An AI classifier is a model that assigns a category to what it receives: an email, a text or a photo. A spam filter, a ticket router or a language detector are classifiers. The newer ones pick between options you define, or answer yes or no, with the probability of each answer.
A question from the app. Pick an answer.
Message: "How much is it to move to the Pro plan?" On your team, prices and plans go to Sales. Jev sent it to Billing, with confidence 0.91 (example number). This is how you asked:
What do you try first?
The threshold decides whether you act, not which team it picks. It would still go to Billing.
With just a name, Jev does not know where Sales ends. Try descriptions that separate prices from charges already made.
Switching models does not clarify your rules. First describe what goes in each team and check the result.
In a test, a model read "A tomato." and answered two yes or no questions: "It is a fruit.", 0.65; "It is a vegetable.", 0.85. Both come out yes and it is not a contradiction: each question is answered on its own, without looking at the other.
If you need a single answer, ask for a choice between options: then the probabilities are split and add up to 1. With fruit and vegetable it came out 50 and 50. In botany it is a fruit and in the kitchen, a vegetable: if you care about one of the two, say so in the question. And a tie is not for your code to pick at random.
An automatic grader gets "Capital of France? Student's answer: Madrid" and the question "The answer is correct.". In the test it gave 0.57: more yes than no, for a wrong answer. Adding "Correct answer: Paris." to the text brought it down to 0.04.
Do not count on it knowing the fact. If the decision depends on a fact, put the fact in the text. It is the same idea as RAG: look up what is needed first and hand it to the model.
In a test, a ticket router sent "Excel closes by itself every time I open a file." to hardware, with 0.96. "The Excel app closes by itself…" went to software, with 0.98. Two ways of saying the same thing, two different teams, both very sure.
A high probability is not a guarantee (calibrated probability explains why). Try describing each option: what goes in, what does not and an example, like "Software: programs and apps, including Excel or Word" instead of a bare "software". And check the result with your own real cases.
Classifiers are among the oldest things in machine learning: a spam filter is trained for a single task. The new ones, like Jev or OpenAI's Decisions API, get the options in each request, so the same model works for different tasks. There are also open models that run on your computer.