Most business decisions are classification, so stop paying for generation

Most of what businesses ask a language model to do is classification, and most of them use it as a text generator. Is this lead serious? Which team owns this email? Is this refund request normal? Each is a choice from a closed set, repeated hundreds or thousands of times a day. The answer wanted is a label, not an essay.

Classification returns a distribution, not prose

Generation produces open-ended text. Classification picks among options you define in advance. There are three useful shapes: a yes or no, a choice among several categories, and an ordinal score where the options form a scale, such as zero for no interest up to three for leading the round.

The natural output of a classifier is a probability distribution over those options. A yes-or-no question might return 94% yes and 6% no. A choice question returns one probability per category, and they sum to 100%. An ordinal score can be reported as an expected value on the scale. A system that works this way never answers in sentences, so there is nothing to parse and nothing to drift.

Calibration matters more than eloquence

The distribution is useful only if the numbers mean something. A calibrated model that says 90% is right about nine times in ten. That property is what lets you route: act automatically above a threshold, send everything below it to a person.

A model that writes a fluent, well-argued paragraph and cannot say how sure it is gives you nothing to route on. Eloquence reads as confidence either way. Calibration is the property you can test, by running a labelled sample and checking how often the 90% answers were correct.

Pick the model by cost and latency per decision

Accuracy is one axis. For a decision made thousands of times a day, latency and cost per decision are the other two, and they compound. A second of waiting becomes a queue. A few cents become a budget line.

A smaller, faster model that is slightly less accurate can be the right choice for classification, provided the uncertain cases reach a human. Reserve the large reasoning model for work where a wrong answer is expensive or the question is open. This is the same logic as staffing: you route by what each decision deserves.

System 1 and System 2, used precisely

Kahneman's System 1 is fast, automatic and low effort. System 2 is slow, deliberate and effortful. Applied to models, the distinction is about compute per decision. A classifier that returns one answer and a probability is spending System 1 effort. A reasoning model that writes out a long chain of thought before answering is spending System 2 effort.

The mapping says nothing about intelligence, and System 1 is not a better version of System 2. It is a cheaper one, with the same failure modes: it is fast, and it can be confidently wrong. That is why the threshold and the human fallback are part of the design.

A model called Jev reportedly follows this design: structured input in, a distribution over your options out, no chat. I haven't tested it, and the numbers circulating come from its maker and a few testers, so I treat them as claims. In the one comparison I watched, the largest model was the most accurate and the fast one was close behind. That is consistent with the tradeoff above and proves nothing beyond it.

Try it with Claude tomorrow

Take one sorting task you do by hand. Paste a real message and ask for a label and a percentage:

Classify the message as serious, not serious or unclear. After the label, write in percent how sure you are. No explanation.

Read the low-confidence cases yourself and let the rest through.

Where it breaks

The percentage Claude states is a rough signal, not a measured probability. Treat 95 and 60 as very different, and 87 and 84 as the same. And a classification is only as good as its options. If you cannot define the closed set in a sentence, the task is not classification yet, and it belongs with the slower model.