Learn AI without jargon
AI glossary for your business
42 words you hear (or will hear) when people talk about AI, explained without jargon and with your business in mind: what they mean, why they matter to you and an example.
42 terms
If you only read six, make it these
- 1PromptThe text you write to ask an AI for something.Read more ↓
- 2HallucinationWhen AI confidently states something false or nonexistent: a figure, a law, a quote, a link.Read more ↓
- 3ContextEverything you tell the AI so it answers well: who you are, what it’s for, who it’s aimed at, what data you have.Read more ↓
- 4Personal dataAny detail that identifies a person: name, ID number, email, phone, address, bank details, photos, purchase history.Read more ↓
- 5Human reviewA person reads and approves what the AI produced before it leaves your business or becomes a decision.Read more ↓
- 6Privacy settingsThe assistant’s options that decide whether it keeps your history and whether it uses it to improve the model.Read more ↓
A9 terms
AI agent
Also: agent, agentic AI
An AI that doesn’t just answer: it also takes steps on its own (searching, opening files, sending an email) to complete a task.
A regular assistant answers you; an agent acts. That saves time, but it also multiplies the risk if it gets something wrong, because it can send an email or change a record without you seeing it. When you start, give it tasks that end in a draft you review, not an automatic send.
ExampleYou ask it to find three packaging suppliers, compare prices and leave you a table. You decide who to write to.
AI assistant
Also: chatbot, conversational assistant, AI chat
A program you talk to in writing (or by voice) that answers, drafts, summarizes or explains.
ChatGPT, Claude, Gemini and Copilot are examples. Most have a free version with limits and a paid one with more capacity. A free one is enough to get started in your business: what matters isn’t the brand but knowing how to ask (the prompt) and checking what you get back.
AI use policy
Also: internal AI rules, shadow AI
A short sheet with the rules your team follows when using AI: which tools, what can be pasted and what always gets reviewed.
If you have employees or contractors, they’re already using it, with or without your rules (“shadow AI”). One page is enough: allowed tools, data that is never pasted, who reviews what goes out and who to ask. It’s the cheapest protection against data leaks and mistakes.
Anonymize
Also: pseudonymize, remove names, hide data
Removing or replacing the details that identify a person before you paste a text into an AI.
The practical way to use AI with real cases without handing over your customers’ data. Replace names with “Client A”, remove ID numbers, bank details, phone numbers and addresses, and keep only what AI needs to help you. If the text still lets someone recognize the person, it isn’t anonymized.
ExampleInstead of “Martha Ross, 4 Main Street, owes $1,200”, you write “Client A, individual, has owed $1,200 for 45 days”.
Personal dataData protection laws (GDPR and others)AI use policy
API
Also: API connection
The door through which one program talks to another. Many tools use it to connect to an AI without you copying and pasting.
You don’t need to code for it to affect you: when an invoicing tool, a CRM or a website “includes AI”, it usually uses an AI provider’s API behind the scenes. It’s worth asking what data of yours is sent and what the provider says about storing it or training on it.
Artificial intelligence (AI)
Also: AI
Programs that can do tasks that used to require human intelligence: understanding text, recognizing images, writing, translating, predicting.
It’s a very broad term. What almost every business uses today is generative AI in the form of an assistant. It doesn’t think or understand like a person: it’s very good at spotting and reproducing patterns, which is why it needs supervising.
Assistant memory
Also: remembers conversations
A feature that lets the assistant keep details about you from one conversation to the next.
It’s handy (you don’t repeat who you are), but it stores information about you and your business. Check what it has saved, delete what you don’t want and turn it off if you handle sensitive information or share the account with others.
Assistant web search
Also: browsing the internet, with sources
When the assistant checks web pages on the spot to answer, instead of using only what it learned.
It’s what you need for current questions: prices, rules, deadlines. Even so, always ask for the links and open the important ones: it can misread a page or trust a bad source. For tax or legal matters, the right source is the official one or your accountant.
Automation
Also: automated workflow, Zapier, Make
Letting a repetitive task run on its own when something happens: an email arrives, an invoice is paid, a form comes in.
