Guides

An AI glossary for business owners: forty terms explained by what they mean for your firm.

By Zain Masroor · 16 October 2026 · 12 min read

This glossary explains the forty AI terms a UK business owner is most likely to meet in a proposal, a contract or a pricing page, in the order they usually come up: the model and how it is priced, how it is given your information, what can go wrong, the shapes of system a supplier might propose, and the data protection words that apply. Each entry says what the term means and what it means for your firm. Where a term has a cost or a rule attached, the entry links to the guide that covers it.

Talk it through: book a free 30-minute consultation
EntriesForty, grouped by when you meet them
StyleWhat it means for the firm, not the maths
UpdatedWith each new guide

The model and how it is priced

The terms on every pricing page, and what each does to the bill.

01Large language model (LLM). Software trained on very large amounts of text that predicts and generates language. Claude, GPT and Gemini are families of them. For your firm: the engine behind every assistant and most automations; you rent it, you do not own it.
02Model. A specific version of an LLM, with a name, a price and a capability level. Providers ship several sizes; the smallest that passes your test is usually right.
03Frontier model. The largest, most capable model a provider offers, priced accordingly. Needed for hard judgement, wasteful for sorting and extraction.
04Token. The unit models are priced in: roughly three-quarters of a word in English. An invoice is a few thousand tokens; a book is a few hundred thousand.
05Input and output tokens. What you send the model and what it sends back, priced separately, with output usually several times dearer. Long documents in, short answers out is the cheap direction.
06Context window. How much text a model can consider at once, measured in tokens. Large windows let you paste whole contracts; they also make each call dearer.
07API. The programmatic door to a model, used by software rather than people, priced per token and covered by a business data processing agreement. Every build uses it.
08Seat or licence. A per-person monthly price for an assistant such as Claude Team or ChatGPT Business, as opposed to per-token pricing on the API.
09Rate limit. A cap on how many calls or tokens a provider allows per minute; a build has to be designed around it.
10Spending ceiling. A limit you set on what a feature may cost per day or month, tested before go-live. The control that stops a loop becoming an invoice.

How the model is given your information

The terms behind every “answers from our own documents” proposal.

01Prompt. The instruction and material given to a model. Most poor output is a poor prompt.
02System prompt. Standing instructions a builder gives a model about its role, tone and rules, invisible to the user. Where “you are the practice’s assistant; never quote prices” lives.
03Grounding. Making a model answer from supplied material, your documents or your records, rather than general knowledge. The difference between a useful assistant and a plausible one.
04Retrieval-augmented generation (RAG). The standard way to ground: relevant passages from your documents are found and given to the model with each question. Not training; nothing about your documents is learned into the model.
05Embedding. A numerical representation of text that lets software find passages similar in meaning. The index behind retrieval.
06Vector database. Where embeddings are stored and searched. A component of a knowledge assistant, not a product you need to understand.
07Fine-tuning. Training a model further on your own examples to change how it behaves. Rarely worth it for an SME; grounding and prompting solve most cases more cheaply.
08Training data. What a model learned from. On consumer plans your conversations may become training data; on business plans and the API they do not.
09Connector. A permission that lets an assistant read a mailbox, a drive or a system. Switched on deliberately after deciding what it may do there.
10Model Context Protocol (MCP). An open standard, now under the Linux Foundation, for connecting AI assistants and agents to tools and data sources in a uniform way. Why a connector built once works across several assistants.

What can go wrong

The failure modes a proposal should name, and the controls that go with each.

01Hallucination. Confident output that is wrong: an invented case, a made-up figure, a policy clause that does not exist. Structural, not a bug that gets fixed; the answer is grounding, checks and a person.
02Acceptance threshold. The accuracy rate a process needs, agreed as a number before a build and measured on real examples. What makes “finished” a fact.
03Prompt injection. Text in something the model reads (an email, a web page, a document) that instructs it to do something you did not ask. The NCSC says it may never be fully fixed; the controls are least privilege and a person on consequential actions.
04Data leakage. Your information reaching somewhere it should not, most often through a personal AI account. Fixed by business plans and a policy.
05Shadow AI. Staff using AI tools the firm has not approved. The default state in most SMEs.
06Bias. Systematically worse outcomes for a group, which the Equality Act treats as the employer’s discrimination regardless of the tool. Tested before go-live and monitored after.
07Drift. A system’s accuracy falling over time as inputs, documents or models change. Why the weekly sample continues after go-live.
08Human in the loop. A person deciding or checking at a defined point. Required in practice for anything that affects an individual or moves money.

