during training. The model is instructed to answer from that material and, ideally, to cite it. Retrieval-augmented generation (RAG) is the most common way to ground a language model. Grounding reduces hallucinations, keeps answers current and lets people check where each statement comes from.

Runtime controls and policies that check what goes into and comes out of an AI model, to keep its behaviour safe, on topic and within the rules.

AI guardrails are the controls placed around a language model, not inside it, that decide what it may receive, say and do. They combine classifiers that flag harmful or off-topic content, allow and deny lists, structured output schemas, checks on tool calls and escalation to a person when the system is unsure. They work at runtime, between the user and the model, so they can be changed without retraining. They are essential in customer-facing assistants and regulated workflows, where one bad answer has a real cost.

When a language model produces information that sounds plausible but is false, invented or not supported by its sources.

An AI hallucination is an output from a large language model that reads as confident and fluent but is wrong: an invented fact, a citation that does not exist, a number that is not in the document or a claim that contradicts the context the model was given. It happens because these models generate the most likely next words, not verified facts, so a gap in knowledge turns into a plausible guess instead of an admission. Hallucinations cannot be removed entirely, but grounding the model in sources (for example with RAG), clear instructions, citations and confidence checks reduce them in production systems.

An architecture that separates content and business logic from presentation, delivering them through APIs to any channel.

Headless architecture decouples the back end, where content, data and business logic live, from the front end that presents them. The back end has no built-in presentation layer, the 'head'; instead it exposes everything through APIs, usually REST or GraphQL. Websites, mobile apps, kiosks, voice assistants and AI agents can then consume the same content and services and present them in their own way. This makes omnichannel experiences easier to build and lets new channels, personalisation and AI be added without rebuilding the core.

A design in which people review, approve or correct an AI system's outputs before they take effect, or step in when the system is unsure.

Human-in-the-loop (HITL) is a way of designing AI systems so that a person reviews, approves or corrects the output before it has an effect. The person may check every decision, only the uncertain or high-impact ones, or a sample. HITL is common in high-stakes decisions, such as credit, claims, hiring or medical support, and in agents that act on real systems. It also produces corrections that can be used to improve the model. The EU AI Act requires high-risk systems to be designed so that people can oversee them effectively.

A retrieval approach that combines keyword search (such as BM25) with semantic vector search and merges both result lists.

Hybrid search runs two kinds of retrieval over the same content and merges the results. Keyword search, usually with a ranking function such as BM25, finds documents that contain the exact words of the query, which matters for product codes, policy numbers, names and jargon. Vector search finds documents with a similar meaning even when they use different words. Combining them gives better coverage than either alone, which is why hybrid search is a common default for enterprise search and retrieval-augmented generation (RAG).

An approach that automates whole business processes, end to end, by combining RPA, AI, workflow orchestration and system integration.

Hyperautomation is the effort to automate as many business processes as makes sense, from start to finish, rather than one task at a time. Gartner, which popularised the term, describes it as rapidly identifying, vetting and automating processes that used to require human judgement or action. It combines several technologies: robotic process automation (RPA) for repetitive screen work, AI for reading documents and making decisions, workflow orchestration to coordinate the steps and integration to connect the systems involved. The goal is to transform end-to-end operations, not just to speed up isolated tasks.

The phase in which a trained model is used to produce outputs, such as a prediction, a classification or generated text, from new inputs.

Inference is the moment a trained model is put to work: it receives a new input and produces an output, such as a forecast, a label, an image or an answer from a language model. Training happens once or from time to time; inference happens every time someone uses the system. That is why, in production, most of the running cost and the user's experience of speed depend on inference. Optimising it means balancing latency, throughput, cost and quality through choices of model size, hardware and serving techniques.

Connecting AI models and agents to a company's existing systems, data and processes so they can read real information and act in real workflows.

AI integration is the work of connecting AI capabilities to the systems a company already runs: ERP, CRM, data warehouses, document repositories, identity and messaging. A model on its own only sees what it is sent; integration lets it read current data, respect each user's permissions and write results back where people work. It covers APIs, data pipelines, authentication, error handling and monitoring. In most enterprise projects it is integration, not the choice of model, that decides whether AI moves from a demo to daily use.

The goal behind what a user says or writes, such as booking a flight or requesting a refund, as detected by a conversational system.

In conversational AI, an intent is the user's goal behind a message: for example, check an order, book an appointment or request a refund. Intent recognition, also called intent classification, maps each utterance to the intent that best fits it, and that decision drives what the chatbot or voicebot does next: which flow to start, which data to ask for, which system to call or when to hand over to a person. Intents usually come with entities or parameters, the specific details such as a date, a product or an account number. Large language models make recognition more flexible, but a clear catalogue of intents is still what keeps a service predictable.