Automations follow fixed rules; AI can be added as one more step (for example, summarizing the message). Start with small, low-risk tasks, and always keep a way to see what the automation has done.
ExampleEvery time someone fills in the form on your website, it’s saved to a spreadsheet and you get an alert with a summary.
B1 term
Bias
Also: AI prejudice, discrimination
An unfair or unrepresentative tendency in AI’s answers, inherited from the data it learned from.
It matters when AI touches people: job ads, candidate screening, rating customers. Check that what it writes doesn’t exclude anyone without reason (age, sex, origin) and don’t let it decide who to hire or who to serve. That takes your judgment.
C4 terms
Context
Also: background information
Everything you tell the AI so it answers well: who you are, what it’s for, who it’s aimed at, what data you have.
The number one cause of a weak answer is too little context. If you ask for “an email to chase an invoice”, you get something generic. If you add who the client is, how late they are and the tone you want, you get something usable. That’s why the prompts on this website have gaps to fill in.
Example“Write a reminder” versus “Write a friendly reminder for a 3-year client who is 15 days late on an $800 invoice”.
Context window
Also: conversation memory, length limit
The amount of text the assistant can keep “in view” at once, between what you give it and what it answers.
If the conversation or document is very long, the assistant may forget the beginning or miss a detail. For long jobs, split the document into parts, ask for summaries by section and repeat the important instructions.
Custom instructions
Also: profile instructions, projects, assistant memory
A fixed text you give the assistant once (who you are, your business, your tone) so it takes it into account in every conversation.
It saves you repeating the same thing in every request. A good starting text includes what you do, who you sell to, how you talk to your customers and what it must never do (make up figures, give legal advice). Review it now and then: if your business changes, change it.
Example“I run a neighborhood hair salon in Denver with 4 employees. I keep things friendly and informal with customers. Don’t make up prices: ask me.”
Customer service chatbot
Also: website bot, WhatsApp assistant
An assistant placed on your website or messaging app to answer your customers automatically.
It can save you the repeated questions (opening hours, prices, order status). To work, it needs verified answers, clear instructions on what it can’t promise and always a way to reach a person. A bot that invents a price or a guarantee commits you.
D3 terms
Data analysis with AI
Also: analyze a spreadsheet, upload an Excel file
Uploading a table (Excel, CSV) to an assistant so it summarizes it, finds patterns or prepares charts.
It helps you understand your sales, expenses or customers without being an Excel expert. Two precautions: check the important math in your spreadsheet (AI gets calculations wrong more often than writing) and remove customers’ personal data first.
Data protection laws (GDPR and others)
Also: GDPR, data privacy law, CCPA
The laws that protect people’s personal data, including your customers’: the GDPR in Europe and similar laws in many other countries and states.
They usually apply to your business even if you work alone, when you handle customer data, and also when you put it into an AI tool. In practice: use only the data you need, anonymize when you can, choose tools with clear terms and keep a record of what you do with the data. For the specifics of your case, talk to a professional.
Deepfakes and AI impersonation
Also: cloned voice, fake video, CEO fraud
Fake audio, video or messages created with AI that imitate a real person.
They’re already used to scam small businesses: a call in “your boss’s” or a supplier’s voice asking for an urgent transfer. Simple defense: for any unexpected payment or change of bank details, confirm through another channel you already knew (you call the usual number) and agree a double-check rule with your team.
E1 term
EU AI Act
Also: AI Act, AI law
The European Union law that regulates AI according to the risk of how it’s used. Other countries are passing their own rules.
If you use an assistant to draft, summarize or organize your work, your use is normally low risk. Get advice if AI is going to make or influence decisions about people (hiring, scoring customers, granting credit) or if you’re going to offer an AI tool to others. Obligations apply in phases and vary by country: check the current ones with your advisor.
Data protection laws (GDPR and others)AI use policyHuman review
F3 terms
File uploads
Also: upload a document, upload a PDF
The option to upload a PDF, an image or a spreadsheet to the conversation so the AI can work with it.
