What AI is useful for

AI is most useful when the work has many possible answers and you can inspect the result. It can explain unfamiliar code, draft a first version, compare approaches, transform data, and help you explore a problem.

Notice the second part: you can inspect the result. Asking AI to rename a function is low risk because you can read the diff and run the tests. Asking it whether a medical symptom is harmless has a very different cost when the answer is wrong.

I use three questions before reaching for AI:

• Can I describe the result I want?
• Can I give it the information needed for the task?
• Can I check whether the result is correct?

If the answer to all three is yes, AI is often a good fit. If you cannot verify the output, the model may still help you explore, but it should not make the final decision.

Four useful kinds of work

Transformation changes something you already have. You can ask for a summary, a translation, a refactor, or a different data format. The original gives you something concrete to compare against.

Generation creates a first draft. This is useful for code, documentation, test cases, emails, and plans. Treat the output as raw material, not finished work.

Exploration helps you see options. Ask for three database designs, likely causes of an error, or questions you have not considered. Here the value is expanding the search space.

Review looks for omissions or problems. A model can inspect a diff, challenge an assumption, or suggest edge cases. It is useful as an extra reviewer, but it cannot prove that nothing is wrong.

Where AI is a poor fit

Be careful when the task requires a guaranteed fact, private data you cannot share, a final decision with serious consequences, or an action that is difficult to undo. A fluent answer can still be invented, outdated, or based on a misunderstood requirement.

AI also has little value when doing the task yourself is faster than describing and checking it. A two-line edit may not need a conversation.

Try this

Take one task from your current project. Classify it as transformation, generation, exploration, or review. Then write down how you would verify the result. If you cannot name a check, reduce the scope until you can.