What is AI classification?
Classification means the model assigns an item to one of a list of labels you defined. It does not write the customer reply. It decides which pile the message belongs in, so the right next step can run.
In plain English
Labels might be: schedule, price, complaint, vendor, unknown. You write the labels and one sentence each. The model picks. Unknown is a real label. If you do not offer it, the model will force a bad fit. Write the labels on a card before you let a model use them. Each label needs an example and a 'not this.' Overlapping labels like 'angry' and 'urgent' will be applied at random. Start with a handful. Route the unsure pile to a person. A classifier that must always pick is how a complaint gets filed as a compliment. Recount the labels after a month. If one label has three items and another has three hundred, the set is wrong, not the inbox. Merge labels nobody can tell apart. Ambiguous buckets create fake precision.
How an owner would use it
Start with labels that match how the work is already divided. Route unknown and anything safety-related to a person. Check a day of labels before you let a label trigger a customer email.
Common mistakes
Dozens of overlapping labels create random sorting. Owners skip the unknown bucket and then automate the wrong pile. Classification is not extraction. Extraction pulls the address out. Classification only names the pile. Do not let a classifier be the only safety check on a roadside message.
A worked example
A towing company
Injury notes never land in the quote pile
This is an illustration, not a customer result. Inbound texts are labeled impound question, private property tow, or needs a person. A text that mentions an injury is labeled needs a person even if it also asks for a price. The quote template does not send.
Related terms
Prompts and guides
Questions owners ask
What is classification?
Classification is labeling an input with a category you defined so the next step knows which path to take.
How do I set up classification for a shared inbox?
Write a short list of labels, include unknown, and review a day of results before a label is allowed to send anything.
What if two labels both fit?
Prefer the safer pile. A message about money and a complaint should follow the complaint path if that path gets a person. Say that rule in the prompt.
Product guides are on AI apps for business.