Automated Admissions Checks with AI for a University in Italy

The client ranks among the biggest private universities in Europe. It is based in Italy and has multiple branches. Candidates apply on an online admissions portal, uploading identity documents and a photo. The admissions office reviews each application and gets back to the candidate if anything is missing or incorrect. One branch handles admissions in a separate system that exposes an API. With two intakes a year, about 20,000 admissions need validating.

All of this was done manually. For each application, staff opened the uploaded ID, compared it with the data typed into the form, decided whether the document came in a format the Italian state recognises, and then reviewed the photo twice. First against the portrait rules (plain background, no bathroom selfie), then against the face on the ID.

The price of doing it by hand

• Staff locked up at peak. Some four to eight people were busy twice a year, for one to two months each time, right when the admissions office had the most on its plate.
• Eyeballing identities. About 20,000 identity matches a year, decided by whoever happened to be on shift.
• Wrong identities on file. One missed mismatch, a document format that should have been rejected or a photo of a different person ends up as a wrong identity in the student record.
• Candidates kept waiting. Each day of manual backlog is another day before a candidate learns that something is missing.
• Two channels, the same manual work. The same hand checks had to cover both the portal and the API-based system.

The original project documents no longer exist. What follows are working numbers and recollections of the delivery team, not audited measurements, so treat every figure as approximate.

• About 20,000 admissions validated each year over two intakes from 2022 to 2025, which adds up to more than 60,000 applications across the engagement.
• The robot got around 90-95% of applications right. The remainder, plus any case where information was missing or a field could not be extracted, was passed to people.
• Manual checking shifted to the robot. Staff deal only with incorrect or incomplete cases. The delivery team calls the savings enormous, although nobody measured them formally.
• A single validation standard applied to every applicant and every intake, in all branches and on both channels.
• More than 3 years of continuous operation, with our support, upgrades and maintenance the whole way.

Where admissions robots usually break

The failure tends to come at the worst time: week one of an intake, when volume jumps and a portal change, a Windows update on the machine the robot runs on or a new document format brings it to a halt. Our support covered both the robot and its infrastructure, and we handled corner cases as they surfaced, so every intake began with a robot that worked.