Deciding what to hand to a language model is usually approached as a question of difficulty. The more reliable test concerns how you will know whether the answer is right.
Compare the cost of checking against the cost of doing
A task is worth delegating when verifying the result is substantially cheaper than producing it.
Drafting a summary of a document you have read qualifies, because you can check it in a glance against knowledge you already hold.
Researching a subject you know nothing about does not, because you have no basis for judging the answer and checking it means doing the work.
Prefer tasks where the source is in front of you
Work performed on text you supply — extracting, reformatting, summarising, comparing — can be verified against the input.
Work performed from the model's own knowledge has no accompanying source, and confident wrong answers in that mode are the most common failure users encounter.
Supplying the material rather than asking the model to recall it converts an unverifiable task into a verifiable one, and it is the single highest-value habit in using these tools.
Watch for tasks with one exact answer
Arithmetic, counting, precise citation and anything requiring an exact identifier are poorly suited, because the process generates plausible sequences rather than computing them.
Tasks with a range of acceptable answers — phrasing, structure, first drafts, alternatives to consider — suit the mechanism directly.
Where a tool can call a calculator or a search index, the exact tasks become viable again, because the exactness is being handled by something that computes.
Consider what a wrong answer costs
A wrong suggestion in a brainstorm costs nothing, since you discard it.
A wrong figure that enters a report and is cited later costs a great deal, and the error is discovered by someone else.
The reasonable rule is to delegate freely where mistakes are caught immediately by you and to delegate cautiously where they are caught later by somebody downstream.
Test on work you have already completed
Before adopting a tool for a recurring task, run it against last month's work where you know the correct output.
This gives a realistic error rate on your actual material, which is a far better guide than any demonstration or benchmark.
It also shows the shape of the errors, and knowing which kind of mistake to look for is most of what makes the checking step fast enough to be worth doing.