AI Glossary

AI, explained in plain English.

The words that come up whenever AI is on the agenda, from agents to inference, each explained in a sentence or two you can share with your team.

140 terms

A

Agent HandoffAgents & Orchestration: The moment one AI agent passes a task, along with what it has learned so far, to another agent or to a person.; Good handoffs carry the full context, so nobody has to start again or ask the customer the same questions twice. In a bank, an agent that has gathered a mortgage applicant's documents might hand the case to an underwriting agent, or to a human underwriter when something looks unusual.
Agent SkillsAgents & Orchestration: Packaged instructions, examples and tools that teach an AI agent how to carry out a specific kind of task well.; Skills let you add new abilities to an agent without rebuilding it, much like giving a new colleague a well-written procedure manual. A maintenance agent might have one skill for reading vibration reports and another for raising work orders in the plant's maintenance system.
Agentic AIAgents & Orchestration: AI that can plan steps, take actions in other systems and adjust its approach to reach a goal, rather than only answering questions.; The shift from answering to acting is what makes agentic AI useful for real operational work, and also what makes careful permissions and oversight essential. Instead of drafting a reply about a missed payment, an agentic system could check the account, propose a payment plan and log the outcome for review.
Agentic WorkflowAgents & Orchestration: A business process where one or more AI agents handle several steps in sequence, deciding what to do next based on what they find.; Unlike a fixed script, an agentic workflow can cope with variation, such as a missing document or an unexpected answer, by choosing a sensible next step. A supplier onboarding workflow might gather certificates, check them against requirements and chase the supplier only for what is missing.
AI AgentAgents & Orchestration: A piece of software that uses an AI model to work towards a goal by reasoning, using tools and taking actions on your behalf.; An agent is defined by what it is allowed to do: which systems it can read, which it can change, and when it must ask a person first. A well-scoped agent at an energy retailer might investigate billing complaints by pulling meter readings and tariff history before suggesting a resolution.
AI AutomationAutomation & Process: Using AI to carry out tasks that previously needed human judgement, such as reading, classifying, summarising or deciding.; Traditional automation follows fixed rules; AI automation can handle messy inputs like emails, scanned forms and free-text notes. A trade finance team might use it to read incoming letters of credit and flag the clauses that need a specialist's attention.
AI ConsultingBusiness & Engagement: Outside expertise that helps an organisation decide where AI fits, what to build first and how to adopt it safely.; Useful AI consulting ends in working systems and changed processes, not just a slide deck. The best engagements test ideas quickly on real data so that decisions about investment rest on evidence rather than enthusiasm.
AI CopilotAgents & Orchestration: An AI assistant that works alongside a person inside their usual tools, suggesting drafts, answers or next steps while the person stays in charge.; Copilots speed people up without removing their judgement, which makes them a common first step in regulated settings. A relationship manager might use one to prepare a client briefing from account history and recent news, then edit it before the meeting.
AI GovernanceGovernance & Risk: The policies, roles and controls an organisation uses to make sure its AI is safe, lawful, fair and used as intended.; Governance answers practical questions: who approves a new use of AI, how it is tested, who monitors it and what happens when it goes wrong. For banks and utilities, it is also how you show regulators that AI decisions are understood and under control.
AI HarnessAgents & Orchestration: The surrounding software that runs an AI model: feeding it context, giving it tools, enforcing limits and recording what it does.; The model is only one part of a working AI system; the harness decides how reliable, safe and auditable it is in practice. Two teams using the same model can get very different results depending on how well their harness manages instructions, data access and error handling.
AI IntegrationDeployment & Ops: Connecting AI to the systems, data and workflows an organisation already uses, so it can work with real information and fit into daily operations.; Most of the value, and most of the effort, sits in integration rather than in the model itself. An AI that drafts maintenance reports is far more useful when it can read sensor history from the plant historian and file the report straight into the asset management system.
AI Maturity ModelBusiness & Engagement: A framework describing the stages an organisation passes through as it moves from experimenting with AI to running it reliably at scale.; A maturity model helps leaders see where they really are and what has to change next, whether that is data quality, skills, governance or operating model. It is most useful as a conversation tool for setting priorities, not as a score to chase.
AI Readiness AssessmentBusiness & Engagement: A structured review of an organisation's data, systems, skills and controls to judge where AI can deliver value soon and what blocks it.; A good assessment produces a short, ranked list of opportunities and the specific gaps to close for each. For a manufacturer, it might reveal that quality data is ready to use but maintenance records are too inconsistent without some clean-up first.
AI SprawlGovernance & Risk: The uncontrolled spread of separate AI tools, pilots and subscriptions across an organisation without shared standards or oversight.; Sprawl leads to duplicated spending, inconsistent quality and data flowing to places nobody has checked. A simple inventory of who is using what, and for which data, is usually the first step towards bringing it under control.
AI-NativeBusiness & Engagement: Describes a business where AI is built into how work and decisions flow, rather than bolted on as separate tools.; In an AI-native organisation, processes are redesigned around what AI does well, with people focusing on judgement, relationships and exceptions. The difference is like rewiring a building versus plugging in extension leads: one changes how everything runs, the other just adds sockets.
API(Application Programming Interface)Deployment & Ops: A defined way for one piece of software to request information or actions from another, without needing to know how it works inside.; APIs are how AI systems read from and write to your existing platforms, such as a core banking system or an energy trading platform. Where good APIs are missing, integration becomes slower and more fragile, which is why they often shape what is realistic to build first.
Artificial General Intelligence(AGI)Foundations: A hypothetical AI that could learn and perform almost any intellectual task a person can, across fields, without being built for each one.; AGI does not exist today, and experts disagree on whether or when it will. For business planning, it is more useful to focus on what current AI can reliably do in specific, well-defined tasks.
Attention MechanismModels & Architecture: The part of a modern AI model that works out which words in the input matter most to each other when interpreting meaning.; Attention is what lets a model understand that "it" in a long contract clause refers to the borrower rather than the bank. It is the core idea behind the models used in today's language and document tools.