The network of connected sensors and devices that collect and transmit data about the physical world, data that feeds AI models and automation.

The Internet of Things (IoT) is the set of physical devices, from sensors and meters to machines and vehicles, that measure something about the world and send that data over a network. On its own, IoT produces streams of readings; AI is what turns them into decisions. IoT data feeds predictive maintenance, anomaly detection, demand forecasting and autonomous systems in manufacturing, logistics, energy and smart buildings. The combination is sometimes called AIoT, and it increasingly runs part of the AI on the device or at the edge, close to where the data is produced.

Techniques that try to make an AI model ignore its safety rules and produce content or actions it was trained or instructed to refuse.

Jailbreaking is the use of crafted inputs to make an AI model bypass its safety restrictions and produce output it would normally refuse, such as harmful instructions, confidential information or policy-breaking content. Typical techniques include role-play, hypothetical framing, encoding the request, splitting it into innocent-looking parts or overwhelming the model with examples. Research shows it works because safety training competes with the model's goal of being helpful and does not always generalise to unusual inputs. For companies, jailbreaking is a security and reputational risk, and defending against it is part of AI governance.

A structured repository of an organisation's information, such as documents, FAQs and policies, that people and AI assistants can search and answer from.

A knowledge base is a structured repository of an organisation's knowledge: documents, FAQs, policies, procedures, product sheets and records of past cases. Traditionally it was a help centre that people searched; today it is also the source that retrieval-augmented generation (RAG) systems and AI assistants use to answer questions from company data. For that, the content is split into fragments, indexed (often as embeddings in a vector database) and kept with its metadata and access permissions. Thinkia Knowledge Core is an example of this kind of system.

Training a smaller student model to reproduce the behaviour of a larger teacher model, so it runs faster and cheaper while keeping much of its quality.

Knowledge distillation transfers what a large teacher model has learned to a smaller student model. Instead of learning only from the correct labels, the student learns from the teacher's outputs, which carry richer information, for example how likely the teacher thought each alternative answer was. The result is a compact model that is cheaper and faster to run in production, often with a modest loss in quality on the target task. It is commonly combined with quantisation and small language models (SLMs) to build efficient assistants.

A way of storing information as a network of entities and the relationships between them, which gives AI systems structure and context that plain text lacks.

A knowledge graph represents information as nodes (entities such as customers, products, contracts or people) connected by labelled edges (relationships such as "signed", "supplies" or "reports to"). Because the connections are explicit, a system can follow them to answer questions that span several sources, such as which suppliers are affected by a delayed part. Google popularised the term in 2012 to describe search based on "things, not strings". In enterprise AI, knowledge graphs improve retrieval and reasoning in RAG by adding structure and meaning beyond what text search finds.

Finding and returning the most relevant documents, passages or facts for a query, so that a person or an AI model can use them.

Knowledge retrieval is the process of finding the information that best answers a query in a body of knowledge: documents, wikis, tickets, databases or knowledge graphs. It uses keyword search, semantic search based on embeddings, or a hybrid of both, often followed by a reranking step. In generative AI it is the 'R' in RAG: the retrieved passages are given to a language model so that its answer is based on the organisation's own, current knowledge. The quality of the final answer depends heavily on retrieving the right content in the first place.

A specification kept in version control next to the code and updated with every change to the system, rather than a static document that goes out of date.

A living specification is a description of what a system must do and how its parts fit together, kept in the same repository as the code and updated whenever the system changes. It usually combines readable documents with machine-readable artefacts: API contracts such as OpenAPI files, architecture decision records, policies and evaluation definitions. Because it is versioned and reviewed like code, it stays true to the running system instead of drifting like a PDF. AI tools can read it to generate code and tests, spot differences between spec and implementation and propose updates when contracts change.

A neural network trained on very large amounts of text to predict the next word, which lets it write, summarise, translate and answer questions.

A large language model (LLM) is a neural network, usually based on the transformer architecture, trained on vast text corpora to predict the next token in a sequence. Doing that at scale gives it a broad ability to generate coherent text, follow instructions, answer questions, summarise, translate and write code. Well-known families include GPT, Claude, Gemini, Llama and Mistral. On its own an LLM only knows what was in its training data, so in business it is usually combined with retrieval of company documents, tools and guardrails.

Tracing, metrics and logs for prompts, retrieval, tool calls, cost and quality, so you can see what an LLM application did and why.

LLM observability is the practice of capturing what happens inside an application built on language models: the prompts sent, the documents retrieved, the tools called, the responses returned, and the latency, token use and cost of each step. It applies the general idea of observability, understanding a system's internal state from its outputs, to systems whose behaviour is probabilistic. Quality scores and user feedback are recorded alongside the technical signals. It closes the loop between experimentation and reliable production agents, because problems that only appear with real users become visible and traceable.

