Finding an outlier is a statistical step, but deciding what to do with it is a judgement about what the data means.
Every option changes the results, so the choice has to be deliberate and recorded.
The first question is always whether the value is an error or a real event.
Errors often have a recognisable cause: a misplaced decimal point, a mix of units, or a placeholder such as 999.
A mix of units, such as one weight in grams among weights in kilograms, can be corrected with confidence once spotted.
A real event that matters to the question, such as a fraudulent payment, should be kept and studied rather than removed.
A real event that distorts a summary can be handled by reporting the median instead of the mean.
Capping, also called winsorising, keeps every row but limits how far any one value can pull.
Taking logarithms shrinks large values more than small ones, which reduces their pull without removing them.
Removing outliers makes results look tidier and less variable than the real process is.
Removing them in repeated rounds is a trap, because each round creates new extremes from what is left.
Treatment should be decided before seeing how it changes the headline result, or it becomes a way to pick an answer.
In practice teams write the treatment down, such as capping at the 99th percentile, and apply it on every run.
Reporting how many values each rule changed lets readers see how much the cleaning shaped the result.
The common misconception is that outliers are always bad data, when they are often the most important rows in the table.