B

Batch InferenceDeployment & Ops: Running an AI model over a large set of items in one scheduled job, rather than responding to requests one at a time.; Batch processing is usually cheaper and simpler when results are not needed instantly. A bank might classify the previous day's transactions overnight, or a utility might score every asset for failure risk once a week.
BenchmarkModels & Architecture: A standard set of tests used to compare how well different AI models perform on particular tasks.; Public benchmarks are a rough guide, but they rarely reflect your documents, your language or your edge cases. The comparison that matters most is a custom test built from your own real work.
BiasGovernance & Risk: A systematic tendency for an AI system to produce unfair or skewed results, often because of patterns in the data it learned from.; Bias matters most where AI influences decisions about people, such as lending, hiring or fraud checks. Detecting it means testing outcomes across different groups and fixing the data or design when results differ for the wrong reasons.
Business Process Automation(BPA)Automation & Process: Using software to run repeatable business processes end to end, reducing manual hand-offs, re-keying and waiting.; The aim is a process that flows from start to finish with people stepping in only where their judgement adds value. Invoice approval, customer onboarding and purchase requisitions are typical starting points.

C

Chain-of-Thought Prompting(CoT)Prompting: Asking an AI model to work through a problem step by step before giving its answer, which often improves accuracy on complex tasks.; Showing the working also makes it easier for a person to check the reasoning, not just the conclusion. It is useful for tasks such as checking whether a set of covenants has been breached or reconciling figures across reports.
ChatbotAgents & Orchestration: Software that holds a text or voice conversation with people, typically to answer questions or guide them through simple tasks.; Older chatbots followed fixed scripts; newer ones use language models and can handle a far wider range of questions. A chatbot is most valuable when connected to accurate, current information and clear routes to a human when needed.
ChunkingRetrieval & Data: Splitting long documents into smaller passages so an AI system can search them and pull back only the relevant parts.; How you split a document affects whether the right information is found later. Cutting a safety procedure in the middle of a numbered step, for example, can separate a warning from the action it applies to.
Closed-Source Model(Proprietary Model)Models & Architecture: An AI model whose inner workings are kept private by its developer and which you typically access as a paid service.; Closed models are often the most capable and easiest to start with, but you depend on the provider for pricing, availability and data handling terms. Regulated firms weigh this against the extra control that self-hosted options offer.
Computer VisionModels & Architecture: AI that interprets images and video, recognising objects, defects, text or activity in what a camera sees.; In manufacturing, computer vision can spot surface defects on a production line that a tired inspector might miss. In energy, it can review drone footage of power lines or turbines to flag corrosion and damage.
Context EngineeringPrompting: Deciding exactly what information, instructions and tools an AI model sees at each step, so it has what it needs and nothing distracting.; Many AI failures come from the model being given the wrong, stale or too much information rather than from the model itself. Careful context engineering is often what turns an impressive demo into a system that works reliably every day.
Context WindowModels & Architecture: The maximum amount of text, measured in tokens, that an AI model can consider at once when producing a response.; Anything outside the context window is invisible to the model for that request. This is why long document sets, such as a full loan file or years of maintenance logs, usually need search and selection rather than being pasted in whole.
Conversational AIAgents & Orchestration: The broad family of technology that lets computers understand and respond to people in natural, back-and-forth dialogue.; It covers chat, voice and messaging channels, and ranges from simple question answering to systems that complete transactions. Success depends as much on knowing when to hand over to a person as on what the AI can do itself.
Cosine SimilarityRetrieval & Data: A mathematical measure of how closely two pieces of text, turned into lists of numbers, point in the same direction of meaning.; It is one of the main ways search systems decide which passages are most related to a question. You rarely need to touch it directly, but it explains why AI search can match "turbine fault" with "generator trip" even without shared words.
Custom AIBusiness & Engagement: AI built or adapted for one organisation's specific data, processes and rules, rather than an off-the-shelf product used as is.; Custom work makes sense where the process is a genuine differentiator or where generic tools cannot meet security and compliance needs. It usually combines existing models with bespoke integration, data handling and controls.

