Sorting the Vocabulary
Five terms cover most of what any vendor will ever say to you, and knowing them is most of the defense.
A contractor I have worked with asked me a question this spring that sounded almost sheepish. His new scheduling software had relaunched with “AI” stamped across the website, the price had gone up accordingly, and he wanted to know what, exactly, he was now paying for.
It took twenty minutes to find out, and the answer was: a rules engine he already had, three years old, wearing a new label. Nothing in the product had learned anything. The marketing had.
His embarrassment was misplaced, because the confusion is manufactured. The industry has every incentive to blur its own vocabulary; precision is the buyer’s job now. The good news is that the working glossary is short. Five terms, properly sorted, will carry you through nearly any sales call, board discussion, or trade-show floor in the country.
The Nesting Dolls
The terms nest inside each other, and the nesting is the map.
Artificial intelligence is the outermost doll and the least precise, the seventy-year-old umbrella term for any software performing tasks we associate with human intelligence. Because it describes an aspiration rather than a mechanism, it is the term most beloved by marketing departments and the one that should trigger your follow-up questions rather than end them.
Machine learning is the first meaningful layer inside. The defining feature is in the name: the system derives its own patterns from data instead of executing rules a programmer wrote by hand. Show it ten thousand invoices labeled “paid late” and “paid on time” and it will find the predictive signals itself, including ones no human thought to specify. Most of the quiet, profitable AI of the last twenty years, fraud detection, demand forecasting, spam filtering, lives here.
Deep learning is a further layer in: machine learning built on neural networks, loose mathematical sketches of the brain stacked many layers deep. This is the technique that conquered images, speech, and language over the past decade, and it is the engine inside everything currently making headlines.
Generative AI is the newest doll, the deep-learning systems that produce new content, text, images, code, audio, rather than merely classifying or predicting from what exists. Large language models are its most consequential branch, and the bulk of this series concerns them.
And then there is automation, which sits outside the nesting dolls entirely, and the distinction is the one that would have saved my contractor his twenty minutes. Automation executes rules: when the form is submitted, create the ticket; when the invoice ages past thirty days, send the reminder. It learns nothing, which is its glory. Rules-based automation is predictable, auditable, cheap, and for an enormous range of small-business drudgery it is the correct tool, no intelligence required. The skill of this era is knowing which problems want rules and which genuinely want learning, and refusing to pay learning prices for rules work.
The Horseless Carriage Problem
If the muddle feels new, it is at least traditional. Every transformative technology spends its first years named for what it resembles. The automobile was the horseless carriage, radio was the wireless telegraph, and early films were photographed stage plays, until each found its own grammar. “Artificial intelligence” is this era’s horseless carriage, a name that describes the technology by pointing at something else, and the vagueness is now a sales instrument. A 2019 survey by a London venture firm found that a large share of European startups classified as AI companies showed no evidence of using the technology in any material way. The label sold; the mechanism was optional.
I know the objection forming: does the taxonomy actually matter, or is this pedantry? It matters for one reason, and the reason is money. These categories carry radically different costs, failure modes, and maintenance burdens. Rules automation fails loudly and fixes cheaply. Machine learning fails statistically and needs ongoing data care. Generative systems fail fluently, with the confidence problem this series keeps returning to, and need human oversight wired in from the start. A vendor who cannot or will not tell you which one you are buying is asking you to price all three risks blind.
Three Questions That Sort Everything
So borrow the test I ran for my contractor. Ask any vendor three questions. What does the system learn from, and what happens if it learns nothing? If the honest answer is that nothing is learned, you are buying automation; pay automation prices and be glad. What does it produce, a prediction, a classification, or new content? That places it on the map and tells you the failure mode. And third: walk me through what happens when it is wrong. The vendors with a crisp answer have deployed before. The vendors who insist it is rarely wrong have not, or worse, have and will not say.
The vocabulary will keep mutating; the marketing departments are well staffed. But the structure underneath, rules versus learning, prediction versus generation, moves slowly, and it is the structure that sets the price of being wrong. My contractor now asks the three questions before every renewal. He reports that the meetings have gotten shorter, and the discounts larger.
