You have probably had both experiences by now. Once, the assistant turned forty minutes of note-tidying into two, and you thought: right, this is the thing everybody means. Another time it produced a paragraph so bland you rewrote it from scratch, and that tab has been open and unused ever since.

The difference is rarely the tool, and almost never the wording of your first sentence. It is the shape of the task. Some jobs sit exactly where a language model is strong and some sit exactly where it is weak, and the two look identical from the outside until you have seen enough of each.

The tasks it does well

Every one of these has the same shape: you already have the information, and the work is turning it into something else. The model is not being asked to know anything you have not given it.

  • Reshaping text you supply. A long thread into five lines. Meeting notes into a list of decisions. A rambling draft into something a director will actually read.
  • Changing register. The same message written for a colleague, for a customer, and for a legal review — three versions in the time it took you to write one.
  • First drafts of things you were going to edit anyway. The job description, the FAQ answer, the polite version of the email you actually want to send.
  • Explaining something at your level. A clause, a formula, an acronym in somebody else's deck, a paragraph from a team whose work you do not follow.
  • Structured extraction. Forty survey responses into a table of themes; an invoice thread into a list of dates and amounts, in the columns you asked for.
  • Being a first reader. "What questions will this raise?" is a genuinely useful thing to ask before you send anything into a room full of people.

The tasks it does badly

  • Anything that turns on a fact you have not given it. Your renewal dates, last quarter's real numbers, what your head of sales agreed on Tuesday. It will answer anyway, and the answer will read exactly like the truth.
  • Arithmetic. A model can lay out a calculation beautifully and still get the number wrong. Check every figure that ends up in a document with your name on it.
  • Judgement calls with real stakes. Who to hire, whether to escalate, what a contract exposes you to. It can list considerations. It cannot carry the consequence.
  • Anything where being confidently wrong is expensive and you would not personally catch the error.
  • Work that needs your specific taste. It writes the median version of everything. Where sounding like everybody else is the failure, this is the wrong tool.