D

Data Annotation(Data Labelling)Retrieval & Data: Adding labels to raw data, such as tagging images or marking parts of documents, so AI systems can learn from or be tested against it.; Annotation by people who know the domain is often the difference between a usable model and a poor one. A quality engineer labelling photos of real defects, for instance, teaches a vision system far more than generic examples could.
Data GovernanceGovernance & Risk: The rules and responsibilities for how data is collected, stored, accessed, shared and kept accurate across an organisation.; AI amplifies whatever data it is given, so weak data governance quickly becomes an AI problem. Clear ownership and access rules decide which data an AI system is allowed to use and for what purpose.
Data PipelineRetrieval & Data: An automated chain of steps that moves data from where it is created, cleans and reshapes it, and delivers it where it is needed.; Reliable pipelines keep AI systems fed with current, consistent information. A utility might pipe smart meter readings, weather data and outage records into a single store that its forecasting models read from.
Data ReadinessRetrieval & Data: How prepared an organisation's data is for AI: whether it is accessible, accurate, complete enough and legally usable.; Data readiness is often the biggest factor in how quickly an AI project delivers. It is rarely all or nothing; some datasets can be used immediately while others need cleaning, linking or permissions sorted first.
Deep LearningFoundations: A type of machine learning that uses many-layered neural networks to learn complex patterns from large amounts of data.; Deep learning powers most of today's advances in language, image and speech recognition. It tends to need more data and computing power than simpler methods, so it is not always the right tool for smaller, structured problems.
DeploymentDeployment & Ops: Moving an AI system from testing into everyday use, where real people and processes depend on it.; Deployment brings new requirements: security reviews, monitoring, support, training and a plan for when things go wrong. Many AI projects stall here, which is why it pays to design for production from the first week.
Diffusion ModelModels & Architecture: A type of AI model that creates images or other media by starting from random noise and gradually refining it into a clear result.; Diffusion models are behind most modern image-generation tools. In industry they are also explored for generating synthetic training images, such as rare defect types that are hard to photograph in real life.
Digital WorkerAgents & Orchestration: A software agent set up to perform a defined role or set of tasks, often described in job-like terms with clear responsibilities.; Framing automation as a digital worker helps teams think about scope, supervision and accountability, just as they would for a new hire. A finance team might have one handling supplier statement reconciliations, with a named person reviewing its exceptions.
Distillation(Knowledge Distillation)Models & Architecture: Training a smaller, faster AI model to imitate a larger one, keeping much of its ability at a fraction of the running cost.; Distillation is useful when a task is narrow and runs at high volume, or when a model must run on local hardware. A plant might use a distilled model on site to classify machine alerts without sending data to the cloud.

E

Edge CaseDeployment & Ops: An unusual or extreme situation that a system was not mainly designed for and that can cause unexpected behaviour.; AI systems often handle typical cases well and stumble on rare ones, such as a handwritten amendment on a contract or a meter reading in the wrong unit. Collecting and testing edge cases is one of the most valuable things a team can do before and after launch.
EmbeddingRetrieval & Data: A list of numbers that represents the meaning of a piece of text, image or other data, so similar things end up with similar numbers.; Embeddings let computers compare meaning rather than exact words. They are what allow a search for "early repayment charges" to find a policy that talks about "prepayment fees".
Embedding ModelRetrieval & Data: An AI model whose job is to turn text or other content into embeddings that capture its meaning.; The choice of embedding model affects how well search understands your industry's language and abbreviations. Testing a few on your own documents is a quick way to improve retrieval quality.
Evaluation(Evals)Deployment & Ops: Systematically testing an AI system against agreed examples and criteria to measure how well it performs before and after changes.; Evaluation turns "it seems to work" into evidence that a risk committee or operations lead can rely on. A good test set is built from real cases, including the awkward ones, and is rerun whenever the model, prompts or data change.
Explainability(Explainable AI, XAI)Governance & Risk: The ability to understand and describe why an AI system produced a particular output or decision.; Explainability matters when people must justify decisions to customers, auditors or regulators, such as why a credit application was declined. Practical approaches include showing the sources used, the key factors considered and the reasoning steps.

F

Few-Shot PromptingPrompting: Giving an AI model a handful of worked examples in the prompt so it can copy the pattern for a new case.; A few good examples often teach a model a house style or format faster than long instructions. For instance, showing three correctly categorised customer complaints helps it categorise the fourth the same way.
Fine-TuningModels & Architecture: Further training an existing AI model on a smaller, specialised set of examples so it performs better on a particular task or style.; Fine-tuning can improve consistency and reduce costs for narrow, repeated tasks, but it needs good example data and ongoing upkeep. Many needs are met more simply by better prompts and retrieval, so it is worth trying those first.
Forward Deployed Engineering(FDE)Business & Engagement: A way of working where engineers embed inside a client's teams, build on real data and workflows, and stay involved after launch.; Sitting alongside the people who do the work means problems are understood first-hand and solutions fit how the business actually runs. Staying after go-live means the system keeps improving as edge cases appear, rather than being handed over and left to drift.
Foundation ModelModels & Architecture: A large, general-purpose AI model trained on broad data that can be adapted to many different tasks.; Most business AI today is built on top of a foundation model rather than trained from scratch. The skill lies in connecting it to the right data, instructions and controls for your specific use.
Frontier ModelModels & Architecture: One of the most advanced AI models available at a given time, typically from a small number of leading developers.; Frontier models are often best for complex reasoning and difficult documents, but they can be slower and more expensive. Many systems use a frontier model only for the hardest steps and smaller models for routine ones.
Function CallingAgents & Orchestration: A way for an AI model to request that a specific software action be run, such as looking up an account, with details filled in.; Function calling is how AI moves from writing text to doing things in other systems. The surrounding software decides whether to actually run each request, which is where permissions and safety checks belong.

