Most AI conversations get lost in vocabulary before they get anywhere useful. You do not need to code or read a research paper to make good decisions about AI, but you do need to know what the words mean when a vendor, a consultant or your own team uses them. This glossary gives you plain-English definitions of the terms that come up most often in mid-market AI work, grouped so you can find them quickly. Read it end to end, or dip in when a term trips you up in a meeting.
Artificial intelligence (AI). Software that performs tasks we used to think needed human judgement, such as understanding language, spotting patterns or making predictions. It is a broad umbrella term, not a single technology.
Narrow AI. AI built to do one specific job well, such as flagging fraudulent transactions or transcribing calls. Almost every AI in business use today is narrow AI, not the general-purpose intelligence of science fiction.
Machine learning. A method where software learns patterns from examples rather than following rules a person wrote by hand. Show it enough past invoices and it learns to predict which ones will be paid late.
Generative AI. AI that produces new content, such as text, images, code or audio, rather than just classifying or predicting. The tools behind ChatGPT and similar products are generative AI.
Large language model (LLM). A type of generative AI trained on huge amounts of text so it can understand and produce human language. It powers chatbots, drafting tools and summarisers, and predicts likely words rather than looking up facts.
Agentic AI. AI that can take a goal, plan the steps and carry out actions across tools, not just answer a single question. An agent might read an email, check your system and draft a reply on its own.
Algorithm. A set of instructions for solving a problem or making a decision. In machine learning, the algorithm is the recipe the software follows to learn patterns from data.
Parameter. An internal setting the model adjusts as it learns. Large models have billions of them, and the count is a rough guide to how much a model can capture, though bigger is not always better for your use.
Training data. The examples a model learns from. Its quality and coverage shape what the model can do and where it falls short, so poor training data means a poor model no matter how clever the method.
Training. The process of feeding data to a model so it adjusts its parameters and improves at a task. This is the expensive, compute-heavy stage that happens before you ever use the model.
Fine-tuning. Taking a model that has already been trained and adjusting it further on your own narrower data, so it fits your language, domain or task more closely. It is cheaper than training from scratch.
Hallucination. When a model produces something that sounds confident and plausible but is simply wrong or made up. LLMs predict likely words, so they can invent facts, figures or sources, which is why human checking still matters.
Prompt engineering. Writing and refining the instructions you give a model to get better, more reliable results. Clear context, examples and constraints usually matter more than clever wording.
Data quality. How accurate, complete, current and consistent your data is. AI trained or run on weak data produces weak results, so this is usually the first thing to fix.
Data bias. When your data over-represents or under-represents certain groups or situations, so the model learns a skewed view of the world. Biased data leads to biased decisions, often without anyone noticing.
Data privacy. Handling personal and confidential data lawfully and safely, including what you feed into AI tools. Under UK GDPR, putting customer or staff data into a public AI tool can be a breach.
Data lineage. A record of where a piece of data came from, how it has moved and how it has changed along the way. It matters when you need to prove to an auditor or regulator that a number is trustworthy.
Accuracy. How often the model gets the answer right. It is a useful headline number but can mislead, because a model can score high overall and still fail badly on the cases you care about.
Bias. A systematic tendency in a model to get things wrong in one direction, often against a particular group. It can come from the data, the design or how the model is used, and it carries legal and reputational risk.
False positive and false negative. A false positive flags something that is not really there, such as blocking a genuine payment as fraud. A false negative misses something that is there, such as letting real fraud through. The two errors have different costs, and you usually have to trade one against the other.
Robustness. How well a model holds up when the real world differs from its training data, such as unusual inputs or messy edge cases. A robust model degrades gracefully rather than failing without warning.
Interpretability. How easily a person can understand why a model produced a given output. High interpretability makes it easier to trust, debug and defend a model’s decisions.
Production. AI that is live and doing real work in your business, as opposed to a pilot or experiment. The jump from a promising pilot to reliable production is where most initiatives stall.
