Sorting incoming messages into piles is one of those tasks that looks trivial until you have done it several hundred times. It never gets faster, it is repetitive enough to make errors, and it is the kind of work where the tenth item in a pile gets filed carelessly.
How the classification works
You describe what each category looks like by giving it the words and phrases that appear in that kind of message. When you paste something in, the tool compares the two and picks the closest match. This is not a trained model with knowledge of your business; it is matching against the examples you supplied, which has a useful consequence: the quality of the result depends almost entirely on how well you set it up.
Setting it up well
The words you choose are doing all the work, so choose the ones that genuinely appear. If your invoices say “remittance advice” and “invoice number”, those belong in your billing category, not a generic word like “payment” that appears everywhere. Three or four specific phrases beat a list of ten broad ones.
Overlap is the usual reason a message lands in the wrong place. If two categories both contain “please” or “account”, expect ambiguity. Give each category the terms that distinguish it rather than the ones it shares.
Where it works well
Short text with clear vocabulary, which in practice means support tickets, form submissions, subject lines, survey answers and short enquiries. Those have a small number of recurring phrases and a fairly stable structure, which is exactly the condition where matching works well.
Long documents are the opposite case. A thousand words contains far more incidental vocabulary, so the signal weakens and the result becomes less predictable. Extract the relevant sentence first and classify that instead.
Getting the edge cases right
Expect some messages to be genuinely ambiguous, and resist the urge to force them. A complaint that mentions a refund is a complaint. Most systems work better with a clear “not sure” option than with a confident wrong answer, because the cost of a mistake is usually higher than the cost of a second look.
Because it runs in the browser, you can paste customer messages containing personal details without sending them anywhere, which is a genuine advantage over hosted alternatives.
Try next
Related guide: url slug generator seo free
Building a category set that holds up
The quality of the whole system is decided by the categories, so it is worth resisting the urge to define eight of them. Fewer categories that people actually use beat a complete taxonomy nobody trusts. Start with three or four, look at what the system actually misfiles, and only then consider adding more.
Give each category at least four or five genuinely distinct phrases, and include the awkward phrasings rather than only the formal ones. Real messages contain typos, lowercase typing and slang, and a category built only from tidy language will miss them.
Then check for overlap between categories, because that is where misfiling comes from. If two categories both include broad words like please, account or help, expect ambiguity on almost every item that contains them. Each category should be defined by what makes it different rather than by what it has in common with everything else.
Finally, decide what happens to something that matches nothing well. A deliberate unsure option is more useful than forcing every input into the nearest category, because the cost of a wrong confident answer is usually higher than the cost of a second look.