G

Generative Adversarial Network(GAN)Models & Architecture: An AI approach where two models compete: one creates fake examples and the other tries to spot them, improving both over time.; GANs were an important early method for generating realistic images and are still used for some synthetic data tasks. They have largely been overtaken by diffusion models for image generation.
Generative AI(GenAI)Foundations: AI that creates new content, such as text, images, code or audio, based on patterns learned from existing examples.; Generative AI is well suited to drafting, summarising and transforming information, such as turning a long inspection report into a clear briefing. Because it generates rather than looks up, its output needs checks against trusted sources where accuracy matters.
GPU(Graphics Processing Unit)Deployment & Ops: A type of computer chip that performs many calculations at once, making it well suited to training and running AI models.; GPU availability and cost shape where and how AI can be run, especially for organisations that want to host models in their own data centres. For most early projects, renting capacity through a cloud provider avoids large upfront hardware decisions.
GraphRAGRetrieval & Data: A form of retrieval-augmented generation that uses a knowledge graph to find connected facts, not just similar passages of text.; It helps with questions that depend on relationships, such as which suppliers share a parent company or which assets are affected by a single substation fault. Building the graph takes extra effort, so it is best used where those connections really matter.
GroundingRetrieval & Data: Tying an AI model's answers to specific, trusted sources of information so its output reflects facts rather than guesswork.; Grounded answers can usually point to the document or record they came from, which makes them easier to check and trust. It is one of the most effective ways to reduce made-up answers in regulated settings.
GuardrailsGovernance & Risk: Rules and checks built around an AI system to stop it producing harmful, off-topic or non-compliant outputs or actions.; Guardrails can block certain topics, mask personal data, require approvals for risky actions or check outputs before they are sent. A customer-facing assistant at a bank, for example, might be prevented from giving personalised investment advice.

H

HallucinationGovernance & Risk: When an AI model states something false or invented with apparent confidence, such as a made-up figure, quote or reference.; Hallucinations happen because models generate plausible text rather than looking facts up. Grounding answers in trusted sources, asking for citations and keeping people involved in key decisions are the main defences.
Human-in-the-Loop(HITL)Governance & Risk: A design where people review, approve or correct AI outputs at key points before they take effect.; Human review keeps accountability where it belongs and catches mistakes before they reach customers or operations. The art is placing reviews where risk is highest, so people check the payment release or safety decision, not every routine step.
Hybrid SearchRetrieval & Data: Combining traditional keyword search with meaning-based search to get results that are both precise and relevant.; Keyword search is good at exact terms like part numbers or account codes, while meaning-based search catches different phrasings of the same idea. Using both usually gives the most reliable results over business documents.
HyperautomationAutomation & Process: An approach that combines several automation technologies, including AI, to automate as much of a process as is sensible.; Rather than automating single tasks in isolation, hyperautomation looks at whole processes and the tools needed for each step. The risk is chasing automation for its own sake, so it works best when tied to clear operational goals.
HyperparameterModels & Architecture: A setting chosen by people before or during training that controls how an AI model learns, such as how fast it adjusts.; Hyperparameters are tuned by engineers rather than learned from data. Business users rarely see them, but they affect how well a custom model performs and how long it takes to train.

I

InferenceDeployment & Ops: The act of running a trained AI model to produce an answer, prediction or piece of content from new input.; Training happens once or occasionally; inference happens every time the system is used, so it drives most ongoing running costs. Choosing the right model size for each task is the main lever for keeping inference costs in check.
Intelligent Automation(IA)Automation & Process: Automation that pairs rule-based software with AI, so processes can handle judgement calls and messy inputs as well as fixed steps.; Rules handle the predictable parts, while AI deals with reading documents, interpreting requests or spotting anomalies. A claims process might use rules to route cases and AI to read the supporting evidence.
Intelligent Document Processing(IDP)Automation & Process: Using AI to read documents such as invoices, contracts and forms, pull out the key information and pass it into business systems.; IDP copes with varied layouts, scans and handwriting far better than older template-based tools. Banks use it for know-your-customer documents and loan files; manufacturers use it for delivery notes, certificates and supplier invoices.
Intent RecognitionModels & Architecture: Working out what a person is trying to achieve from what they say or write, such as reporting a fault or disputing a bill.; Correctly identifying intent lets a system route requests to the right team or workflow straight away. A utility's contact centre might use it to separate outage reports from billing queries the moment a message arrives.

