Knowing where a tool stops being useful is part of using it well. A few categories come up repeatedly.

When you cannot check the answer

If you have no way to verify the output and the cost of being wrong is real, you are not saving time, you are moving risk somewhere you cannot see it. This applies to figures you will publish, legal or medical specifics, and anything a customer will rely on.

When the task is the thinking

Some writing exists to force you to work something out – a strategy memo, a difficult performance conversation, a decision you are unsure about. Getting a draft in ten seconds skips the part that had the value in it.

When it needs to be genuinely yours

Condolences. An apology. Feedback to someone who reports to you. People can tell, and the cost of being caught outsourcing it is much higher than the time saved.

When the input does not exist

A model cannot tell you what your customers think, what happened in a meeting it has no record of, or what your numbers were last quarter. Asking anyway produces confident text with nothing behind it, which is worse than no answer.

When a simpler tool would do

• A find and replace is faster than a prompt.
• A spreadsheet formula is more reliable than a model doing arithmetic.
• A saved template beats regenerating the same document every time.

The useful question is not can AI do this, but is this the cheapest correct way to get it done.