AI glossary: key terms explained

The language around AI has moved faster than most people have had time to keep up with, and a lot of it is used loosely.

This glossary is a plain English reference to the terms a UK business actually runs into when it looks at AI and automation, written so a non technical manager can read any entry aloud and understand it.

Each term has a short, self contained definition and a note on what it means in practice, grouped by theme rather than listed A to Z so related ideas sit together. Where a term deserves a fuller treatment, the entry links to a deeper guide.

Use it however suits you: read it through once to build a mental map, or jump to the single term that sent you here.

The terms are grouped into four clusters, the foundations, the language model vocabulary, the agent vocabulary, and the automation and governance terms, moving from the general to the specific. Nothing here assumes prior knowledge, and nothing is defined in terms that need their own glossary.

How this AI glossary is organised

Treat this as a reference to return to rather than something to memorise. When a term comes up in a proposal, a tool's marketing, or a conversation with a supplier, look it up here for a plain definition, then follow the link if you need the fuller picture.

If you are early in learning about AI, the foundation and language model clusters are the ones worth reading first, since almost everything commercial is built on them.

For definitions from a neutral, high trust UK source, the UK Parliament's Parliamentary Office of Science and Technology maintains its own artificial intelligence glossary, which is a sensible cross reference.

And when a term stops being abstract and becomes a decision, whether a particular AI or automation is worth adopting for your own work, that is the point to move from vocabulary to judgement.

Plain definitions, written for someone deciding rather than building Updated July 2026

A

Agentic loopAI agent terms: The agentic loop is the repeating cycle an agent runs through: plan a step, act through a tool, observe the result, then decide the next step. This loop is what lets an agent handle a task that cannot be done in one shot. Understanding it explains both why agents can tackle complex jobs and why they can go wrong if a step is misjudged.
AI agentAI agent terms: An AI agent is a system that uses a model to pursue a goal over multiple steps, deciding what to do, taking actions through tools, and reacting to the results, rather than answering a single prompt. It is the difference between a tool that responds and one that carries out a task. Our complete guide to AI agents covers the types and where they earn their place.
AI governanceAutomation and governance terms: AI governance is the set of policies and controls a business puts around its use of AI: who may use which tools, what data they may use, how output is checked, and how it stays compliant with rules such as UK GDPR. For any organisation handling personal or sensitive data, governance is not optional; it is what makes AI use defensible as well as useful.
AlgorithmFoundation terms: An algorithm is a fixed set of steps for turning an input into an output. Ordinary software is written as algorithms a person specified in advance, which is why it does exactly what it was told and nothing else. The distinction matters because machine learning replaces some of those hand written steps with patterns learned from data.
API integrationAutomation and governance terms: An API integration is a connection that lets two software systems exchange data or trigger actions directly, through a defined interface, rather than a person copying between them. APIs are the plumbing of automation. When systems connect cleanly this way, a workflow can run end to end without manual handoffs, which is usually more robust than automation that imitates a human clicking around.
Artificial intelligence (AI)Foundation terms: Artificial intelligence is the broad field of building computer systems that perform tasks normally requiring human intelligence, such as understanding language, recognising patterns, or making decisions. It is an umbrella term, not a single technology, covering everything from a spam filter to a chatbot. Our guide to what artificial intelligence is covers the wider picture.
Autonomous agentAI agent terms: An autonomous agent is an AI agent that operates with limited human intervention, making decisions and taking actions towards a goal on its own. Full autonomy is rarer and riskier in practice than the label suggests; most reliable business deployments keep a human checkpoint. The useful question is not whether an agent is autonomous but how much oversight a given task genuinely needs.

B

BenchmarkEvaluation and risk terms: A benchmark is a standardised test used to compare models against each other. Benchmarks are useful for narrowing a shortlist and misleading as a promise, because they measure general capability rather than performance on your documents and your process. Your own evaluation set beats any published league table.
BiasEvaluation and risk terms: Bias is systematic unfairness in a system's outputs, inherited from patterns in the data it learned from. It matters most in decisions affecting people, such as screening applicants, where a model can reproduce historic discrimination at speed and scale. Where a decision affects someone's rights or livelihood, a person needs to own it.