J

JailbreakGovernance & Risk: A deliberate attempt to trick an AI system into ignoring its rules and producing content or actions it is meant to refuse.; Jailbreaks often use role-play, unusual wording or long, confusing instructions. Any AI exposed to customers or the public should be tested against them and backed by controls that do not rely solely on the model behaving well.

K

Knowledge BaseRetrieval & Data: An organised collection of information, such as policies, procedures and FAQs, that people and AI systems can search for answers.; An AI assistant is only as good as the knowledge base behind it, so keeping content current and owned matters. Retiring outdated procedures, for example, stops the assistant from confidently quoting last year's process.
Knowledge CutoffModels & Architecture: The date after which an AI model has no built-in knowledge, because its training data stops there.; Anything that happened after the cutoff, such as a new regulation or tariff change, is unknown to the model unless you supply it. This is a key reason to connect AI to current internal and external sources.
Knowledge GraphRetrieval & Data: A way of storing information as things and the relationships between them, such as customers, accounts, assets and who owns what.; Knowledge graphs make connections explicit, which helps with questions that span many records. A bank might use one to trace beneficial ownership across companies; a utility to map which customers depend on which network assets.

L

Large Language Model(LLM)Foundations: An AI model trained on vast amounts of text that can understand and produce human language, from answering questions to drafting documents.; LLMs are the engine behind most modern AI assistants and document tools. They are flexible and fluent, but they need access to your data and sensible controls to be accurate and safe in business use.
LatencyDeployment & Ops: The delay between sending a request to an AI system and receiving its response.; Acceptable latency depends on the task: a live customer call needs answers in moments, while an overnight report can take minutes. Larger models and long inputs usually mean more delay, which affects design choices.
LLMOpsDeployment & Ops: The practices and tools for running language-model systems reliably in production, including testing, monitoring, versioning and cost control.; LLMOps covers the specific challenges of language models, such as managing prompts, tracking output quality and watching for unsafe responses. It is what keeps a system dependable after the launch excitement fades.
Long ContextModels & Architecture: The ability of an AI model to take in and work with very large amounts of text in a single request.; Long context makes it possible to analyse whole contracts or lengthy reports in one go. Models can still overlook details buried in the middle, so important tasks benefit from careful structuring and checks.

M

Machine Learning(ML)Foundations: A branch of AI where computers learn patterns from data to make predictions or decisions, instead of following hand-written rules.; Machine learning has long been used for tasks such as fraud detection, demand forecasting and predicting equipment failures. It works best where there is plenty of good historical data and a clear outcome to predict.
MemoryAgents & Orchestration: The ability of an AI system to retain and reuse information across steps or conversations, such as preferences, past decisions or case history.; Memory lets an agent pick up a case where it left off rather than starting again each time. It also raises questions about what should be remembered, for how long and who can see it, especially with personal data.
Mixture of Experts(MoE)Models & Architecture: A model design where many specialised sub-networks exist, but only a few are switched on for each piece of input.; This lets a model be very capable without using all of its capacity on every request, which keeps running costs down. For business users it is mostly invisible, but it helps explain why some large models are surprisingly fast.
MLOpsDeployment & Ops: The practices for building, deploying, monitoring and updating machine learning models reliably over their whole lifetime.; MLOps applies the discipline of software operations to models, so changes are tested, tracked and reversible. A credit risk or load forecasting model, for example, needs version control and monitoring just like any critical system.
Model Context Protocol(MCP)Agents & Orchestration: An open standard that gives AI applications a common way to connect to tools and data sources.; Instead of building a one-off connection for every system, teams can expose a tool once and let any compatible AI application use it. Access still needs to be controlled carefully, since a connection is also a door into your systems.
Model DriftDeployment & Ops: The gradual decline in an AI model's accuracy as the real world changes and drifts away from the data it was trained on.; Drift happens when customer behaviour, market conditions or equipment change over time. A fraud model trained before a new scam pattern appeared, for example, may start missing cases unless it is monitored and refreshed.
Multi-Agent SystemAgents & Orchestration: A setup where several AI agents, each with its own role, work together and pass tasks between them to complete a larger job.; Splitting work across specialised agents can make complex processes easier to build, test and control. A deal review might have one agent summarising financials, another checking legal terms and a coordinator combining their findings.
Multimodal AIModels & Architecture: AI that can work with more than one type of input or output, such as text, images, audio and video together.; Multimodal systems can read a scanned form with a diagram, or combine a technician's photo with their spoken notes. That makes them a good fit for field operations and document-heavy processes.

N

Named Entity Recognition(NER)Models & Architecture: Automatically spotting and labelling specific things in text, such as names, companies, dates, amounts or locations.; NER turns unstructured text into data that systems can use. A compliance team might use it to pull counterparty names and amounts from emails, or to find personal details that need masking.
Narrow AI(Weak AI)Foundations: AI designed to perform a specific task or a limited set of tasks, rather than general human-like intelligence.; All AI in practical use today is narrow, even when it seems broad. Recognising this helps set realistic expectations and design systems with clear boundaries.
Natural Language Processing(NLP)Foundations: The field of AI concerned with helping computers read, understand and produce human language.; NLP covers everything from classifying emails to translating documents and summarising reports. Language models have made it dramatically more capable and easier to apply to everyday business text.
Neural NetworkFoundations: A computing approach loosely inspired by the brain, made of layers of connected units that learn patterns by adjusting their connections.; Neural networks are the building blocks of modern AI, from image recognition to language models. Their strength is learning complex patterns; their weakness is that their reasoning can be hard to inspect directly.
No-Code / Low-CodeBusiness & Engagement: Tools that let people build applications or automations through visual interfaces and settings, with little or no programming.; These tools let business teams move quickly on simple workflows without waiting for developers. For critical or regulated processes, they still need proper testing, access controls and someone accountable for maintaining them.