The shapes of system a supplier might propose

Eight shapes cover almost every proposal an SME receives. Knowing which one is on the table sets the price band.

01Assistant. A chat tool a person uses: Claude, ChatGPT, Copilot, Gemini. Bought per seat.
02Automation or workflow. Software that moves data between systems on a trigger, with or without a model in the middle. n8n, Make and Zapier are workflow tools; custom code is the alternative.
03Agent. Software that reads, decides and acts across your systems towards a goal, using a model for the deciding. Needs scoped credentials, ceilings, logging and a person on consequential actions.
04Knowledge assistant. A grounded assistant that answers from your own documents with sources. The most common first build for professional firms.
05Document processing. Reading what arrives (invoices, applications, referrals) into structured records with an exceptions path for a person.
06Voice agent or AI receptionist. A model on the phone that answers, qualifies and books, with honesty that it is an AI and an escalation path.
07Copilot. Microsoft’s name for its assistants across Microsoft 365, and a general word for an assistant embedded in a tool you already use.
08Off the shelf versus custom. A product with a subscription versus software built for you that you own. The right answer depends on whether your process is standard.

The data protection words

The eight words a DPA, a DPIA or an ICO enquiry will use.

01Controller. The organisation that decides why and how personal data is processed: you, for your customers’ and staff’s data.
02Processor. An organisation that processes personal data on your instructions: the AI provider on a business plan or the API, under a data processing agreement.
03Data processing agreement (DPA). The contract that makes a provider a processor, setting what it may do with your data. On file for every provider.
04Sub-processor. A provider’s own suppliers who touch your data, such as its cloud host. Listed by the provider; part of your record.
05Data protection impact assessment (DPIA). A written assessment of the risks of a processing activity, required for high-risk processing such as scoring, monitoring or significant automated decisions.
06Automated decision-making (ADM). A decision made without meaningful human involvement. Under Articles 22A to 22D, as amended by the Data (Use and Access) Act 2025, allowed with safeguards: the person is told, can contest, and can obtain human review.
07Lawful basis. The legal ground for processing personal data. For AI in a business it is usually legitimate interests or contract; consent is rarely right for staff.
08Data residency. Where your data is stored and processed. Some providers offer UK or EU storage on certain plans; inference often still runs elsewhere.

Common questions

What is a token in AI pricing?

The unit models are priced in, roughly three-quarters of an English word. Input and output tokens are priced separately, with output usually several times dearer.

What is the difference between RAG and fine-tuning?

RAG gives the model relevant passages from your documents at question time; nothing is learned into the model. Fine-tuning trains the model further on your examples. For an SME, RAG solves most cases more cheaply and keeps your data out of the model.

What is an AI hallucination?

Confident output that is wrong: an invented reference, figure or clause. It is structural to how models work, so the controls are grounding, checks and a person, not waiting for a fix.

What is an AI agent?

Software that reads, decides and acts across your systems towards a goal, using a model for the deciding. It needs scoped credentials, spending ceilings, logging and a person on consequential actions.

What is prompt injection?

Instructions hidden in something the model reads, such as an email, that make it act against your intent. The NCSC says it may never be fully fixed; reduce the impact with least privilege and human confirmation.

What is MCP?

The Model Context Protocol, an open standard for connecting AI assistants and agents to tools and data, donated by Anthropic to the Linux Foundation’s Agentic AI Foundation in December 2025 and supported across the main assistants.

Is an AI provider a controller or a processor?

On a business plan or the API, a processor acting on your instructions under a data processing agreement. On a consumer account, the position is the provider’s own terms, which is why personal accounts are not used for work.

Met a term in a proposal that is not here?

Send it over. We explain every term in our own proposals in plain English, and we add the ones people ask about to this page.

Book a free 30-minute consultation Get a fixed quote
5.0 on Google, every review five starsFree consultation · Fixed price in writing · Code and accounts yours from day one