Inference. The moment a trained model is actually used to produce an answer, such as generating a reply or scoring a lead. Every inference costs compute, so at scale it costs real money.
Latency. The delay between asking a model for an answer and getting one. Low latency matters for anything customer-facing, where a slow response is as good as no response.
Scalability. Whether a system keeps working, at acceptable cost and speed, as usage grows from a handful of users to thousands. Something that runs fine in a pilot can buckle under real volume.
Governance. The rules, roles and controls that decide how AI is chosen, approved, used and monitored in your organisation. It is the single biggest predictor of whether AI pays off, which we return to below.
Explainability. Being able to give a clear, human account of how an AI system reached a decision. Regulators, auditors and affected customers increasingly expect it, especially for decisions about people.
Fairness. Making sure an AI system does not treat people or groups unjustly, for example in hiring, lending or pricing. Fairness has to be checked deliberately, because a model will not correct its own bias.
Transparency. Being open about where and how you use AI, what data it relies on and what its limits are, with both staff and customers. It is the difference between AI people trust and AI they quietly work around.
Accountability. Naming who is responsible when an AI system gets something wrong. Without a clear owner, problems fall between teams and no one fixes them.
Risk management. Identifying what could go wrong with an AI system, judging how serious and likely it is, and putting controls in place before it goes live. It covers security, compliance, accuracy and reputation.
Multi-modal AI. AI that works across more than one type of input or output at once, such as text, images, audio and video together. You might show it a photo and ask a question about it in words.
Few-shot learning. Giving a model a small number of examples inside your prompt so it picks up a task without any retraining. It is a fast way to steer a general model towards your specific job.
Retrieval-augmented generation (RAG). A technique that lets a model pull in your own documents or data at the moment of answering, so it responds from your facts rather than its general training. It reduces hallucination and is a common way to build AI on private company knowledge.
Transfer learning. Reusing a model trained on one task as the starting point for a related one, rather than building from nothing. It saves time, data and cost, and underpins most practical AI today.
Agentic workflows. Chains of steps an AI agent carries out to complete a larger job, often calling several tools and making decisions along the way. The value tends to come from joining systems together, not from any single clever step.
Knowing the vocabulary is not academic. In aibl’s State of UK AI Adoption Survey 2026 (755 UK mid-market leaders), 49.6 per cent report a measurable return on AI today, but only 14 per cent have scaled it across three or more functions. The gap between trying AI and running it properly is where leaders need to be sharpest, and that starts with understanding what you are being sold and what you are signing off.
One term earns its place above the rest. In the survey, measurable ROI climbs from 22.2 per cent in companies with no governance to 85.3 per cent in those with mature, embedded governance, on much the same tools. So when you read this glossary, pay closest attention to the governance section. The words that decide whether AI pays off are less about the technology and more about how you control it. Pick two or three terms your team currently fudges, agree what they mean in your business, and use them consistently.
AI is the broad umbrella for software that does things we associate with human judgement. Machine learning is one way of building it, where the software learns patterns from data instead of following hand-written rules. All machine learning is AI, but not all AI is machine learning.
No. You can use these tools well without knowing the maths, in the same way you drive without understanding the engine. What you do need is to know their limits: they can sound confident and still be wrong, so anything that matters should be checked before it goes out.
Generative AI is the wider category of tools that produce new content, including images, audio and code. A large language model is the kind that specialises in text and language. Every LLM is generative AI, but generative AI also covers models that make things other than words.
Governance, accountability and data quality, more than any of the technical ones. Our survey found governance to be the biggest single predictor of return, so if your team is going to get fluent in a few terms first, start there.
Achilleas Kasimidis is Global Director of Rev Ops at GoStudent, the online tutoring marketplace operating across more than 10 European countries...
Watch video
At GoStudent, Achilleas Kasimidis and his team set out to automate the first sales call, taking a new enquiry...
Read more
You may have had enough of AI productivity studies, but a new working paper published by the National Bureau of...
Read moreGet ahead with the most actionable insights, playbooks and real-world AI use cases you can adopt right now, in your inbox every week