O

ObservabilityDeployment & Ops: Being able to see what an AI system is doing in production: its inputs, outputs, decisions, errors, speed and costs.; Good observability lets teams trace exactly why a system gave a particular answer and spot problems before users report them. It is also the raw material for audits and for proving to regulators that controls work.
OntologyRetrieval & Data: A formal map of the key concepts in a business area and how they relate, giving people and systems a shared vocabulary.; An ontology settles questions like whether a "site", a "plant" and a "facility" mean the same thing. Agreeing these definitions helps AI combine data from different systems without confusion.
Open-Source Model(Open-Weight Model)Models & Architecture: An AI model whose trained internals are published so organisations can download, run and adapt it on their own infrastructure.; Running a model yourself gives more control over data, cost and availability, which appeals to regulated firms. In exchange, you take on the work of hosting, securing and updating it.
Optical Character Recognition(OCR)Automation & Process: Technology that converts text in images or scanned documents into machine-readable text.; OCR is usually the first step in digitising paper-heavy processes, such as old engineering drawings or signed loan agreements. On its own it captures the words but not their meaning, so it is often paired with AI to understand what the text says.
OrchestrationAgents & Orchestration: Coordinating multiple AI models, agents, tools and people so the steps of a process happen in the right order with the right information.; Orchestration handles sequencing, retries, approvals and error handling across a workflow. It is what keeps a multi-step process, such as month-end close or outage response, running predictably.
OverfittingModels & Architecture: When an AI model learns its training examples too closely, including their quirks, and then performs poorly on new data.; An overfitted model can look excellent in testing yet fail in real use. Testing on data the model has never seen is the standard way to catch it.

P

ParameterModels & Architecture: One of the internal numbers a model adjusts during training; together, parameters store what the model has learned.; Model size is often described by its number of parameters. More parameters can mean more capability, but also higher running costs, so bigger is not automatically better for a given task.
PII(Personally Identifiable Information, Personal Data)Governance & Risk: Any information that could identify a specific person, such as a name, address, account number or national insurance number.; AI systems that handle PII must follow data protection law and internal policy on storage, access and sharing. Common safeguards include masking personal details before they reach a model and keeping clear records of where data flows.
Pre-TrainingModels & Architecture: The first, large-scale stage of training where a model learns general patterns of language or images from a huge body of data.; Pre-training gives a model its broad knowledge and is enormously expensive, so very few organisations do it themselves. Businesses typically start from a pre-trained model and adapt it with their own data and instructions.
Predictive AIFoundations: AI that uses historical data to forecast outcomes or estimate likelihoods, such as demand, risk or the chance of equipment failure.; Predictive AI is well established in credit scoring, fraud detection, load forecasting and predictive maintenance. It complements generative AI: one estimates what is likely to happen, the other helps explain it and decide what to do.
Process MiningAutomation & Process: Analysing the digital trail left in business systems to see how a process really runs, including delays, loops and workarounds.; Process mining often reveals that the real process differs from the documented one. It helps pick the best candidates for automation, such as the approval step where invoices routinely sit for days.
PromptPrompting: The instructions, question or information you give an AI model to tell it what you want it to do.; Clear prompts with context, examples and a defined output format produce far better results than vague requests. In business systems, prompts are usually written and tested by the team rather than typed fresh by each user.
Prompt ChainingPrompting: Breaking a complex task into a sequence of smaller prompts, where each step's output feeds into the next.; Chaining makes each step simpler, easier to test and easier to fix when something goes wrong. A credit memo might be built by first extracting figures, then analysing trends, then drafting the narrative.
Prompt EngineeringPrompting: The practice of designing, testing and refining prompts so an AI model reliably produces the output you need.; Good prompt engineering is closer to writing a precise brief than to finding magic words. It involves trying variations against real examples and keeping the versions that perform best.
Prompt InjectionGovernance & Risk: An attack where hidden or malicious instructions in text, such as an email or web page, trick an AI system into doing something unintended.; The risk grows when AI agents read outside content and can take actions, for example an agent processing supplier emails that contain hidden commands. Defences include limiting what agents can do, separating trusted instructions from outside content and requiring approval for sensitive actions.
Prompt TemplatePrompting: A reusable prompt with blanks that are filled in automatically, such as customer name or document text, each time it runs.; Templates keep AI behaviour consistent across thousands of uses and make changes easy to manage and review. A bank might use one template for every complaint summary so that each follows the same structure.

