Make your agent fail in private: the ten-question test set

Every AI agent looks brilliant in the demo. But in the demo, you ask the question you designed it to answer. Your customers will ask something else — at eleven at night, in the wrong language, about the one product you discontinued six months ago. The cheapest hour in the whole project is the hour you spend before launch trying to break it yourself.

Write ten questions down in a document. Not in your head — in a document, so that next month, after you change something, you can run the exact same ten.

• Three easy ones — the questions you built it for: price, opening hours, do you deliver to X. If it wobbles here, nothing else matters.
• Two it should refuse — a discount request, a legal or medical question. The right answer is “let me get a human”, not a confident guess.
• Two with information missing — half a sentence, a bare “how much?”. A good agent asks one clarifying question instead of inventing an answer.
• Two hostile ones — a rude complaint, and someone trying to talk it out of following its instructions.
• One in another language — even if you never advertised it. Someone will.

Mark each answer pass or fail, fix what failed, then run all ten again. Ten minutes per round — and “I think it works” turns into something you can actually put in front of a customer.