An early-warning AI that flags at-risk students 6 weeks before they drop out

Industry

EdTech

Market

India

Client

Education provider in India

Sponsor

Director

What we built

Student retention early-warning AI

The situation

For the client, an education provider in India, dropout was not an abstract metric. The Director put the cost at ₹4 crore a year in lost tuition and reputation damage. Each student who left represented fees that stopped, a seat that went empty, and a story about the institution that travelled.

The frustrating part was that dropout rarely comes without warning. Students disengage gradually. They log in less, submit late, miss sessions, fall behind on assessments, and sometimes fall behind on fees. The signals existed across the provider's systems, but nobody saw them together in time to act. By the time a student was visibly gone, the conversation that might have kept them was weeks too late.

What we built

An early-warning system that brings the signals together, identifies students at risk while there is still time, and puts them in front of the people who can help.

Signals from the systems that already exist. Attendance, learning-platform activity, assignment submission, assessment performance, support requests and payment status are combined into one picture per student, updated continuously.

Risk scoring with reasons. A model trained on the provider's own history estimates each student's risk of dropping out and explains why: "attendance down sharply over three weeks, two assignments missed, no platform activity since Tuesday." Counsellors act on reasons, not scores.

An intervention workflow. At-risk students appear in a prioritised list for counsellors and faculty mentors, with a suggested next step and a record of every outreach attempt and outcome. A flag is not the end of the process; a conversation is.

Learning from outcomes. Which students were flagged, who was contacted, and who stayed or left feeds back into the model, so the warnings get sharper each term.

Student data handled carefully. The system uses only the data needed for retention, with access limited to staff responsible for student support, in line with India's data protection requirements.

How it went live

The model was built and validated on past terms first: would it have flagged the students who actually left, and how early? That established the lead time before any live student was scored. The first live semester ran with counsellors working the flagged list alongside their usual caseload, so the team could see the warnings translate into conversations.

Results

The system flags at-risk students six weeks before they disappear, early enough for a counsellor to make a real difference. In the first semester, the provider retained 23 students who would otherwise have left, and the Director reported that the system paid for itself in month three.

What made the difference

Explanations counsellors can use. A reason opens a conversation; a score does not.

Validated on history before going live. Proving the lead time on past terms is what made the institution trust the flags.

Built around the intervention. The value is in the outreach workflow and the follow-through, not the prediction alone.