A branch of AI in which systems learn patterns from data to make predictions or decisions, instead of following rules written by hand.

Machine learning (ML) is the part of artificial intelligence in which a system improves at a task by learning from examples rather than from explicit programming. A model is trained on data, adjusts its internal parameters to reduce its errors, and is then used to predict or classify new cases: whether a transaction is fraud, how much demand to expect, which product a customer might want. The main families are supervised, unsupervised and reinforcement learning. ML is the foundation of modern AI, including the deep learning behind today's large language models.

An open standard that defines how AI applications connect to tools, data and enterprise systems through a common client-server protocol.

The Model Context Protocol (MCP) is an open standard, introduced by Anthropic, that defines how an AI application discovers and uses external tools, data sources and prompts. Instead of writing a bespoke integration for every model and every system, a team wraps a system once as an MCP server, and any MCP-compatible application can use it. The protocol follows a client-server model: the host application opens one client per server, and each server exposes what it offers in a structured, self-describing way. This reduces integration work and makes agent architectures more composable, but it does not remove the need for permissions, authentication and governance.

Data that describes other data, such as author, date, type, owner or access rights, used to find, filter and govern content.

Metadata is data about data: the information that describes a document, a dataset or a record, such as who created it, when, what it is about, what format it has and who may see it. In search and RAG systems, metadata lets you filter results by date, department, product or permissions, and rank recent or authoritative sources higher. In data and AI governance it records lineage, ownership and licensing. Poor or missing metadata is one of the most common reasons enterprise assistants return outdated or unauthorised content.

The set of practices for deploying, monitoring and maintaining machine learning models in production in a reliable, repeatable way.

MLOps (machine learning operations) applies DevOps practices to machine learning: version control, automated pipelines, testing, deployment, monitoring and rollback, extended to data and models. The model code is only a small part of a production ML system; around it sit data collection and validation, feature pipelines, training, serving and monitoring. MLOps automates that whole chain so that models can be retrained, checked and released without manual steps. Without it, models that worked in a notebook tend to degrade or break once real data changes.

A short, structured document that describes an AI model's intended use, training data, evaluation results, limitations and ethical considerations.

A model card is documentation that travels with a trained model and explains what it is for and how well it works. The idea was proposed by researchers at Google in 2018: a standard format to report a model's intended use, the data it was trained and evaluated on, its performance across relevant groups and conditions, and its known limitations and risks. Model cards help people decide whether a model fits their use case, and they support governance, procurement and audit readiness alongside technical specifications. Model hubs such as Hugging Face now attach a model card to every published model.

AI that can take in and relate more than one type of data, such as text, images, audio or video, often within a single model.

Multimodal AI handles several types of data, text, images, audio and video, and relates them to each other, often in the same model. A multimodal model can describe a photo, answer questions about a scanned document, transcribe and summarise a call or find images from a text description. This lets AI work with information as it actually exists in a company, which is rarely text alone. Quality still varies by modality and task, so each use case needs its own testing.

The field of AI that enables computers to understand, analyse and generate human language, in text and speech.

Natural language processing (NLP) is the branch of AI concerned with human language. It covers understanding text and speech (what is being said, by whom, with what intent and tone) and producing it (answers, summaries, translations). The field moved from hand-written rules to statistical models and then to neural networks; today large language models handle many NLP tasks with a single model guided by instructions. NLP underpins chatbots and virtual assistants, search, document processing, translation and sentiment analysis.

A way of serving customers so that web, app, phone, chat, email and in-person channels work as one, with shared context and a consistent experience.

Omnichannel describes a customer experience in which all channels, web, mobile app, voice, chat, email and in person, work as one connected service rather than separate ones. A customer can start in one channel and continue in another without repeating themselves, because history, preferences and open requests are shared. That requires shared customer data and identity, integrated systems and consistent rules across channels. AI helps by keeping context between channels, personalising responses and letting virtual and human agents pick up where the last interaction left off.

Coordinating several models, agents, tools and services so they work together as one workflow, with routing, state, governance and error handling.

AI orchestration is the layer that decides which model, agent or tool does what, in what order, with what data, and what happens when something fails. A single model call rarely solves a business process; real workflows combine retrieval, several model calls, tool actions, human approvals and integrations with existing systems. Orchestration coordinates those steps, keeps track of state, routes each task to the right component and applies policies on cost, security and compliance. Thinkia Synapse is an orchestration platform that unifies governance, routing and execution.

When a model learns its training data too closely, including the noise, and then performs poorly on new data it has not seen.

Overfitting happens when a model fits its training data so closely that it memorises details and noise instead of learning the general pattern. It scores well on the data it was trained on but makes poor predictions on new data. The usual sign is a gap that opens between performance on training data and on a separate validation set. Common causes are a model that is too complex for the amount of data, too little or unrepresentative data, and training for too long. Regularisation, more and better data, early stopping and proper validation help prevent it.

