Glossary
Twenty bits of AI jargon, in plain English.
Every term a vendor might use on you, defined the way we'd explain it across a table, so nobody can baffle you into buying.
The jargon, translated
Each term gets a definition, a worked example, why it matters, and the mistake we see most.
Large language model (LLM)
Definition. A large language model is software trained on enormous amounts of text so it can read, write and summarise language in a way that feels human.
Worked example. Ask one to turn forty rambling customer emails into a tidy complaints summary and it will draft it in seconds, in your house style if you show it examples.
Why it matters. LLMs are the engine inside most of the AI tools now sold to businesses, including ChatGPT.
Common mistake. Treating one as a database. An LLM predicts plausible text; it doesn't look facts up unless you connect it to your data.
Related: Prompt, Token, Hallucination
Prompt
Definition. A prompt is the instruction you give an AI system: the question, the request and any background you include with it.
Worked example. "Summarise this contract" is a weak prompt. "List the payment terms, notice periods and penalties in this contract as a table" gets a usable answer.
Why it matters. In a built system the prompts are written once, tested and reused, which is why results stay consistent instead of depending on who typed the question.
Common mistake. Blaming the model for vague output when the instruction was vague.
Related: Large language model (LLM), Context window
Fine-tuning
Definition. Fine-tuning means taking an existing AI model and training it further on your own examples so it gets better at one specific job.
Worked example. A model fine-tuned on two years of your quotes learns your pricing language and layout, so drafts arrive sounding like you wrote them.
Why it matters. It's one way to make general AI specific to your business, though retrieval is often cheaper and easier to update.
Common mistake. Assuming every custom build needs it. Most business jobs are solved with good prompts and your data retrieved at the right moment.
Related: Retrieval-augmented generation (RAG), Training data, Model
Retrieval-augmented generation (RAG)
Definition. Retrieval-augmented generation means the AI looks up the relevant parts of your own documents or data first, then writes its answer from what it found.
Worked example. Asked about a client's contract, the system fetches that contract, quotes the actual clause and links to it, rather than answering from general memory.
Why it matters. It grounds answers in your real information, which cuts hallucination sharply and means you can check the source.
Common mistake. Thinking the model has "learned" your documents. It reads them at answer time; change the documents and the answers change too.
Related: Hallucination, Fine-tuning, Context window
Hallucination
Definition. A hallucination is when an AI system states something false with complete confidence, because it generates plausible text rather than checking facts.
Worked example. Asked for a supplier's terms it hasn't seen, a bare model may invent a plausible-sounding notice period, formatted beautifully and entirely wrong.
Why it matters. It's the main reason unmanaged AI can't be trusted with work that matters, and the reason well-built systems retrieve real data and keep a person checking output.
Common mistake. Believing an answer because it's fluent. Fluency is what these models do; accuracy has to be engineered in.
Related: Retrieval-augmented generation (RAG), Human in the loop
Automation
Definition. Automation is having software carry out a repetitive task that a person currently does by hand, on a schedule or a trigger.
Worked example. Payroll checking: instead of a person reading every line each month, a system checks all of them and flags the handful that genuinely need a human eye.
Why it matters. The savings come from tasks done weekly or monthly, forever; that's where days a month come back.
Common mistake. Automating a broken process. If the manual version is muddled, the automated version is muddled faster.
Related: Workflow, Human in the loop
Workflow
Definition. A workflow is the sequence of steps a piece of work moves through, from the thing that kicks it off to the finished result.
Worked example. An invoice arrives, someone checks it against the purchase order, codes it, gets approval, then pays it: five steps, three people, one workflow.
Why it matters. AI is fitted to workflows, not bolted onto businesses. Mapping the real steps, including the unofficial ones, is where every good build starts.
Common mistake. Describing the official process rather than what actually happens; the spreadsheet someone keeps on the side is usually the real system.
Related: Automation, Integration
Integration
Definition. An integration is a connection that lets two pieces of software pass information between themselves automatically.
Worked example. When a deal closes in your CRM, an integration can create the invoice in your accounts package and tell your team's chat, without anyone retyping anything.
Why it matters. Most wasted hours in small businesses are people acting as the integration: copying data from one screen into another.
Common mistake. Assuming your software can't connect. Most modern tools have an API, and even the stubborn ones usually have a workable route in.
API
Definition. An API is the doorway a piece of software provides so other software can ask it for data or tell it to do things.
Worked example. Your accounts package's API lets a built system pull this month's invoices automatically, instead of a person exporting a spreadsheet every Friday.
Why it matters. APIs are what make custom builds possible without replacing your existing tools; we connect to what you already pay for.
Common mistake. Treating "has an API" as a technical detail. When you're choosing software, it decides whether you can ever automate around it.
Related: Integration, Lock-in
Training data
Definition. Training data is the information a model learned from while it was being built.
Worked example. A general model learned from a huge slice of the public internet; a fine-tuned quoting model also learned from the two thousand quotes your team wrote.
Why it matters. What a model was trained on decides what it's good at, and whether your own information was involved is a fair contractual question for any vendor.
Common mistake. Assuming a tool trains on your data by default. Well-configured business AI doesn't, and ours never trains anything outside your business.
Related: Fine-tuning, Machine learning, UK GDPR
Machine learning
Definition. Machine learning is software that improves at a task by finding patterns in examples, instead of following rules a programmer wrote by hand.
Worked example. Show a system thousands of past invoices marked "fine" or "query this" and it learns what a suspicious one looks like, without anyone writing the rules out.