C

Chain of thoughtLanguage model terms: Chain of thought is prompting a model to work through its reasoning step by step before answering. It measurably improves results on anything involving several stages of logic. The reasoning it shows is a useful audit trail, though it should be read as the model's account of its answer rather than proof of how it arrived there.
ChunkingData and retrieval terms: Chunking is splitting long documents into passages small enough to retrieve and feed to a model. How you split matters more than people expect: cut a policy mid clause and the retrieved passage answers the question wrongly while looking perfectly confident. Poor chunking is a common and quiet cause of bad answers.
Computer visionFoundation terms: Computer vision is machine interpretation of images and video. In day to day business automation it usually turns up as document work: reading a scanned invoice, checking a photograph was taken where it claims, spotting that a form has been signed. It is mature technology and considerably more reliable on documents than on open ended scenes.
Confidence scoreRunning AI in a business: A confidence score is a number a system attaches to its own output indicating how sure it is. Used well it is the basis of a sensible escalation rule: handle the confident cases automatically, send the rest to a person. Treat it as a useful signal rather than a guarantee, since a model can be confidently wrong.
Context windowLanguage model terms: The context window is the amount of text, measured in tokens, a model can consider at once, covering both your input and its reply. If a document exceeds the window, the model cannot see all of it in one pass. Larger context windows let a model work with longer documents, which matters for tasks like summarising a lengthy report or contract.

D

Data pipelineAutomation and governance terms: A data pipeline is an automated flow that moves data from where it is created to where it is used, cleaning and reshaping it along the way. Reliable AI and reporting both depend on one. Much of the unglamorous work in a real automation project is getting the pipeline right, because a model or a dashboard is only as good as the data feeding it.
Data residencyRunning AI in a business: Data residency is the question of which country your data physically sits in and which laws therefore apply to it. For UK businesses handling personal data it drives real decisions about which providers and regions are acceptable. Settle it before a build starts, because it constrains the architecture rather than decorating it.
Deep learningFoundation terms: Deep learning is a type of machine learning that uses layered networks, loosely inspired by the brain, to learn complex patterns from very large amounts of data. The extra layers let it handle messy inputs like images, audio, and natural language far better than earlier methods. It is the approach behind most recent breakthroughs, including the language models businesses now use daily.

E

EmbeddingData and retrieval terms: An embedding is a list of numbers standing for the meaning of a piece of text, produced by a model so that similar meanings end up as similar lists. It is what makes searching by meaning possible rather than searching by exact words. Everything described as semantic search or retrieval rests on embeddings underneath.
EvaluationEvaluation and risk terms: Evaluation, often shortened to evals, is a fixed set of test cases with known good answers, run against a system to measure whether it works and to catch it getting worse. Without it, "the AI seems better this week" is an opinion. Any automation you intend to trust in production needs one, and it is the first thing missing from most projects.
ExplainabilityEvaluation and risk terms: Explainability is how far you can account for why a system produced a particular answer. Rule based automation is fully explainable by construction; models much less so. Where you are regulated, or simply need to defend a decision, favour designs that show their sources and keep the consequential judgements with a person.

F

Fine tuningLanguage model terms: Fine tuning is further training of an existing model on a specific set of examples so it adapts to a particular style, domain, or task. It is more involved and costly than prompting or retrieval, and for most business needs it is not the first tool to reach for. It matters most when you need a model to consistently match a very specific tone or specialised vocabulary.
Foundation modelLanguage model terms: A foundation model is a large, general purpose model trained on broad data that can be adapted to many tasks, rather than built for one. The large language models behind mainstream AI tools are foundation models. The term matters because it explains why the same underlying model can power a writing assistant, a coding tool, and a customer chatbot with only light adaptation.

G

Generative AILanguage model terms: Generative AI is AI that creates new content, text, images, audio, or code, rather than only classifying or predicting from existing data. It is the category behind writing assistants and image tools. Our guide to what generative AI is covers how it differs from the predictive AI that came before it, which is a distinction worth holding.
GroundingLanguage model terms: Grounding means tying a model's output to a reliable source of truth, such as your own data, so the answer can be traced and trusted rather than freely generated. Retrieval augmented generation is one way to achieve it. Grounding is what separates an assistant that guesses from one that answers from your actual records, which is usually the difference that makes it usable.
GuardrailsAI agent terms: Guardrails are the rules and limits placed around an AI system to keep its behaviour safe and on task, such as what it may access, what it must not do, and when it should hand off to a human. They are essential for any business deployment, because they turn a capable but unpredictable model into something you can trust to run against real data and real customers.