A parameter is a numerical value a model learns during training; large language models have billions of them, and their count gives the model's size.

Parameters are the numerical values, mainly the weights of the connections in a neural network, that a model adjusts during training to fit its data. Together they encode what the model has learnt, and once training ends they stay fixed unless the model is fine-tuned. Large language models have billions of parameters, and the size labels in model names, such as 7B or 70B, refer to that count. More parameters generally allow more capability, but they also require more memory and compute to run.

A sequence of processing steps, such as ingestion, preparation, model and post-processing, that turns input data into a result.

A pipeline is a chain of processing steps in which the output of one step becomes the input of the next: for example, data ingestion, cleaning, feature preparation, model inference and post-processing. In machine learning, pipelines cover both how models are trained and how they are used in production. In generative AI, a pipeline can chain retrieval, prompt construction, model calls, validation and actions. Building AI systems as pipelines makes each step testable, repeatable and replaceable, which is what makes them reliable and scalable.

Shared infrastructure that brings models, agents, data connections, workflows and governance together in one place, so teams build AI use cases on a common base.

An AI platform is the common layer on which an organisation builds and runs its AI use cases. It centralises access to models, the connections to company data and systems, agents and workflows, and the controls around them: security, permissions, monitoring and cost. Instead of each team assembling its own stack, they reuse the same components and rules, which makes new use cases faster to launch and easier to govern. Thinkia Synapse is an enterprise AI platform for orchestration, security and cost control.

A document that describes the problem, users, scope, success metrics and requirements of a product or feature, so the whole team builds the same thing.

A product requirements document (PRD) sets out what a product or feature should do and why: the problem it solves, who it is for, what is in and out of scope, the requirements and how success will be measured. Its job is to align business, design and engineering before and during development. Agile teams favour short, living PRDs over exhaustive specifications. In AI initiatives and spec-driven development, a clear PRD also feeds acceptance criteria, evaluation rubrics and system prompts, so that model behaviour matches the intended outcome.

Using historical data, statistics and machine learning to estimate what is likely to happen next: demand, churn, failures, fraud or sales.

Predictive analytics uses historical data together with statistical and machine learning models to estimate future outcomes or unknown values. Typical questions are how much demand there will be next month, which customers are likely to leave, which machine is likely to fail or which transaction looks fraudulent. The output is usually a probability or a forecast, which is then used to act earlier: adjust stock, contact a customer or schedule maintenance. It sits between descriptive analytics, which explains what happened, and prescriptive analytics, which recommends what to do.

AI that takes the initiative, anticipating needs and acting or alerting before being asked, instead of only responding to queries.

Proactive AI anticipates needs and takes the first step without waiting for an explicit request: it raises an alert, suggests an action, starts a workflow or steers a conversation towards a goal. It shifts AI from reactive, answering when asked, to anticipatory, acting on signals such as a pattern in data, a change in a customer's situation or an upcoming deadline. Done well, it saves time and prevents problems; done badly, it interrupts and erodes trust. The design questions are when to act, how sure the system must be and how much it may do without a person.

The live environment where an AI system serves real users and real data, as opposed to development, testing or a pilot.

Production is the live environment where a system does its real job: real users, real data, real consequences. For AI, reaching production means more than deploying a model. The system needs monitoring of quality and cost, scaling, security, access control, logging and compliance, plus a team that owns it and a way to update it safely. Many AI initiatives work in a demo or a pilot and stall before this point, because the hard part is integrating with real systems and keeping behaviour reliable over time, not the model itself.

The input given to an AI model, such as instructions, context, examples and the question, that guides what it produces.

A prompt is everything a generative AI model receives as input before it responds: the instructions, the context, any examples and the user's request. It can be text alone or include images, documents and other files. The model has no other way of knowing what is wanted, so the content and structure of the prompt strongly shape the quality and consistency of the answer. In applications, the prompt is usually assembled by software from several parts, such as a fixed system prompt, retrieved documents and the user's message, rather than typed by hand each time.

The practice of designing and testing the instructions given to a language model so that its answers are better and more consistent.

Prompt engineering is the practice of writing, structuring and testing the inputs given to a large language model (LLM) so that it produces reliable, useful outputs. It covers clear instructions, worked examples (few-shot), step-by-step reasoning (chain of thought), roles and structured output formats such as JSON. In a business setting it is less about clever wording and more about method: define what a good answer looks like, build a set of test cases and change the prompt until it passes them. It is usually the cheapest way to improve a model's behaviour before considering fine-tuning or a different model.

An attack that hijacks a language model's instructions through untrusted input, from a user or from content the model reads.