Why it matters. It's the broad field underneath everything else in this glossary; large language models are one branch of it.
Common mistake. Using it interchangeably with AI in vendor claims. Ask specifically what learns, from what data, and what happens when it's wrong.
Related: Model, Training data, Large language model (LLM)
Model
Definition. A model is the trained thing itself: the file of learned patterns that takes an input and produces an output.
Worked example. The system behind ChatGPT is a model; a payroll checker might use one general model for reading documents and a smaller one for spotting odd lines.
Why it matters. Models are swappable parts, not the product. A well-built system can change model when a better or cheaper one appears.
Common mistake. Buying "the best model" rather than the right system. The plumbing around a model decides whether it's useful.
Related: Large language model (LLM), Machine learning, Fine-tuning
Token
Definition. A token is the small chunk of text, roughly three-quarters of a word in English, that AI models read and write in, and that usage is billed by.
Worked example. The sentence "Please chase the unpaid invoices" is about seven tokens. A twenty-page contract is several thousand, which costs pennies to process, not pounds.
Why it matters. Token pricing is why the running costs of most business AI are small next to the salary time it saves.
Common mistake. Fearing usage bills without doing the arithmetic. Processing a document typically costs less than the paper it was printed on.
Related: Context window, Large language model (LLM)
Context window
Definition. The context window is how much text a model can hold in mind at once: the instructions, your documents and the conversation so far.
Worked example. A large window can take in a full year of board minutes in one go; a small one has to work through them in chunks.
Why it matters. It sets what a system can consider in a single pass, and it's why "just paste everything into ChatGPT" stops working as the documents pile up.
Common mistake. Assuming the model remembers yesterday. Outside the window, nothing persists unless the system is built to store and retrieve it.
Related: Token, Retrieval-augmented generation (RAG), Prompt
Human in the loop
Definition. Human in the loop means a person reviews or approves the system's work at a deliberate checkpoint before it counts.
Worked example. A payroll checking system reads every line but a person signs off the flagged ones; the machine does the reading, the human does the deciding.
Why it matters. It's the difference between AI you can trust with real work and AI you have to hope about. Every system we ship keeps a person in charge.
Common mistake. Treating the review as a formality. The checkpoint only protects you if the person genuinely looks.
Related: Audit trail, Hallucination, Automation
Audit trail
Definition. An audit trail is a record of everything a system did and why: what it read, what it decided, and who approved it.
Worked example. When a payroll checker flags a line, the trail shows which rule was tripped, what data it saw, and who cleared it, months later if you need it.
Why it matters. If a regulator, client or accountant asks "why did this happen?", you answer with a record rather than a shrug.
Common mistake. Accepting AI tools that can't explain their own actions. If a vendor can't show you a trail, that's your answer.
Related: Human in the loop, UK GDPR
Co-ownership
Definition. Co-ownership means the system we build is yours as well as ours: you can see inside it, run it where you choose, and keep it if we part ways.
Worked example. If we stopped working together tomorrow, your report-drafting system keeps running, with your team or another developer holding the keys.
Why it matters. It removes the quiet threat in most software relationships: that leaving your supplier means losing your tools.
Common mistake. Not asking about ownership until the exit is tested. Read what happens when you leave before you sign anything, with us or anyone else.
Related: Lock-in, Off-the-shelf vs custom
Lock-in
Definition. Lock-in is when leaving a supplier would cost so much in money, data or disruption that you effectively can't.
Worked example. Five years of client history in a system with no export, priced per seat, on a contract that auto-renews: that business isn't a customer any more, it's a hostage.
Why it matters. Fear of lock-in is a sensible reason to hesitate over software, and the reason our builds are co-owned with your data staying yours.
Common mistake. Only pricing the way in. The way out is part of the cost.
Related: Co-ownership, API, Off-the-shelf vs custom
Off-the-shelf vs custom
Definition. Off-the-shelf means software built for everyone and rented; custom means software built for your specific job and owned.
Worked example. A generic AI email tool answers like a generic company. A custom one, built on your policies and your past replies, answers like you, and doesn't charge per seat forever.
Why it matters. Off-the-shelf is right for standard jobs. When the job is your competitive edge, or oddly shaped, custom pays for itself.
Common mistake. Comparing only the first month's price. Compare three years of subscriptions against owning the thing outright.
Related: Co-ownership, Lock-in, our build vs buy guide
UK GDPR
Definition. UK GDPR is the UK's data protection law, setting the rules for how businesses collect, store and use personal information.
Worked example. Sending your customer list to a free AI tool with no data agreement could put personal data outside your control, which is exactly what the law exists to prevent.
Why it matters. Any AI system touching names, salaries or customer records must be built to this standard from the start; ours are.
Common mistake. Assuming AI is a legal grey zone. The same rules apply, and the tool being clever is not a defence.
Related: Training data, Audit trail
The one thing to do next
We won't promise AI fixes your whole business overnight. Anyone who says that is selling you something.
What we promise is a straight answer. On a free Impact Call you say where your week goes, we say where AI could hand you days back and where it couldn't, and you leave with a one-page opportunity map either way.
Prefer email? Write to jim@ainativ.es and we'll set it up.
- It's free, with no obligation. No pitch deck, no follow-up you didn't ask for.
- You leave with a one-page opportunity map of where AI could help, and where it couldn't.
- Honest arithmetic on a real job in your business, not a generic demo.
- You deal with the founders who scope and build the work, not a sales team.