What people mean when they say AI

Say “AI” in a meeting and ten people picture ten different things. One thinks of a chatbot. Another thinks of the system that flags fraudulent transactions. A third pictures robots. They are all partly right, and that is the problem.

The word became an umbrella. Inside it are tools with very different strengths, costs, and ways of failing. When the term is fuzzy, the decisions made with it get fuzzy too.

This post gives you a practical mental model. Not technical for its own sake, but useful for deciding. The goal is simple: next time someone says “we should use this for the problem,” you can ask a better question back.

Why “AI” is such a slippery word

Even the official definitions are broad on purpose. The OECD, whose wording has been adopted by the European Union, the United States, and others, defines these systems around a simple idea: they infer from inputs to produce outputs such as predictions, content, recommendations, or decisions (OECD AI Principles).

Notice how much fits inside that definition. A spam filter. A tool that writes marketing copy. A system that recommends prices. A bot that books meetings.

That breadth works for public policy. It works less well when you need to decide where to spend money. For that, it helps to split the umbrella into three categories.

Three categories worth keeping straight

1. Predictive models

This is the oldest and least glamorous branch, and it quietly runs a lot of the economy. You give a system many past examples. It learns patterns, then estimates what may happen in a new case.

Will this customer churn? Does this transaction look fraudulent? How many units will we sell next month?

Predictive models work well with narrow questions, reasonably clean historical data, and a clear variable to predict. They work badly when data is thin, messy, or when the past stops resembling the present. They can also absorb bias from the data without warning you.

2. Generative models

This is the branch that reached the headlines. These systems produce new content: text, images, audio, code. The large language models behind chat assistants are the most familiar example.

Adoption grew quickly. In Stanford’s AI Index, the share of organizations reporting that they used generative tools in at least one business function rose from 33 percent to 71 percent in a single year (Stanford HAI AI Index 2025).

Generative models are useful for first drafts, summaries, translation, rephrasing, early ideas, and turning rough notes into something readable. They are fast and flexible. Their problem is different: they are not reliably correct.

A language model does not “know” in the human sense. It predicts likely text. That is why it can state something false with full confidence. It also struggles with exact arithmetic, with recognizing what it does not know, and with any task where “sounds plausible” is not enough.

3. Automation and agents

The third category does not just answer. It takes actions.

Plain automation follows fixed rules: when an invoice arrives, file it in this folder and notify this person. Agents are a newer pattern: a generative model gets tools and some room to decide steps, like searching a database, filling a form, or sending an email.

Automation with fixed rules works well when the process is stable. It is predictable and easy to audit. Agents can help with messier, multi-step tasks, but they inherit the problems of the model behind them. They also add a new risk: a wrong action is usually more expensive than a wrong answer.

Most real products mix categories. A support tool might predict which tickets are urgent, draft replies, and close simple cases on its own. To decide well, you need to know which part you are evaluating.

A few terms, defined simply

Model. The trained system itself, the thing that takes an input and produces an output. Think of it as a recipe learned from examples rather than written by hand.

Training data. The examples the model learned from. Their quality sets the ceiling. If the data is narrow, old, or biased, the model will carry that too.

Token. The unit a language model reads and writes in, usually a word or part of a word. It matters for two practical reasons: many services charge by token, and every model has a limit on how many tokens it can consider at once. As a sign of how quickly costs have fallen, Stanford reports that running a model at a 2022 quality level went from about 20 dollars per million tokens to 7 cents in roughly two years (Stanford HAI AI Index 2025).

Hallucination. When a generative model produces content that sounds right but is false. A peer-reviewed survey defines it as plausible but nonfactual content (ACM Transactions on Information Systems). NIST prefers the term “confabulation” and treats it as a structural trait of these systems, not a rare bug (NIST AI 600-1, Generative AI Profile).

The important idea for decision-makers is simple: confident and correct are not the same thing.

The gap between a demo and a Tuesday

A demo is designed to succeed. It uses a clean example, a forgiving audience, and someone who knows exactly what to type. Real work is different: incomplete data, exceptions, tired people, and a customer waiting on the other end.

A tool can impress for ten minutes and still be wrong often enough to cost more than it saves. The cost appears when someone has to review, correct, explain, or undo.

The useful questions are not flashy. How often is it wrong? How will we detect it? What happens with the cases the demo skipped? Who is accountable if it fails?

Treat the demo as a starting hypothesis, not a verdict.

Where this leaves you

You do not need to become technical. You need a model clear enough to ask better questions.

When someone says “let’s use AI,” you can ask: for predicting, generating, or acting? Where is that category strong? Where does it fail? What is the distance between this demo and reliable use in a normal work week?

That is the foundation. Once the word “AI” means something specific in your head, the next question becomes practical: where does it actually earn its keep in a business? We take that up in where AI fits in a business. And because your team is very likely using these tools already, with or without a plan, it is worth reading what your team is already doing with AI too.

First understand what you have. Then decide what to do with it. That order saves more money than it seems.