Large Language Model (LLM)
beginnerA large language model is an AI trained on vast amounts of text to predict and generate language. It powers tools like ChatGPT, Claude and Gemini — it can write, summarise, classify and answer, but it predicts plausible text rather than 'knowing' facts.
What it means for your business
The LLM is the engine inside most AI tools you'll buy. You rarely choose one directly, but it shapes cost, quality and data handling — so it's worth knowing whether a vendor uses a frontier model or a cheap one behind the scenes.
Hallucination
beginnerA hallucination is when an AI states something false as if it were true — an invented statistic, a fake citation, a made-up policy. It happens because language models generate plausible text, not verified facts.
What it means for your business
This is the single biggest risk in customer-facing AI: a confident wrong answer about your pricing, availability or policy. It's manageable — by grounding the AI in your real data and keeping a human on high-stakes replies — but never assume it's 'solved'.
The 2026 reality
Hallucination is reduced, not eliminated, in 2026. Any vendor who tells you their AI 'doesn't hallucinate' either doesn't understand the technology or is hoping you don't. The right answer is grounding plus guardrails, not a promise.
Machine Learning
beginnerMachine learning is software that learns patterns from data rather than following rules a programmer wrote by hand. You show it thousands of examples — past invoices, support tickets, sales outcomes — and it works out the patterns itself, then applies them to new cases it has never seen.
What it means for your business
Most of the 'AI' a UK SME actually benefits from is plain machine learning, not chatbots: predicting which leads will close, flagging late-paying customers, forecasting stock. It needs decent historical data, so if your records live in someone's head or a messy spreadsheet, that's the first job, not the model.
The 2026 reality
By 2026 the word has been swallowed by 'AI', but the distinction matters commercially. A boring ML model that scores your leads is often cheaper, faster and more reliable than wiring an LLM into the same job. Don't let a vendor sell you a GPT-shaped solution to a problem a regression would solve for a fraction of the cost.
Generative AI
beginnerGenerative AI is software that produces new content — text, images, audio, code — rather than just classifying or predicting. You give it a prompt and it generates a fresh response, drawing on patterns learned from vast training data. ChatGPT, Claude and image tools are all generative AI.
What it means for your business
This is the part of AI that creates leverage on output: drafting proposals, replying to enquiries, writing first-pass copy, summarising calls. The win is throughput per person, not replacing staff. The catch: it generates plausible content, not correct content — so it speeds up the first draft, it doesn't remove the need to check.
The 2026 reality
The 2026 reality is that generative AI is brilliant at the first 80% and dangerous at the last 20%. Businesses that win use it to compress drafting time, not to ship unreviewed output to clients. The firms quietly getting hurt are the ones treating 'it sounds confident' as 'it's right'.
Prompt
beginnerA prompt is the instruction you give an AI — the question, request or context you type in. The quality of the prompt heavily shapes the quality of the answer: clear instructions, relevant context and examples produce far better results than a vague one-liner.
What it means for your business
Prompting is a skill your team can learn in an afternoon and it changes the return on every AI tool you pay for. The practical move is to build a small library of tested prompts for your recurring jobs — quote follow-ups, enquiry triage, meeting notes — so output is consistent rather than depending on whoever's typing.
The 2026 reality
'Prompt engineering' was wildly over-mystified, and by 2026 the better models forgive sloppy prompts far more than they did in 2023. The durable skill isn't clever phrasing — it's being specific about what you actually want and giving the model your real context. Anyone selling a £500 'prompt pack' is selling a sticking plaster.
Token
beginnerA token is the unit of text an AI model reads and writes — roughly a word or part of a word. A rough rule is 100 tokens to around 75 English words. Models price by tokens and have token limits, so tokens are the currency of AI.
What it means for your business
Tokens are how your AI bills land. Every prompt and reply is metered, so a tool that quietly stuffs huge instructions or whole documents into each request can cost ten times more than a lean one for the same job. When comparing vendors, ask how many tokens a typical task consumes — that's the number that drives your monthly bill.
Context Window
advancedThe context window is how much text an AI model can hold in mind at once — the prompt, any documents you've supplied, and the conversation so far — measured in tokens. Once you exceed it, the earliest content falls out of view and the model effectively forgets it.
What it means for your business
This is why an AI assistant 'forgets' what you said earlier in a long chat, and why feeding it your entire knowledge base in one go is the wrong design. A large window lets you drop a long contract in for analysis, but relying on it for memory across thousands of conversations is expensive and unreliable — that's a job for retrieval.
The 2026 reality
By 2026 vendors brag about million-token windows, but bigger is not the win it sounds. Models suffer 'lost in the middle' — they attend well to the start and end and skim what's between — and you pay for every token whether it's used well or not. The mature pattern is RAG: retrieve only the relevant slice.
Training vs Inference
advancedTraining is teaching a model — feeding it enormous data so it learns patterns, done once and costing millions. Inference is using the trained model to answer a single request. Training builds the engine; inference is each journey. Almost everything a business does with AI is inference.
What it means for your business
This distinction kills a costly misconception: you almost never need to 'train your own AI'. You're buying inference — paying per request to use a model someone else trained. When a vendor pitches a 'custom-trained model for your business', press them: they usually mean retrieval or fine-tuning, both far cheaper than training from scratch, which no SME should attempt.
The 2026 reality
The honest 2026 state: training a frontier model is a job for a handful of labs with eye-watering budgets, and 'we'll train an AI on your data' is mostly marketing for grounding the model in your data at inference time — a completely different, much cheaper thing. Your data doesn't need to be 'in the model'. It needs to be retrievable when the model answers.
Fine-Tuning
advancedFine-tuning takes an already-trained model and trains it a little further on your own examples, so it absorbs a specific style, format or task. It changes the model's behaviour — unlike RAG, which feeds the model facts at the moment of answering without altering the model itself.
What it means for your business
Fine-tuning is the right tool for teaching consistent style or structure — replying in your brand voice, always returning a fixed format — not for teaching facts. The trap is reaching for it first: it needs hundreds of clean labelled examples, costs to retrain whenever things change, and most problems people bring to it are solved better and cheaper by a good prompt or retrieval.
The 2026 reality
The 2026 pattern that catches people out: they fine-tune to inject knowledge — prices, policies — then wonder why the model still gets facts wrong and can't be updated without retraining. Fine-tuning shapes how the model responds; RAG controls what it knows. Most 'we need to fine-tune' instincts should be 'we need retrieval'.
Multimodal AI
beginnerMultimodal AI works with more than one type of input or output — text, images, audio and sometimes video — in a single model. You can show it a photo and ask a question about it, hand it a PDF, or have it listen to a voice note, rather than being limited to typed text.
What it means for your business
This quietly removes friction in everyday jobs: read a photographed receipt or a scanned invoice, pull figures off a screenshot, transcribe and summarise a call, check a product image against a description. A lot of manual data entry and 'someone reads it and types it in' work can be handed off without building anything custom.