AI does the first review. The teacher decides.

A music academy in Mexico City, with two branches and more than 200 students, had its teachers busy with the same thing every week: basic and introductory classes. That was 20 hours the academy wanted elsewhere, on the intermediate and advanced levels, where a good teacher makes the difference and where it is decided whether a student stays.

The obvious way out was to record the basic classes and let students watch them alone. The academy did not want that, and it was right. A video explains; it does not listen. Someone learning an instrument needs to be told what went wrong in their own attempt, not in someone else’s.

What we built follows four steps. The teacher sets the exercise and shows it, in the academy’s own method and not a generic syllabus. The student practises and records a video. The recording is checked against the criteria for that exercise. And progress is recorded, criterion by criterion.

The content arrives in the format each student is most comfortable with: video, reading, text or image. A mobile app shows progress through levels, so the student knows what is missing to move up and arrives at the next class prepared.

AI does the first review of each video and proposes feedback. Its job ends there. The teacher validates the proposal, corrects it if needed and assigns the next exercise.

That boundary is not a technical detail; it is the reason this works. If AI answered alone, the student might receive something that contradicts what their teacher teaches, and the teacher would stop trusting the tool at the first mistake. With the teacher at the end of every review, AI takes the repetitive work without taking the teacher’s voice. The student still gets their teacher’s feedback; it now arrives between classes instead of once a week.

Teacher-led introductory classes went from 20 hours to 5. The teachers’ workload did not fall: those hours moved to intermediate and advanced classes and jam sessions, which is exactly where the academy wanted them.

Attendance in the rainy months rose 20%. In June and July, when getting to the academy is hard, students kept practising and getting feedback in the app instead of losing the month.

Retention went from 60% to 80%. A student who can see their progress, and knows someone is waiting for their next video, has a reason to come back.

The system has run since January 2026. One of our engineers built it, working inside the academy for three months.

Three things, which hold outside music.

First, the useful question is not what AI can do, but which part of an expert’s work is repetitive and which part is judgment. The first can be done ahead; the second is left alone.

Second, the method belongs to the client. The criteria each exercise is reviewed against were written by the academy. Without that, any tool teaches what every other tool teaches.

Third, this repeats. Assign, practise, validate and master works the same for a padel shot as for an SME’s deliverable. That is why we are turning it into a product, named Chiron after Jason’s teacher. The first to arrive will be Bandeja, for padel academies, which goes into pilot in January 2027.