H

HallucinationLanguage model terms: A hallucination is when a model produces information that sounds confident and plausible but is wrong or invented. It happens because the model generates likely text, not verified fact. This is the single most important limit to understand: AI output on anything factual needs checking, and workflows that rely on it should be designed with that in mind rather than assuming accuracy.

I

InferenceFoundation terms: Inference is the act of running a trained model to get an answer, as opposed to training it in the first place. Training happens once and costs a great deal; inference happens on every request and is what you actually pay for in production. When someone talks about the running cost of an AI feature, they mean inference.

K

Knowledge cutoffLanguage model terms: The knowledge cutoff is the date after which a model has seen nothing. Ask about events past it and the model either says so or, worse, invents something plausible. It is the single clearest argument for grounding a system in your own current documents rather than relying on what the model happens to remember.

L

Large language model (LLM)Language model terms: A large language model is an AI system trained on vast amounts of text to predict and generate language, which lets it draft, summarise, answer, and translate. It does not "know" facts the way a database does; it produces the most likely continuation of the text it is given. Understanding that single point explains most of both its power and its limits.
LatencyRunning AI in a business: Latency is how long a system takes to respond. It rarely matters for work running overnight and matters enormously for anything a customer waits on, such as a voice agent. It is worth fixing as a requirement early, because the cheapest and most capable options tend to sit at opposite ends of it.

M

Machine learning (ML)Foundation terms: Machine learning is a branch of AI where a system learns patterns from data rather than following rules written by hand. Instead of a programmer specifying every step, you show the system many examples and it infers the pattern. Most modern AI you will encounter, from fraud detection to recommendation engines, is machine learning underneath.
ModelFoundation terms: A model is the file that results from training: the learned patterns, stored as numbers, that let a system make predictions about new inputs. Talking about "the AI" usually means talking about a model plus the software around it. Models can be swapped, which is why a well built system is not tied to one supplier.
Model context protocol (MCP)Language model terms: The model context protocol is an open standard for connecting AI models to external tools and data sources through a common interface, rather than building a custom bridge for each one. It matters because it is becoming the standard way assistants securely reach into the systems a business already runs, which is what turns a chat tool into something that can actually do work.
Model driftEvaluation and risk terms: Model drift is the slow decay in a system's accuracy as the world moves away from the data it learned on. Nothing breaks visibly; results simply get quietly worse. It is the reason an automation needs monitoring and occasional retraining rather than being installed and forgotten.
Multi agent systemAI agent terms: A multi agent system uses several specialised agents that coordinate, each handling part of a larger task, rather than one agent doing everything. One might gather data, another draft, another check. It mirrors how a team divides work. For complex processes it can be more reliable than a single agent, though it also adds coordination and cost worth weighing.
MultimodalLanguage model terms: Multimodal describes a model that can handle more than one type of input or output, such as text, images, and audio together. A multimodal assistant might read a photo of an invoice and answer questions about it. For business use it widens what a single tool can do, from reading scanned documents to describing an image, without stitching several systems together.

N

Natural language processing (NLP)Foundation terms: Natural language processing is the broad field concerned with getting computers to work with human language: reading it, categorising it, extracting from it, producing it. Language models are the current state of that field rather than a separate thing. If a task involves text that a person wrote, it is an NLP task.
Neural networkFoundation terms: A neural network is the structure at the heart of deep learning: layers of simple connected units that each pass a signal forward, adjusting their strength as the system learns. The name is a loose analogy to brain cells, not a literal copy. In practice it is the mathematical machinery that lets a model turn raw data into a useful prediction or output.
No code and low codeAutomation and governance terms: No code and low code are tools that let people build automations and applications through visual interfaces rather than writing software, with low code allowing some custom code where needed. They lower the barrier to building useful automation. They suit standard, well supported tasks; genuinely bespoke or high stakes workflows often outgrow them and need a custom build.

O

Open weight modelRunning AI in a business: An open weight model is one whose trained parameters are published, so it can be downloaded and run on your own hardware rather than called as a service. The appeal is control and data residency; the cost is that you now operate infrastructure. It is worth considering when data cannot leave your estate.
OrchestrationAI agent terms: Orchestration is the coordination layer that decides which agent, tool, or step runs when, and passes information between them so a multi step process holds together. It is the difference between a set of capable parts and a working system. In practice, good orchestration is where much of the reliability of an automation actually lives.
OverfittingEvaluation and risk terms: Overfitting is when a model learns the peculiarities of its training data so closely that it performs worse on anything new. It is the classic failure of a system that tested beautifully and disappointed in production. The guard against it is holding back data the model never saw during training and testing on that.

