You have probably used AI today without calling it that. A phone suggested the next word in a message. A video app proposed something to watch. A translation appeared in a web page. Someone on your team may have asked a chatbot to tidy a paragraph.
These do very different jobs, and that is the first useful thing to know. "AI" is not one machine with one skill. It is a family of methods for getting software to do things that used to need a person's judgement.
A better question than "what is AI?" is therefore: what task is this system doing, what information does it use, and how should I check the result?
Step 1: Task · Overview
What the system is asked to do.
Read all steps as a list
- Task. What the system is asked to do. Classify an email, pull a date out of a contract, summarise a report, draft a reply. A clear task is half the work.
- Information. What it is given or can look up. The email text, a document, a table, a photo. What the system can see limits what it can get right.
- Model or tool. The method that turns information into a result. Often a trained model. Sometimes ordinary software rules, a database search or a calculator working alongside it.
- Output. What comes back. A label, a number, some extracted fields, or a piece of generated text or an image. Generated output is a candidate, not a verified fact.
- Review. How a person checks it. Light checking for a low-risk draft; careful checking where a mistake costs money, trust or safety.
Machine learning: software that learns from examples#
Most of what people call AI today rests on machine learning. Google's introductory course describes it as training a piece of software, called a model, to make useful predictions or to generate content from data.
The contrast with ordinary programming helps. In ordinary programming a person writes the rules: if the invoice total is over a limit, flag it. In machine learning you show a model many examples and let it find the pattern. Nobody writes the rule for "this looks like a spam email". The model learns it from thousands of messages that people have already labelled.
Machine learning comes in a few flavours that are worth recognising:
- Supervised learning learns from examples that come with the right answers. Predicting a number (a delivery time) or choosing a category (spam or not spam) are typical uses.
- Unsupervised learning looks for patterns without being told the answers, for example grouping similar customer messages together.
- Reinforcement learning learns by receiving rewards or penalties for actions, which suits things like game-playing or control problems.
Generative AI: producing candidates#
The tools in the news are mostly generative AI: models that create new content, such as text, images or audio, from an instruction. Google's course describes them as typically trained first on large amounts of data without labelled answers, and then sometimes refined further.
The important word is candidate. A generative model produces something plausible. It does not look up the truth and report it. A draft email is a fine thing to ask for, because you will read and fix it. A date, a price or a legal clause produced from memory is a different matter, and later articles in this series come back to why.
Four everyday jobs#
Here are four ordinary tasks, with the kind of help AI can give. The business is fictional.
| Task | What the AI does | What a person still does | | --- | --- | --- | | Sort incoming email for a small travel agency | Labels each message: booking, complaint, invoice | Checks the complaints queue, fixes wrong labels | | Read a supplier PDF | Extracts invoice number, date and total into fields | Confirms the total before it is paid | | Summarise a 20-page report | Writes a one-page summary | Checks the numbers that matter against the report | | Prepare a first draft of a reply | Writes a polite draft in the agency's tone | Edits, adds facts, decides whether to send |
Notice that in every row the human keeps the decision. That is a good default.
AI and plain automation work together#
Not every clever workflow needs AI. If the rule is simple and never changes, ordinary automation is cheaper, faster and fully predictable: when a form arrives, copy it into the spreadsheet.
AI earns its place where the input is messy or varied: free text, scanned pages, photos, many ways of saying the same thing. In practice the two combine. Software rules handle the predictable parts, a model interprets the messy part, and a person approves anything that matters.
Where a person stays responsible#
Four points deserve a human owner whatever the tool:
- What goes in. Check you are allowed to share the information with the tool. A confidential contract is not a casual thing to paste into a public chatbot.
- Which facts matter. Names, dates, amounts and promises should be verified against a real record.
- The decision. Whether to pay, send, hire or promise is yours.
- The action. Sending an email or changing a record should have an owner.
A quick test for a sensible first use#
Before using AI on a task, ask:
- Is the task repeated often enough to be worth improving?
- Is the input something I am allowed to share with this tool?
- Would a wrong answer be easy to spot and cheap to fix?
- Can someone name what a good result looks like?
- Is there a person who will check it?
If most answers are yes, you have a reasonable candidate. If the task is rare, the input is confidential or a mistake would be costly and hard to see, begin with a smaller step or none at all.
Where to go next#
The next article explains how language models, the engine behind many of today's chat tools, differ from the applications built around them: How language models work, and how they differ from AI applications. When you are ready to pick a real task, read How to choose your first business AI workflow.