Prompt injection is an attack in which text crafted by an attacker makes a language model ignore its original instructions, leak data or take actions it should not. The malicious instructions can be typed directly by a user or hidden in content the model processes, such as a retrieved document, an email or a web page. It works because a model receives instructions and data in the same channel and cannot reliably tell them apart. Defences include isolating untrusted content, filtering outputs, allowlisting tools and keeping the model's permissions strictly scoped.

Reducing the numerical precision of a model's weights or activations, for example to INT8 or INT4, to shrink it and make inference faster and cheaper.

Quantization is a technique that stores and computes a model's numbers with fewer bits, for example 8-bit or 4-bit integers instead of 16- or 32-bit floating point. The model takes less memory, moves less data and can run faster on GPUs, TPUs, NPUs or even CPUs. The trade-off is some loss of accuracy, which modern methods keep small for many tasks. Quantization is central to deploying large language models at reasonable cost and to running them on smaller hardware, on premises or on devices.

A technique that retrieves relevant passages from a knowledge base and gives them to a language model as context, so its answer rests on those sources.

Retrieval-augmented generation (RAG) combines a search step with a language model. When a question arrives, the system looks up the most relevant passages in a document collection, usually through a vector index, and passes them to the model together with the question. The model then writes its answer from that material instead of relying only on what it learnt in training. This keeps answers tied to current, private sources, lets the system cite where each answer came from and reduces hallucinations, without retraining the model.

A security practice where a dedicated team simulates adversarial attacks against an AI system to expose vulnerabilities before it reaches production.

Red teaming is a proactive security measure adapted from cybersecurity to evaluate AI models and applications. Instead of waiting for users to find flaws, a specialised group actively tries to break the system, bypassing safety filters, extracting sensitive data or triggering hallucinations. This adversarial testing helps you map the actual limits of your large language models (LLMs) and apply precise guardrails. As AI adoption scales, red teaming shifts from a manual exercise to an automated, continuous requirement.

A way of training an agent by trial and error: it takes actions, receives rewards or penalties, and learns to maximise the total reward over time.

Reinforcement learning (RL) is a branch of machine learning in which an agent learns by interacting with an environment. At each step it observes the situation, takes an action and receives a reward or penalty; over many trials it learns a policy that maximises cumulative reward. Unlike supervised learning, nobody gives it the right answer for each case; it has to discover which actions pay off, sometimes only after a long delay. RL is used in robotics, games, control and optimisation, and a variant that uses human feedback is a key step in training today's chat models.

Developing and using AI with care for ethics, legal compliance and social impact, through clear principles, roles and controls.

Responsible AI is the practice of developing, deploying and using AI with attention to ethics, legal compliance and social impact. It includes principles such as fairness, transparency, accountability, privacy, safety and human oversight, and above all the processes that make them real: governance, risk assessment, testing, documentation and monitoring. Frameworks such as the NIST AI Risk Management Framework and the OECD AI Principles give organisations a structure, and laws such as the EU AI Act make part of it mandatory. In practice, it turns good intentions into decisions someone owns and can show.

The search step that fetches the documents or passages most relevant to a query, before they are shown or passed to a model.

Retrieval is the step in RAG and in search systems that finds, among a large collection, the documents or text passages most relevant to a query. It usually returns a small ranked list, the top-k results, from an index built in advance. It can work by matching keywords (sparse retrieval), by comparing embeddings in a vector space (dense retrieval) or by combining both. In a RAG system, what retrieval returns is all the model sees of your knowledge, so its quality sets the ceiling for the quality of the answer.

Return on investment: the net gain from an investment relative to its cost; for AI projects, the key measure for justifying and sustaining the spend.

Return on investment (ROI) compares what an investment returns with what it costs, usually as net gain divided by cost. For AI projects it is the measure that decides whether a pilot gets funded, scaled or stopped. Showing it means linking the system to measurable value, such as cost savings, revenue growth, faster processes or fewer errors, and counting the full cost, not just the model. Because many AI benefits are indirect or arrive over time, a clear baseline and agreed metrics before launch matter more than the formula itself.

Software bots that repeat rule-based tasks a person does on screen, such as copying data between systems or filling in forms.

Robotic process automation (RPA) uses software bots that mimic how a person interacts with applications: they click, type, copy and paste, read screens and move data between systems. It suits repetitive, rule-based tasks with structured inputs, such as data entry, reconciliations or report generation, especially where systems have no API. On its own RPA follows fixed scripts and breaks when screens or rules change. Combined with AI for reading documents and making simple decisions, it becomes part of intelligent automation or hyperautomation.

Search that finds content by what it means rather than by matching the exact words, usually by comparing embeddings of the query and the documents.