P

ParameterFoundation terms: A parameter is one of the numbers inside a model that training adjusts. Modern language models have billions of them, and the count is often quoted as a rough proxy for capability. Treat it as a loose signal only, because a smaller, better trained model frequently beats a larger, sloppier one on real work.
PromptLanguage model terms: A prompt is the instruction you give an AI model to get an output. The quality of the prompt heavily shapes the quality of the result: a vague prompt yields a vague answer, a specific one yields a usable answer. Writing prompts well is a learnable skill, and for most business tasks it is the single highest leverage thing you can improve.
Prompt injectionEvaluation and risk terms: Prompt injection is an attack where instructions hidden inside content the model reads, such as a web page or an incoming email, hijack what it does next. It matters the moment an agent can both read untrusted material and take actions. The defence is limiting what the agent is permitted to do, not hoping the model spots the trick.

R

Reinforcement learningFoundation terms: Reinforcement learning trains a system by rewarding good outcomes and penalising bad ones, so behaviour improves through trial rather than through labelled examples. It is how models are tuned to be helpful and to follow instructions. In business automation it appears rarely and indirectly, mostly baked into the models you buy rather than something you would run yourself.
Retrieval augmented generation (RAG)Language model terms: Retrieval augmented generation is a technique where a model looks up relevant information from your own documents or data before answering, so its response is grounded in your sources rather than only its training. It is the standard way to make an AI assistant answer accurately from a company's own knowledge, and it reduces hallucination on questions about your specific business.
Robotic process automation (RPA)Automation and governance terms: Robotic process automation is software that mimics the clicks and keystrokes a person would make to move data between systems, following fixed rules. It works well for stable, repetitive tasks but breaks when a screen or format changes. It is rules based, not intelligent, which is why pairing it with AI, or replacing it, is a common step as processes grow more varied.

S

Semantic searchData and retrieval terms: Semantic search finds material by meaning instead of by matching keywords, so a question phrased in your own words still finds the passage that answers it. It is the difference between needing to know the right search term and simply asking. This is usually the single most visible improvement when a company puts its own documents behind AI.
Structured dataFoundation terms: Structured data is anything organised into predictable fields, such as a database table or a spreadsheet with consistent columns. It is the easy case: conventional software has handled it well for decades and you rarely need AI to work with it. Knowing whether your problem is structured is the fastest way to tell whether you need AI at all.
Supervised learningFoundation terms: Supervised learning trains a model on examples that have already been labelled with the right answer, so it learns to reproduce that judgement on new cases. It is the approach behind most practical classification work, such as sorting incoming documents by type. Its cost is the labelling: someone has to produce the correct answers first.
Synthetic dataRunning AI in a business: Synthetic data is data generated by a model rather than collected from the real world, used to train or test a system where real examples are scarce or too sensitive to use. It is genuinely useful for testing edge cases you cannot otherwise obtain. It is not a substitute for validating against real cases before launch.

T

TemperatureLanguage model terms: Temperature is a setting that controls how much randomness a model uses when choosing its next word. Low values give consistent, predictable output; higher values give more varied phrasing. For business automation you generally want it low, because reproducibility matters more than flair when the output feeds a process.
TokenLanguage model terms: A token is the unit a language model reads and generates, roughly a word or part of a word. Models measure input and output length in tokens, and usage is often priced per token. The practical upshot is that longer prompts and longer answers cost more and take longer, which is worth knowing when you scale a tool across a team.
Tool useLanguage model terms: Tool use describes a model calling an external function, such as searching a database, sending an email, or running a calculation, rather than only generating text. It is what lets an AI act rather than just answer. Tool use is the mechanism underneath most useful automation, because it connects the model's language ability to the systems where real work happens.
Training dataFoundation terms: Training data is the collection of examples a machine learning system learns from. Its quality and coverage largely decide how good the resulting model is: narrow or biased training data produces a narrow or biased model. This is why "rubbish in, rubbish out" holds for AI, and why data quality matters more than almost any other factor in a real project.
TransformerFoundation terms: The transformer is the design underneath almost every current language model. Its useful property is attention: when processing any part of the input it can weigh every other part, which is how it keeps track of meaning across a long document. You never interact with it directly, but it explains why these models handle context so much better than what came before.