Q

QuantizationModels & Architecture: Shrinking an AI model by storing its numbers with less precision, so it runs faster and on cheaper hardware with modest quality loss.; Quantization makes it practical to run capable models on local servers or devices at the edge, such as on a factory floor. The trade-off in accuracy should be tested on your own tasks before relying on it.
Question AnsweringModels & Architecture: AI that responds to questions in natural language, ideally drawing on specific documents or data to give accurate answers.; Question answering over internal documents is one of the most common first uses of AI in large organisations. An engineer might ask which torque settings apply to a particular pump model and get the answer with a link to the manual.

R

Rate LimitDeployment & Ops: A cap on how many requests a service, such as an AI model provider, will accept in a given period.; Rate limits affect how quickly large jobs can run and how a system behaves at busy times. Planning for them, through queuing, batching or agreed capacity, avoids slowdowns during critical periods such as month-end.
ReAct(Reason and Act)Agents & Orchestration: A pattern where an AI agent alternates between reasoning about what to do next and taking an action, then checks the result.; This think-act-observe loop helps agents tackle tasks that need several lookups or tools. It also leaves a readable trail of reasoning, which makes the agent's behaviour easier to review.
Reasoning ModelModels & Architecture: An AI model designed to spend extra effort thinking through a problem in steps before answering, improving results on complex tasks.; Reasoning models suit problems such as multi-step analysis, planning or checking calculations. They are usually slower and costlier per request, so they are best reserved for tasks where accuracy outweighs speed.
Red TeamingGovernance & Risk: Deliberately attacking or stress-testing an AI system, as an adversary would, to find weaknesses before real users or attackers do.; Red teams try to make the system leak data, break rules, give harmful advice or take unsafe actions. The findings feed directly into stronger guardrails and are valuable evidence for risk and compliance reviews.
Reinforcement Learning(RL)Foundations: A way of training AI by trial and error, where the system is rewarded for good outcomes and learns which actions lead to them.; Reinforcement learning suits problems involving sequences of decisions, such as optimising energy storage dispatch or production scheduling. It needs a safe environment, often a simulation, to learn in before touching real operations.
RerankingRetrieval & Data: A second pass over search results that reorders them so the most relevant items appear first before the AI uses them.; A first search casts a wide net quickly; reranking then looks more closely to pick the best matches. This often noticeably improves answer quality when searching large document collections.
Retrieval-Augmented Generation(RAG)Retrieval & Data: A technique where an AI system first looks up relevant information from trusted sources, then uses it to write its answer.; RAG lets AI use your current policies, contracts and records without retraining the model, and lets answers point back to their sources. It is the standard approach for internal assistants in banks, utilities and manufacturers.
RLHF(Reinforcement Learning from Human Feedback)Models & Architecture: A training method where people rate a model's answers and the model is adjusted to produce the kinds of answers people prefer.; RLHF is a big part of why modern assistants are helpful and polite rather than just fluent. It shapes general behaviour, but it does not give a model knowledge of your organisation's specific rules or data.
Robotic Process Automation(RPA)Automation & Process: Software bots that mimic a person's clicks and keystrokes to carry out repetitive, rule-based tasks across existing applications.; RPA is useful for moving data between older systems that lack modern connections. It can break when screens change and struggles with judgement, which is why it is increasingly paired with AI.

S

Semantic SearchRetrieval & Data: Search that finds results based on meaning rather than exact word matches, so differently worded content can still be found.; Semantic search helps staff find the right policy or procedure even when they do not know the official terminology. A search for "what to do if a transformer overheats" can surface a document titled "thermal event response".
Sentiment AnalysisRetrieval & Data: Using AI to judge the emotional tone of text, such as whether a customer message is positive, negative or frustrated.; Sentiment analysis helps prioritise complaints and spot emerging issues across large volumes of feedback. A retail bank might use it to flag distressed customers who may be in financial difficulty for faster, more careful handling.
Service Level Agreement(SLA)Business & Engagement: A formal commitment between a provider and a client on the level of service, such as availability, response times and support.; For AI systems that support critical operations, SLAs should cover not just uptime but how quickly issues are fixed and how quality is maintained. They give both sides a clear basis for accountability.
Shadow AIGovernance & Risk: Staff using AI tools at work without the organisation's knowledge or approval, often by pasting company data into public services.; Shadow AI usually signals real demand that approved tools are not meeting. The best response combines clear policy with safe, sanctioned alternatives that are genuinely easy to use.
Small Language Model(SLM)Models & Architecture: A compact language model that is cheaper and faster to run, often good enough for focused tasks and able to run on local hardware.; Small models suit high-volume, well-defined jobs such as classifying tickets or extracting fields, and can run inside your own environment. They are less capable at open-ended reasoning, so many systems mix small and large models.
Structured OutputAgents & Orchestration: AI responses produced in a fixed, machine-readable format, such as a form with defined fields, so other systems can use them directly.; Structured output is what lets AI feed straight into databases and workflows without someone copying and pasting. An invoice reader, for example, might return supplier, amount, date and purchase order number in the same layout every time.
Supervised LearningFoundations: Training a model on examples where the correct answer is already known, so it learns to predict that answer for new cases.; Most traditional business models, such as credit risk scores or defect classifiers, are trained this way. The quality and consistency of the labelled examples largely determine how well the model performs.
Synthetic DataRetrieval & Data: Artificially generated data that mimics the patterns of real data, used for training or testing when real data is scarce or sensitive.; Synthetic data can help test systems without exposing real customer records, or fill gaps such as rare fraud patterns or unusual equipment faults. It must be checked carefully, since unrealistic synthetic data can teach a model the wrong lessons.
System PromptPrompting: Background instructions given to an AI model that set its role, tone, rules and limits for every conversation or task.; The system prompt is where much of an assistant's behaviour is defined, such as which topics to avoid and when to escalate. It should be version-controlled and tested like any other important configuration.