Semantic search retrieves content based on meaning instead of literal keyword matches. Queries and documents are converted into embeddings, and the system returns the documents whose vectors are closest to the query's. That lets it find a policy on "time off" when someone asks about "holidays", or a document in another language. It is the retrieval layer behind most enterprise assistants and RAG systems, often combined with keyword search and metadata filters.

The use of AI to identify the tone or emotion in text, such as positive, negative or neutral, and often what it refers to.

Sentiment analysis classifies the attitude expressed in a piece of text: whether a review, a message or a survey answer is positive, negative or neutral, and sometimes which emotion it shows, such as frustration or satisfaction. More detailed forms, known as aspect-based sentiment analysis, link each opinion to what it refers to, for example a customer who likes the product but complains about delivery. It started with word lists and classic machine learning; today language models handle irony, context and mixed opinions better. Businesses use it to track customer feedback, social media, brand perception and service quality at scale.

The use of AI tools, such as public chatbots or AI features in apps, by employees without approval or oversight from IT, security or compliance.

Shadow AI is the use of AI tools and services inside an organisation without the knowledge or approval of IT, security or compliance. A typical case is an employee pasting customer data or internal documents into a public chatbot to draft a reply or summarise a report. It also includes AI features switched on inside software the company already uses, and personal accounts used for work. The risks are data leakage, regulatory breaches and decisions based on unchecked outputs. The usual answer is not a ban but approved tools, clear data rules and safe defaults.

A governed registry of the skills AI agents can use, with names, versions, owners, required tools and dependencies.

A skill catalog is the registry where an organisation keeps track of the skills its AI agents can load: packaged instructions, scripts and resources that teach an agent how to do a specific task. For each skill it records what it does, who maintains it, which version is current, what tools and permissions it needs and how it combines with other skills. It prevents unmanaged prompts spreading across teams (shadow prompts), makes skills reusable and auditable, and fits with spec-driven change: the spec is updated first, then the skill package, then the deployments that use it.

A compact language model built for narrower tasks, with lower cost and latency than frontier models and small enough to run on modest hardware.

A small language model (SLM) is a language model with far fewer parameters than frontier LLMs, designed to do a narrower set of tasks well. SLMs are often trained on carefully selected data, distilled from a larger model or fine-tuned for one domain or workflow. Because they need less compute, they are cheaper and faster per request and can run on a single server, a laptop or a phone. They suit high-volume assistants, on-device use and hybrid set-ups where a router sends only the hard cases to a larger model.

Using product and system specifications, such as requirements, API contracts and guardrails, as the source that steers AI models, agents and evaluations.

Spec-driven AI treats written specifications as the source of truth for how an AI system should behave. Requirements documents, acceptance criteria, API and tool schemas and policy rules are connected to the prompts, the retrieval content and the evaluation suites, so that a change starts in the spec and flows from there. This keeps behaviour traceable: anyone can see which requirement a prompt, a tool or a test comes from. It is a more governable alternative to adjusting prompts by trial and error, and it is closely related to spec-driven development, where AI coding agents build software from specifications.

A way of building software in which an explicit, reviewable specification is written first and code, tests and docs are derived from it.

Spec-driven development (SDD) treats the specification as the source of truth: requirements, acceptance criteria, API contracts and event schemas are written and reviewed before the code. Teams then generate or write stubs, tests and documentation from the spec, which reduces drift between what was agreed and what was built. With AI coding agents the idea has gained weight, because a precise spec gives the model a bounded contract to work to instead of a vague request. The spec is a living document that is updated as understanding improves, and the code follows it.

Thinkia's agentic platform for enterprise AI: it orchestrates agents, models and workflows with centralised governance, security and cost control.

Thinkia Synapse is a unified agentic platform for enterprise AI. It orchestrates agents, models and workflows, and brings governance, security and cost control into one place instead of leaving them to each project. Thinkia describes it as a central nervous system for AI: the layer that connects the different pieces an organisation uses and keeps them under common control. The underlying idea is that, as companies move from single pilots to many agents, coordination and oversight become as important as the individual models.

Artificially generated data that reproduces the statistical patterns of real data, used to train, test or share models when real data is scarce or sensitive.

Synthetic data is data created by a model or simulation rather than collected from real events, built to keep the useful statistical properties of an original dataset. It is used to enlarge scarce datasets, balance rare cases, test systems without exposing personal data and simulate situations that are hard to observe. Its quality depends on the original data and on the model that generates it: it can miss outliers, inherit the biases of the source and, if poorly made, still leak information about real individuals. Governance is needed to check that synthetic data is fit for purpose and does not encode harmful bias.

Persistent instructions that set a model's role, rules, tools and output format before any user message is read.

