Pilot results
What we know, how we know it, and what we don’t know yet.
Do students who work on a notion with Coach retain it better? Here is what the pilot classes have shown so far, how we know it, and what they have not shown yet. The page will grow as the pilots progress.
Research work begun in 2023
Fed by teachers’ feedback, then built and iterated in class with teachers and students, mostly in mathematics.
Teachers’ feedback from the classroom
In pilot classes, teachers reported better retention of the notions. Detail below.
A published use case
Method and architecture described in our 2026 use case, published in AI for Good’s Innovate for Impact programme (ITU).
No independent trial yet
We are looking for an academic partner to design one.
Pilot classes
What the teachers of the pilot classes reported.
Presented in the format we will use for every study: question, context, participants, what was observed, measure, limits, source.
The research findings the method relies on, and what they change in Coach.
Self-determination theory
Autonomy, competence and relatedness support more durable motivation. Hence the student’s choice over the form of help, and stages they pass when they are ready.
Cognitive load theory
Effective help isolates one obstacle at a time. Hence an interface that shows one thing, and a correction that names one step.
Active recall
Remembering with effort anchors better than rereading. Hence the exam without help before moving on.
Error diagnosis
Three types of difficulty (modelling, reasoning, execution) to work on separately. Inhibition, freezing in front of the problem, is their consequence when the other three don’t hold. Described in the AI for Good use case.
Mistakes as information
A sense of control supports engagement. Hence a correction that describes the mistake and gives the solution, without a red cross.
Preferences the student can revise
Fixed “learning styles” are not supported by research. Coach gathers preferences (how to enter a lesson, which help when stuck), and the student can regenerate them.
Publication.
Our method, our architecture and the feedback above are described in a use case published on the AI for Good platform of the International Telecommunication Union. The text is written by AiCole.
2026 use case · Pedagogical Infrastructure for Academic Mentorship and Education
Category: generative AI for creativity, education and public services · France.
What you will find here as the pilots progress.
Every pilot published with its question, its protocol, its limits and what failed.
No percentage without the sample, the method and the comparison beside it.
The teachers concerned review what is written about their class.
No individual student data is published.
What we do not claim.
“Coach detects whether your child is visual, auditory or kinaesthetic.”
Coach gathers what helps this student enter a lesson and get going again when stuck. These are preferences, which they can change.
“The AI replaces the teacher’s explanation.”
Coach extends classroom work between two lessons. The content follows the Ministry of Education’s official programmes, published in the Bulletin officiel.
“Guaranteed results.”
We publish our pilot classes’ feedback with its limits. An observation reported by teachers does not prove general efficacy.
“Official Brevet and Bac preparation.”
Coach’s exams and challenges are internal checks within the journey. The Brevet and Bac practice papers are AI-generated training papers; the official past papers remain the reference.
Researcher, teacher, school: let’s work together.
The pilot classes’ feedback and its limits are shared on request. We are looking for an academic partner to design an independent evaluation.