T

TemperatureModels & Architecture: A setting that controls how varied or predictable an AI model's responses are; lower is more consistent, higher is more creative.; Business tasks that need consistent, repeatable answers, such as data extraction, usually run at low temperature. Brainstorming or drafting marketing ideas can benefit from a higher setting.
TokenModels & Architecture: A small chunk of text, often part of a word, that AI models read and write; usage and costs are typically measured in tokens.; Knowing roughly how many tokens a task uses helps estimate running costs and whether content fits in the model's context window. Long documents and lengthy conversations consume more tokens and therefore cost more.
TokenizationModels & Architecture: The process of breaking text into tokens so an AI model can process it.; Tokenization explains some odd model behaviour, such as difficulty counting letters or handling unusual codes and part numbers. It can also mean the same content costs more in some languages than others.
Tool UseAgents & Orchestration: An AI model's ability to use external tools, such as search, calculators, databases or business systems, to complete a task.; Tools let AI get accurate figures and take real actions instead of relying on memory. Each tool should come with clear limits on what it can access or change, matched to the risk involved.
Training DataRetrieval & Data: The examples an AI model learns from, which shape what it knows, how it behaves and where its blind spots are.; Gaps or errors in training data show up as gaps or errors in the model. For custom models, using representative data from your own operations matters more than simply having a lot of it.
Transfer LearningModels & Architecture: Reusing what a model learned on one task as a starting point for a related task, saving data and training time.; Transfer learning is why a model trained on general images can be adapted to spot specific weld defects with relatively few examples. It makes custom AI practical for organisations without vast datasets.
Transformer Model(Transformer)Models & Architecture: The model design behind most modern AI, which uses attention to process whole sequences of text or data at once.; Transformers made today's language models possible by handling long passages and relationships between distant words. Most AI tools a business uses now, from chat assistants to document readers, are built on them.

U

UnderfittingModels & Architecture: When an AI model is too simple or poorly trained to capture the real patterns in the data, so it performs badly everywhere.; Underfitting shows up as poor results in both testing and real use. The fix is usually a more suitable model, better input data or more training, rather than more fine adjustment.
Unstructured DataRetrieval & Data: Information that does not sit neatly in tables, such as emails, PDFs, call recordings, images and free-text notes.; Most organisational knowledge lives in unstructured data, and until recently it was hard for software to use. Modern AI makes it possible to search, summarise and extract facts from it, such as reading years of engineer notes to spot recurring faults.
Unsupervised LearningFoundations: Training a model to find patterns or groupings in data without being told the correct answers in advance.; It is useful for discovering structure you did not know was there, such as customer segments or unusual transactions that do not fit normal patterns. People still need to interpret whether the patterns found are meaningful.

V

VectorRetrieval & Data: An ordered list of numbers; in AI, vectors are used to represent the meaning of text, images or other data.; Turning content into vectors lets computers measure how similar two things are by comparing their numbers. This is the basis of meaning-based search and recommendation.
Vector DatabaseRetrieval & Data: A database built to store vectors and quickly find the ones most similar to a query, powering meaning-based search.; Vector databases sit behind many AI assistants that answer questions from company documents. Like any data store, they need access controls, so people only retrieve content they are permitted to see.
Vertical AIBusiness & Engagement: AI built for a specific industry, with its terminology, regulations and workflows designed in from the start.; Vertical AI can deliver value faster because it already understands sector-specific needs, such as trade finance documents or grid operations. The trade-off is less flexibility outside that industry's common patterns.
Voice AIAgents & Orchestration: AI that understands spoken language and responds with natural-sounding speech, enabling phone and voice-based interactions.; Voice AI can handle routine calls such as meter reading submissions or balance enquiries, freeing staff for complex conversations. It needs fast responses, clear disclosure that callers are speaking to AI, and easy transfer to a person.

W

Workflow AutomationAutomation & Process: Using software to move tasks, information and approvals between people and systems automatically according to defined steps.; Workflow automation removes chasing, re-keying and waiting from routine processes such as change requests or purchase approvals. Adding AI lets these workflows handle steps that need reading or judgement, not just routing.

Z

Zero-Shot PromptingPrompting: Asking an AI model to do a task with instructions alone, without giving it any worked examples.; Modern models handle many tasks well zero-shot, which makes it the quickest place to start. If results are inconsistent, adding a few examples usually helps.