When it's better not to use AI

The treasurer of a hiking club has to split the costs of a weekend among twenty-three participants, with different rates depending on age and a deposit already paid by some. He pastes everything into an AI and asks for the amount to charge each person. It answers with an impeccable table. Three lines further down, he notices that the total doesn't match the lodge's invoice. He's lost twenty minutes, and he still has to recalculate everything.

His request was fine; it was the choice of tool that wasn't. AI does a lot of good, but there are cases where avoiding it is the right decision. Six situations come up most often.

When you can't verify the result

This principle holds for everything else: the more an error costs and the less able you are to spot it, the more you should be wary. You'll spot a typo in an email about your own field. You'll spot far less an error in a lease clause or a medication dosage. A wrong but well-written answer is the most dangerous, because nothing flags it.

The practical rule: use AI only for what you could check yourself, or what you will check against a reliable source. For a medical, legal, or tax matter that puts your health or your money at stake, it can help you prepare your understanding and your questions, and the decision stays with a professional. Checking an answer describes how to proceed when AI can still be of use.

When the task takes two minutes by hand

Writing a "thanks, received" or moving an appointment gains nothing from going through an AI. You write the prompt, you read the answer, you touch it up, and you end up with the same email as if you'd typed it yourself. A good test: if you can state the result in one sentence in your head, write it directly.

When the effort is the point

A student who has an essay written for them doesn't learn to argue. A person who asks for a speech for their sister's wedding gets a text that could be anyone's. When what matters is what you retain or what you express, keep the pen. You can, however, use the AI as a partner: to ask you questions, reread your draft, point out confusing passages. Students will find uses in our article on prompts for students.

When the message calls for real presence

Condolences, a sincere apology, telling a longtime colleague that a contract is ending: the recipient can sense the formula. A clumsy, personal text is better than a perfect, generic one. If you're stuck on the first sentence, ask the AI for ways to get started, then write your own message.

When the data must not leave

A relative's medical file, a payslip, an ongoing dispute, customer information, or trade secrets: depending on the tool, the plan, and your settings, what you send may be stored. Before pasting anything, go through the list of what not to entrust to a chatbot. If your task involves personal data in a professional setting, look at your organization's rules. The CNIL, the French data protection authority, also publishes recommendations.

When you need an exact result

Long calculations, totals, splits, counting dates backward: an AI can make mistakes on these while presenting a polished result. Some tools can run a calculation for you, others can't, and you don't always know which one is at work. For a number used to invoice, to file a return, or to claim money, use a spreadsheet.

The same remark applies to recent information: an AI without access to a search doesn't know what happened after its training, and you aren't always sure the search is turned on.

What the AI can do instead

Setting the AI aside for a task doesn't make it useless around that task. The treasurer from the start would have saved time by asking it for the formula, not the result:

I have a table with the names of 23 participants in column A, their category (adult, child, member) in B, and the deposit already paid in C. The rates are: adult €[X], child €[Y], member €[Z]. Write the formula to put in D to calculate the balance due, explain it in two lines, and tell me how to check that the total matches the invoice.

He applies the formula, checks the sum, and the calculation stays in a spreadsheet. The prompts for Excel show other examples of this division of roles.

The test before you dive in

Before your next request, ask yourself these questions:

1. If the answer is wrong, who pays the price, and how would I know?
2. Would I have finished faster doing it myself?
3. Does this work have to come from me for it to count?
4. Is there anything in my request that I wouldn't send to a stranger?

A single worrying answer is enough to change approach: do the task yourself, or limit the AI to preparation.