U

Unstructured dataFoundation terms: Unstructured data is everything without that fixed shape: emails, contracts, call recordings, photographs, notes. Most of what a business actually holds is unstructured, and it is precisely where conventional automation stalls and where language models earn their place. If your bottleneck is somebody reading things, this is why.
Unsupervised learningFoundation terms: Unsupervised learning finds structure in data that carries no labels, typically by grouping similar records together. It is useful when you suspect there are patterns in your data but cannot say in advance what they are. The output needs interpreting by a person, because the system can tell you what clusters exist but not what they mean.

V

Vector databaseData and retrieval terms: A vector database stores embeddings and answers the question "what is nearest to this?" quickly, across millions of records. It is the retrieval half of a retrieval augmented system. For most businesses it is infrastructure rather than a decision: something your build sits on, not a product you evaluate separately.

W

Workflow automationAutomation and governance terms: Workflow automation is the use of software to carry out a multi step business process, moving information between systems and taking actions with little or no manual effort. It predates AI and often does not need it. The recent shift is that AI can now handle the judgement steps that once forced a person into the loop, widening what can be automated.

Z

Zero shot and few shotLanguage model terms: Zero shot means asking a model to do something with no examples; few shot means including a handful of worked examples in the prompt first. Few shot is usually the cheapest reliable improvement available, because showing the shape of a good answer beats describing it. It is the first thing to try before anyone suggests fine tuning.

Questions about AI terminology

An AI glossary is a reference that defines the key terms used in artificial intelligence and automation in plain language, so a non specialist can understand them. This one is grouped by theme, from foundation concepts to agents and automation, with each term written as a short, self contained definition and linked to a fuller guide where useful.

Because AI vocabulary is used loosely and often assumes prior knowledge, which makes it easy to nod along without a firm grasp of terms you are being sold on. A clear glossary lets you check what a word actually means, compare suppliers on equal footing, and make decisions from understanding rather than from marketing language.

Use it to check any term you meet in a proposal, tool, or supplier conversation before acting on it, and to build a shared vocabulary across a team so everyone means the same thing. Reading the foundation and language model terms first gives you the base to evaluate almost any AI tool or claim with confidence.

For most businesses the essentials are large language model, prompt, hallucination, and retrieval augmented generation on the language side, and AI agent, workflow automation, and guardrails on the action side. Together these cover what the common tools do, where they fail, and what it takes to run them safely against your own data.

Artificial intelligence is the broad field of systems that perform tasks needing human intelligence. Machine learning is a branch of it where systems learn patterns from data rather than following hand written rules. All machine learning is AI, but not all AI is machine learning. In practice, most modern AI you meet is machine learning underneath.

A support tool using retrieval augmented generation answers from your own help documents; an AI agent runs an agentic loop to reconcile orders across systems; guardrails stop it committing to a refund. Each abstract term maps to a concrete behaviour, which is why knowing the vocabulary helps you judge what a tool will actually do.

You can evaluate tools and suppliers on their merits rather than on marketing, brief your team with a shared language, and spot when a claim is vaguer than it sounds. Understanding the vocabulary also reveals the limits, such as hallucination, that decide whether a given use of AI is safe to rely on.

The glossary spans four clusters: foundation concepts such as machine learning and neural networks; language model terms such as LLMs, prompts, and retrieval augmented generation; agent terms such as AI agents, agentic loops, and orchestration; and automation and governance terms such as RPA, API integration, and AI governance. Together they cover the vocabulary a UK business actually meets.

Reading an AI product page with these terms

The value of a shared vocabulary is that it turns AI from something that happens to your business into something you can assess and direct.

Once the terms are clear, the useful questions become practical ones: which of these tools or techniques would genuinely help the work your team does, and which are a solution looking for a problem. That judgement is specific to your processes and your data, not to the definitions.

An automation audit is where we help with it: we look at the real work your team repeats, say plainly which of these ideas would earn their place and which would not, and design anything worth building around the systems you already run.

When a build is the right answer, our workflow agents service is how we deliver it and keep it dependable over time.

To go deeper on the terms here, our complete guide to AI agents covers the agent cluster, the automation glossary covers the automation terms, and the AI agent glossary covers the agentic vocabulary in full.

The other glossaries

AI agent glossary Open → Business automation glossary Open → All resources Guides →