Institutions ask us about AI for student success and expect to hear about new platforms. The honest answer is that three datasets they already have do most of the work.
The three signals
Attendance, by week rather than by term. A drop over three consecutive weeks is the single strongest early signal we have seen.
Internal assessment, especially a fall between the first and second assessment rather than a low absolute score.
LMS activity: not grades, but whether the student opens the material at all. Silence for two weeks is informative.
What to do with it
A weekly list per mentor of students whose signals have moved, with the reason. Not a risk score, which nobody acts on, but a sentence: attendance fell from 85% to 60% over three weeks.
The constraint that matters
Student data must stay inside the institution's infrastructure, be minimised, and be visible only to the mentors who act on it. We put this in writing before touching anything.