The AI words, in plain English.

Software that does things we used to think needed a person: reading, writing, summarising, deciding, spotting patterns.

In practiceWhen a tool drafts your meeting notes instead of you typing them, that is AI doing a task, not magic.

Instead of a developer writing every rule, the software learns patterns from lots of examples.

In practiceYour spam filter was never told what spam looks like — it learned from millions of emails people marked as spam.

The finished, trained 'brain' you are talking to. Different models are different brains with different strengths and prices.

In practiceClaude, ChatGPT and Gemini are products; behind each sits one or more models you can switch between.

A model trained on an enormous amount of text so it can read and write human language well.

In practiceIt is what lets you paste a 40-page contract and ask, in normal Finnish or English, what you should worry about.

AI that produces something new — text, images, slides, code — rather than only sorting or scoring what already exists.

In practiceAsking for a first draft of a customer email is generative AI. Ranking leads is not.

The expensive, one-off process of teaching a model from huge amounts of data. Done by the AI companies, not by you.

In practiceYou almost never train a model. You use one someone else trained and give it your context.

A chat window on top of a model. The window is not the intelligence — the model is.

In practiceSame model, different chat windows: on your phone, inside Slack, or embedded in your website.

Whatever you type or paste in. Your instruction plus the material you give it.

In practiceYou paste last quarter's report and write 'summarise this for our board in five bullets' — that whole thing is the prompt.

The small chunks of text an AI reads and writes in — roughly three quarters of a word. Pricing is usually per token.

In practiceA one-page memo is about 500 tokens. When a vendor quotes 'per million tokens', that is around 1,500 pages.

How much it can hold in mind at once — your whole conversation plus every document you pasted.

In practiceIn a very long chat the earliest details start falling out. Start a fresh chat for a new topic.

A dial between predictable and creative.

In practiceLow for a legal summary you want boringly consistent. Higher for campaign ideas where you want range.

The internal settings learned during training. Often quoted in billions as a rough size measure.

In practiceBigger is not automatically better for your job — and it is always slower and pricier.

◇Not the settings you adjust. Those are just settings.

Taking an existing model and training it a bit more on your own material so it adopts your style or domain.

In practiceRarely the first answer. Good prompts and giving it your documents usually get you 90% there for a fraction of the cost.

The model handles more than text — images, PDFs, audio, screenshots.

In practicePhotograph a whiteboard after a workshop and ask for a written summary and action list.

The learnable skill of asking well: giving role, context, examples and the format you want back.

In practice'Write something about our product' gets mush. 'You are our B2B copywriter, here are three past emails, write a fourth for this audience, max 120 words' gets usable output.

Standing instructions set once, behind the scenes, that apply to every conversation with that assistant.

In practice'Always answer in Finnish, always cite the source document, never invent numbers' — set once for the whole team.

Showing two or three examples of what good looks like instead of describing it.

In practicePaste three offers you were proud of, then ask for a fourth. Quality jumps immediately.

Treating the first answer as a first draft and steering it, rather than accepting or giving up.

In practice'Shorter. More direct. Drop the adjectives. Now write it as the CFO would.'

A named person reviews and approves before anything leaves the building or hits a customer.

In practiceAI drafts every reply to a complaint; your service lead approves each one before it sends.

Forcing answers to come from sources you supplied instead of the model's general memory.

In practice'Answer only from the attached policy. If it is not in there, say you don't know.' This is the single biggest quality fix.

AI that takes multiple steps on its own towards a goal — looking things up, using tools, checking its own work.

In practiceNot just drafting the offer, but pulling the pricing, filling the template, and putting it in your drafts folder.

◇An assistant answers. An agent does a sequence of things.

AI sitting inside a tool you already use, helping with the task in front of you.

In practiceThe suggestion pane in your email, your documents, or your CRM.

Answering from your own documents: the system searches your material first, then answers using what it found.

In practice'What is our parental leave policy?' — it finds the right handbook page and answers with a link, instead of guessing.

A standard plug socket that lets an AI safely reach your real tools — calendar, files, CRM, database.

In practiceIt is why an assistant can actually book the meeting rather than telling you how to book it.

The plumbing that lets two pieces of software talk, so AI can run inside your own systems.

In practiceHow AI gets into your webshop's chat instead of living in a browser tab someone forgot to open.

A repeatable process that runs without anyone pressing start.

In practiceEvery inbound enquiry is summarised, categorised and routed to the right person before anyone opens it.

Building working software by describing what you want in plain language and letting AI write the code.

In practiceA sales lead builds their own quote calculator in an afternoon without waiting two quarters for IT.

The model states something false with complete confidence. It is generating plausible text, not looking up truth.

In practiceNever let a number, a name, a date or a legal citation reach a customer unchecked. Ask for the source.

Skew learned from the training data, which quietly shows up in the output.

In practiceMatters most in hiring, credit, and anything touching people. Keep a human deciding.

The model's built-in knowledge stops at a date. Anything after that it has not seen.

In practiceFor current facts, either give it the document or use a tool that can search the web.

Where your text goes, who can see it, and whether it is used to train future models.

In practiceBusiness plans have different rules from personal data. Agree what may be pasted before people start pasting.

Staff using personal AI accounts for work because nothing official was offered.

In practiceIt is already happening in your organisation. A clear policy plus proper tools beats a ban nobody follows.

European rules that sort AI uses by risk. Most everyday office use is low risk; people-affecting decisions are not.

In practiceDrafting emails: fine. Automatically screening job applicants: obligations you need to know about.