A system prompt is the set of instructions an application gives a large language model before the conversation starts and keeps in place for every turn. It defines the model's role, tone, policies, what it may and may not do, which tools it can use and the format of its answers. Users normally do not see it; they only write the user messages that follow. Because it shapes every answer, the system prompt should be versioned, tested and reviewed like code, as part of operating language models and enterprise guardrails.

Breaking a goal into smaller subtasks that an AI agent, a set of agents or a team can carry out and check one by one.

Task decomposition turns a high-level goal into smaller steps, ordered or parallel, each with clear inputs, outputs and a way to tell when it is done. In AI systems, it is what lets a language model or a group of agents handle work too large or varied for a single prompt. The decomposition can be written by people, as in a defined workflow, or produced by the model itself as a plan. It connects specification work, use cases and acceptance criteria, with step-by-step reasoning, tools and orchestration, and it is one of the main ways to keep agents from drifting beyond their scope.

An engineering blueprint that describes how a system will be built: services, data models, APIs, infrastructure, risks and rollout plan.

A technical specification, or tech spec, turns what a system must do into how it will be built. It describes the architecture and components, the data models, the APIs and integrations, the infrastructure, the non-functional targets such as performance and availability, the security measures, and the migration and rollout plan, together with the trade-offs considered. It is written before the main build so that problems surface while they are still cheap to fix and the team agrees on one design. For AI systems it also covers the choice of model, evaluation suites, fallback paths and observability, so that tools and agents built on top have a stable contract to target.

A sampling parameter that controls how random a language model's choice of the next token is: low is focused, high is varied.

Temperature is a parameter applied when a language model generates text, controlling how much randomness goes into choosing each next token. The model assigns a probability to every possible next token; temperature reshapes that distribution before a token is sampled. A low temperature concentrates the choice on the most likely tokens and gives focused, repeatable answers; a higher temperature spreads it out and gives more varied, creative and less predictable text. Choosing it is part of matching a model to the task, from factual extraction to brainstorming.

The allocation of additional processing power during the inference phase, allowing an AI model to generate, evaluate and refine intermediate steps before delivering a final answer.

Historically, improving AI performance meant scaling up the computing power used during pre-training. Test-time compute shifts this focus to the inference phase, when the model is actually running. By dedicating more processing time and resources to a specific query, reasoning models can think through a problem, explore multiple paths, check their own logic and correct mistakes before responding. This approach, popularised by models such as the OpenAI o1 series, allows smaller models to tackle highly complex reasoning tasks that would otherwise require massive pre-training scale.

The basic unit of text a language model reads and writes: a word, part of a word or a symbol. In English, one token is roughly four characters.

A token is the unit of text a language model actually processes. Before the model sees any text, a tokeniser splits it into pieces that can be whole words, parts of words, punctuation or spaces, and maps each piece to a number. In English, a common rule of thumb is about four characters, or three quarters of a word, per token; other languages and code often need more tokens for the same content. Token counts matter in practice because providers charge per token and because the context window, the amount a model can handle at once, is measured in tokens.

The ability of a language model to emit a structured request to run a function, API or query, which the host application validates and executes.

Tool calling, also known as function calling, lets a large language model ask for an action instead of only writing text. The application describes the available tools to the model, each with a name, a description and an input schema. When the model decides a tool is needed, it returns a structured call with arguments; the application validates it, runs the function (an SQL query, an HTTP request, a CRM update) and sends the result back so the model can continue. The model never executes anything itself. Tool calling is what turns a chat model into a copilot or an agent that works with real systems and data.

The process of teaching a model from data by adjusting its parameters, from general pre-training to task-specific fine-tuning.

Training is the process by which a machine learning model learns from data. The model makes predictions, an error measure compares them with the expected result, and an optimisation algorithm adjusts its parameters to reduce that error, over many passes through the data. For large language models, training happens in stages: pre-training on vast amounts of general text, then fine-tuning on specific tasks or domains, and often alignment with human feedback. Training needs data, computing power and expertise; inference is what happens afterwards, when the trained model is used.

Reusing a model trained on one task or dataset as the starting point for a new task, so that it needs far less data and compute.

Transfer learning is the technique of reusing what a model learnt on one task or domain to solve another. Instead of training from scratch, you start from a pre-trained model, which already captures general patterns of language, images or other data, and adapt it to the new task with a smaller dataset. In language, work such as ULMFiT and BERT made this the standard approach, and today's large language models take it further: a single pre-trained model transfers to new tasks through fine-tuning or simply through prompts. This is why companies can build useful AI without huge task-specific datasets.

A concrete business scenario in which AI is applied to a specific task, for specific users, with a measurable result.

An AI use case is a specific business scenario where AI is applied to a defined task: for example, automating first-line customer service, summarising claim files or predicting equipment failures. A well-defined use case says who uses it, what problem it solves, which data and systems it needs and how success will be measured. Identifying and prioritising use cases is the step that turns an AI strategy into projects. The term comes from software engineering, where a use case describes how a user interacts with a system to reach a goal.