It’s the easiest way to get AI to use your real information: a contract, a quote, a statement. It’s also the easiest way to upload data you shouldn’t have. Before attaching, check what it contains and remove anything personal or confidential.
Fine-tuning
Also: training your own model
Training a model further with your own examples so it behaves in a specific way.
It sounds appealing, but for almost any small business it isn’t the first step: it’s costly, needs lots of good examples and goes out of date. It almost always pays off more to start with good prompts, custom instructions and, if needed, giving it your documents (RAG).
RAG (using your documents)Custom instructionsLanguage model (LLM)
Free plan and paid plan
Also: subscription, Pro version
Almost every assistant offers a free version with limits and a paid one with more capacity and more control.
Start free. Move to paid when you notice the limits holding you back, or you need more powerful models or more privacy (for example, a team plan with better guarantees about your data). If several people in the business use AI, look at business terms instead of scattered personal accounts.
G1 term
Generative AI
Also: GenAI
AI that creates new content (text, images, audio, code) from what you ask for.
It’s what you use when you talk to an assistant or generate an image. Unlike a search engine, it doesn’t return something that already existed: it produces a new answer each time, which is why it can vary or be wrong.
H2 terms
Hallucination
Also: made-up answer, invented data
When AI confidently states something false or nonexistent: a figure, a law, a quote, a link.
It’s the most important error to know about. AI doesn’t “know” when it doesn’t know: it generates the answer that sounds most likely. That’s why you should never use a figure, legal deadline, date or name it gives you without checking it. Ask it to tell you what it’s missing to answer and not to make things up.
ExampleYou ask about the deadline for a grant and it gives you a very specific date. If it didn’t look it up in the official source, it may be made up.
Human review
Also: human in the loop, supervision
A person reads and approves what the AI produced before it leaves your business or becomes a decision.
It’s the rule that holds everything together. AI proposes, you decide. Check figures, names, dates, legal references and any promise to a customer above all. The bigger the consequence, the more attention, and for tax, employment or legal matters, have your professional look at it too.
I2 terms
Integration
Also: connecting tools, connector
Connecting two tools so they share data without you copying it by hand.
When you connect your email, calendar or spreadsheet to an assistant, it can work with your real information. Upside: less copying and pasting. Risk: you’re giving it access. Connect only what’s needed and check the permissions you grant.
Iterate
Also: refine, ask for changes, second pass
Improving the answer with follow-up requests instead of expecting the first one to be perfect.
The habit that most separates people who get value from AI from people who get frustrated. Ask for a draft first, then correct it: “shorter”, “warmer”, “remove the third sentence”, “add an example”. Two or three rounds are usually enough.
L1 term
Language model (LLM)
Also: LLM, model, large language model
The engine inside the assistant: a system trained on huge amounts of text to predict and generate language.
The assistant is the app you see; the model is what does the thinking behind it. Some models are faster and cheaper, others more powerful. A simple one is enough to write an email; for analyzing a long document or reasoning with several pieces of data, a more powerful one usually does better.
M2 terms
Machine learning
Also: ML
The technique by which a program learns patterns from many examples, instead of following rules written by hand.
It’s behind spam filters, product recommendations and sales forecasts. Generative AI is one application of this technique. If a tool says it “learns”, ask what data it learns from and whether it’s yours.
Multimodal
Also: understands images, voice and photo
When the assistant works with several types of content at once: text, images, audio, documents.
You can take a photo of an invoice, a whiteboard or a product and ask it to transcribe, summarize or describe it. Useful day to day, with the usual care: check what it read and don’t upload images with third parties’ personal data.
P6 terms
Personal data
Also: customer data, personal information
Any detail that identifies a person: name, ID number, email, phone, address, bank details, photos, purchase history.
These need the most care when you use AI. Your customers didn’t give them to you so they’d end up in a third-party service. The easy rule: don’t paste personal data into an assistant; anonymize first. For specific questions, talk to your data protection advisor.
