Ask whether the task has one correct answer
If it does, and a deterministic tool exists, use the deterministic tool. Counting characters, converting case, sorting a list and validating JSON all have exact answers, and a language model is a slower and less reliable way to get them.
Language models earn their place where the answer is a judgement: drafting, rephrasing, summarising, explaining, and suggesting options.
Ask how expensive an error would be
For low-stakes drafting, an occasional error costs a moment of editing. For financial, medical or legal output, an error is expensive and verification is mandatory.
Match the amount of checking to the stakes, and decide that before you start rather than after you have a fluent draft in front of you.
Ask whether you will do it again
A one-off task rarely justifies building a workflow. A weekly task justifies working out the exact prompt or process once and reusing it.
Automation pays back on repetition, not on novelty.
A quick decision list
Exact, deterministic result needed: use a utility. Text to be produced or reshaped: use a writing assistant. Image or document to be read: use recognition. Code to be explained or drafted: use a code assistant.
Task requires knowing something private, current or specific to your organisation: supply that information yourself, because the tool does not have it.