A low-risk side-project method: validate the problem before building an AI product

Start with a recurring problem, not a model. Talk to 5–10 people in a specific group about how they solve the task today. Ask for real examples, not hypothetical enthusiasm.

Offer a manual or no-code version first, then measure repeat use, willingness to pay, and outcomes. Automate only the parts that repeat. AI makes prototypes faster, but speed does not prove demand.

What unnecessarily manual task do you see repeatedly in your industry?

"Real examples, not hypotheticals" is the whole method in five words, Arenaman. People are hopeless at predicting what they'd pay for but brutally honest about what annoyed them last Tuesday.

One question I add in those 5–10 chats: "would you pay ₹500 right now to make this problem disappear?" The wince (or the instant yes) is worth more than ten survey forms.

Did any interview ever kill a project you were excited about? Those stories are the good stuff.