A short requirement written from the user's point of view, often 'As a… I want… so that…', plus acceptance criteria.

A user story is a small, valuable piece of functionality described from the point of view of the person who will use it. It is often written as 'As a [type of user], I want [something] so that [benefit]', and completed with acceptance criteria that say how to know it is done. Stories link personas to outcomes and are used to order and plan a team's backlog. In spec-driven AI work they also anchor evaluation: each story should map to measurable acceptance criteria and be traceable to the prompts, tools and telemetry that implement it.

A list of numbers that represents a piece of data, such as a sentence or an image, so that similar content ends up with similar numbers.

In AI, a vector is an ordered list of numbers that describes something in a form a machine can compare. An embedding model turns text, images or other data into vectors, usually with hundreds or thousands of dimensions, arranged so that items with similar meaning sit close together. Measuring the distance between vectors then becomes a way of measuring how similar two pieces of content are. This is what makes semantic search possible, and it is the core of retrieval in RAG systems, where the question and the documents are compared as vectors.

A database built to store embeddings and return the stored vectors most similar to a query vector, quickly and at scale.

A vector database stores embeddings: lists of numbers that represent the meaning of a text, an image or another item. Given a query turned into an embedding, it returns the stored items whose vectors are closest, which usually means the most similar in meaning. To do this fast over millions of items it uses approximate nearest neighbour indexes such as HNSW, trading a small amount of accuracy for speed. It is a core building block of semantic search and retrieval-augmented generation (RAG), and most products also store metadata so results can be filtered by date, source or permission.

A model that understands images and text together, so it can describe pictures, read documents and answer questions about what it sees.

A vision-language model (VLM) is an AI model that takes images and text as input and reasons across both. It typically combines an image encoder, which turns pixels into numerical representations, with a language model that turns those representations and the user's question into an answer. This lets one model caption photos, read scanned documents, interpret charts and slides, or answer questions about a screenshot. In enterprise use, VLMs power document AI, visual inspection and multimodal assistants, and they extend RAG to content that is not plain text.

An AI assistant that holds a conversation by voice: it understands what the caller says and answers with synthesised speech.

A voicebot is a conversational AI system that interacts through speech instead of text. It converts the caller's speech into text, works out what they want, takes an action or finds an answer, and replies with a synthetic voice. Voicebots are used in contact centres, phone lines and IVR menus, in-car and hands-free applications. Language models have made them far more flexible than older menu-driven IVR systems, but latency, accents, noise and the hand-off to a person still decide whether callers accept them.

A defined sequence of steps, some run by AI models, that carries out a business process from trigger to result.

A workflow is a sequence of steps that accomplishes a business process, such as handling a claim, onboarding a customer or answering a request. An AI workflow includes steps performed by AI models, for example classifying a document, retrieving relevant information, drafting a reply or extracting fields, alongside ordinary steps such as API calls, rules and human approvals. The order and the branches are designed in advance, which makes workflows predictable, testable and easy to audit. Synapse, Thinkia's platform, orchestrates complex AI workflows of this kind.

Methods and practices that make an AI system's outputs understandable to people, so they can see why it reached a decision and judge whether to trust it.

Explainable AI (XAI) covers the methods and practices that let people understand why an AI system produced a given output. An explanation might show which inputs weighed most in a credit decision, which passage of a document supported an answer or how a model behaves across different cases. Explainability builds trust with users, helps teams find errors and bias, and supports regulatory requirements for transparency in high-stakes decisions. An explanation is only useful if it is accurate and suited to the person who receives it, which is harder than it sounds.

Running systems or operations without interruptions that users notice, both during planned changes and by preventing unplanned failures.

Zero downtime is the goal of keeping a system or an operation available without interruptions that users or customers notice. It has two sides. Planned downtime is avoided with deployment techniques such as rolling updates and blue-green releases, which change software without stopping the service. Unplanned downtime is reduced with redundancy, monitoring and, increasingly, predictive AI that spots signs of failure in equipment or systems before they stop. In practice, zero downtime is an engineering target supported by clear availability objectives, not an absolute guarantee.

A model's ability to perform a task it has not been shown examples of, guided only by an instruction or description of the task.

Zero-shot means a model handles a task without being given examples of that task, either in training or in the prompt. With large language models, zero-shot prompting is simply describing what you want, such as classify this email or summarise this contract, and letting the model do it. It works because large models learn broad patterns of language and tasks during pre-training. Zero-shot is the cheapest and fastest way to start, but for specialised formats or domains, adding a few examples (few-shot) or fine-tuning often improves accuracy and consistency.