How do I decide when to use AI, and how much to hand over?
Sometimes the best prompt is no prompt.
AI could have a go at all three. It's a good idea for maybe one of them.
Made-up examples.
Drew Breunig puts it neatly. Our job is to choose when to be deterministic and when to be probabilistic.
Deterministic: the same input always gives the same output. A spreadsheet formula, a calculator, a script.
Probabilistic: the output involves some chance, so you might get a different answer each time. That's how AI models write (see "AI isn't looking anything up").
If there's one right answer and clear rules for getting it, you want deterministic.
Sales total: one right answer, clear rules. Use a formula. If you don't know the formula, ask the AI to write it for you, then use the formula.
Customer reply: no single right answer, and you can easily judge whether it's good. A great job for AI, with you reading it before it goes.
Contract advice: a wrong answer is costly, and you can't easily tell if it's wrong. AI can help you understand the contract, but the answer needs someone qualified.
Anthropic's AI Fluency course describes three ways of working with AI:
Each step hands over more, and each needs more trust and more checking built in.
Anthropic tracks how people use Claude. Between December 2024 and August 2025, the share of conversations where people handed over a whole task, rather than working on it together, rose from 27% to 39%.
That's not wrong. But handing over more is only safe if you can still tell when the result is wrong.
Before you open the chat, ask: is there one right answer with clear rules? If so, get the AI to write you a formula or script, not to do the sum.
Match how much you hand over to how well you can check it. Easy to check means hand over more. Hard to check means stay involved.
If it's costly to get wrong and hard to check, use AI to understand the problem, and a person to make the call.