We are still early
3 min read
Spend enough time around people building with AI and it starts to feel as if everyone is doing it. Someone has agents running through a codebase. Someone else launched an app over the weekend. You open your feed and find another product doing something you were thinking about building.
I find the more useful question is who could use something I know how to build. That gives me a person to talk to and a piece of work to understand.
A technology circle in São Paulo may make AI feel ordinary. In Cetic.br's TIC Domicílios 2025 survey, 69% of Class A internet users aged ten or older had used generative AI, compared with 16% in Classes D and E. Among internet users in its elementary-education group who had not used AI, 65% cited a lack of skill as a reason.
I would be careful about offering those people an empty chat box and expecting them to work out what to do. A useful opportunity might involve choosing the task with them, preparing their material and making the result easy to use on their phone. The adoption gap gives me a reason to look beyond my feed. I still need to find someone who wants help with a particular job.
Take the guitar learner who keeps practicing songs they already know. They have notes on the chord changes they struggle with, but rarely use those notes to decide what to practice. I could build a small practice tool for myself. To make something useful for other people, I would look for a teacher willing to try it with a few students who have the same difficulty.
Imagine those students understand the exercises during a lesson, then lose track of what to repeat at home. They already have a teacher and material. The missing help is turning that material into today's practice. A teacher who knows them can introduce the tool, explain its limits and notice when it sends someone in the wrong direction.
For a first version, I would ask an agent to turn the teacher's existing notes into a twenty-minute session. It would name the chord change, show the agreed exercise and give the learner somewhere to record what still felt difficult. The teacher could check the session before sharing it. I would leave automatic assessment of their playing for another day; it adds a harder problem before we know whether anyone follows the plan.
Then I would watch someone open it on their own phone. Can they find the exercise? Do they understand what to record? Does starting a session require another explanation from me? If the learner has to create an account, copy notes between apps and invent a prompt, I may have given them more work than the practice plan removes.
After a week, the useful evidence would be the sessions they attempted and the notes they brought back. Perhaps twenty minutes is too long. Perhaps they need one exercise and a recording from the teacher. I would change the tool around that before adding more features.
This could stay a small tool shared by one teacher. If I wanted to sell it, I would also need to find out who pays and whether preparing and checking the sessions leaves the teacher enough time to make it worthwhile. Low adoption cannot answer either question.
Before starting another version, I would ask the teacher to bring the next week's notes and invite the same learners to use it again.