AnonymizeData protection laws (GDPR and others)AI use policy
Privacy settings
Also: don’t train on my data, history
The assistant’s options that decide whether it keeps your history and whether it uses it to improve the model.
Spend five minutes reviewing them in every tool you use: whether history is saved, whether it’s used for training and how to delete it. It’s the simplest control you have, and hardly anyone touches it.
Prompt
Also: instruction, request, question to AI
The text you write to ask an AI for something.
The quality of the answer depends a lot on the quality of the prompt. A good prompt says who you are, what you want, who it’s for, with what data and in what format, and asks the AI to ask for anything it’s missing. You don’t need to be technical: you need to be specific.
Example“You are my admin assistant. Write a friendly but firm email to chase an $800 invoice that is 15 days late. Ask me anything you need before you write.”
Prompt engineering
Also: knowing how to ask
The craft of writing good requests to get good results from an AI.
More than a secret technique, it’s habits: give context, state the format you want, include examples, ask it to ask questions and correct over a few rounds. That gets you most of the improvement; you don’t need any special course.
Prompt template
Also: prompt with gaps, reusable prompt
A prompt written once, with gaps you fill in each time with the details of the case.
It’s the format of this library: the text is already thought out and you only fill in the gold gaps. When a prompt works for you, save it as your own template: it’s the fastest way for your team to work the same way and for the result to be predictable.
Prompt with examples
Also: few-shot, giving examples
Including one or two examples of what you want in your request, so the AI copies the format and tone.
It’s one of the cheapest improvements: instead of describing how you write, paste two of your messages you like and ask “write a new one in this style”. It works for emails, posts, product descriptions and review replies.
R1 term
RAG (using your documents)
Also: retrieval-augmented generation, chat with your documents
A way for AI to answer by consulting your own documents (catalog, manual, FAQs) instead of making things up.
It’s what’s behind a chatbot that answers with your real prices and terms. It works only as well as your documents are good and up to date. If your rates change, you have to update the source or the bot will keep saying the old ones.
S1 term
System prompt
Also: background instructions
The background instructions that define how an assistant behaves before the user writes anything.
It’s what decides, for example, that the bot on your website is friendly, formal or never promises discounts. If you set up a chatbot, this text is the most important thing you’ll write: what it knows, what it can’t do and when it hands over to a person.
T4 terms
Temperature
Also: model creativity
A setting that makes AI more predictable (low) or more varied and creative (high) when it answers.
In everyday assistants you almost never touch it, but it explains why the same question gets different answers. For precision tasks (summaries, extracting data) you want a stable answer; for name or campaign ideas, a more varied one.
Token
Also: chunk of text
The unit AI reads and writes in: a piece of a word. A common word is usually one or two tokens.
It matters because pay-per-use plans and limits are measured in tokens, and because there’s a maximum number of tokens the model can handle at once. You don’t have to count them: just know that very long texts cost more and may get cut short.
Training data
Also: training on your data
The texts, images and other material the model was taught with, and sometimes also what users write.
It matters for a practical reason: depending on the provider and the plan, what you write may or may not be used to improve the model. Check your assistant’s privacy settings and, if you work with client or confidential information, choose a plan or setting that doesn’t train on your conversations.
Personal dataPrivacy settingsData protection laws (GDPR and others)
Transcription
Also: speech to text, dictation, meeting notes
Turning a recording (meeting, voice note, call) into written text.
One of the most useful helps for a small business: you record a meeting or dictate an idea, then ask for a summary, a task list or a draft. Always tell the people you record and avoid recording conversations with sensitive data without a clear basis for doing so.
U1 term
Usage limits
Also: quota, messages per day
The cap on messages or capacity each assistant gives you on its free plan (and on paid ones).
If you run out of messages halfway through a job, this is why. Free plans usually have daily caps and less powerful models. Plan heavy tasks (long documents, data analysis) for when you have room.
No term matches that search. See the best practices or search the prompt library.
Keep learning
Last reviewed: November 2026. General guidance to help you organise your business. It is not tax, employment, accounting, legal or financial advice: for those matters, talk to your accountant or lawyer.