Admissions Chatbot Research: A View from Ireland

What the research establishes

Published research shows that a conversational text-message assistant can help admitted applicants finish their paperwork before term starts. It does not show that a chatbot on a college website increases enrolments. The strongest trials are American, test messages sent by the institution to admitted students, and use non-generative AI.

A note on method: no paid keyword-volume tool was available for this piece, so the topic was scoped from public search results. Every study below was traced to its source. No figure comes from Skolbot or any client.

Georgia State: the one large trial of an admissions chatbot

Page and Gehlbach remains the reference study. In 2016 Georgia State deployed "Pounce", a text-message assistant built with a vendor, to guide admitted students through the summer: financial aid forms, immunisation records, transcripts, course registration.

Admitted students were randomly assigned to receive the messages or not. Among those who had committed to the university, on-time enrolment rose by 3.3 percentage points. The authors say that is comparable to heavier interventions tested earlier, with far less staff effort.

Two limits matter. The effect concerns finishing tasks after admission, not the decision to apply or to choose a college. And the trial dates from 2016, well before today's generative assistants, so its result does not transfer directly to an open-ended generative chatbot.

Summer melt: targeted texts work where help is scarce

Before chatbots, Castleman and Page tested personalised texts on low-income school leavers. Results varied by site: positive in three of four districts, null in Boston, where students already had plenty of school and nonprofit support.

For an institution, the lesson is that the effect of an automated message depends on what the applicant would have received anyway. Where an adviser already answers fast, the added value is small.

Targeting changes the result

Nurshatayeva, Page, White and Gehlbach replicated the Georgia State design at another university (4,442 students). The overall effect on loan acceptance was about 4 points, rising to about 8 points for first-generation students, with about 3 points on course registration and enrolment.

The authors' practical message is that proactive outreach pays most for applicants with the least family or school backing. They also report that the community-college strand of their project could not be run properly because mobile numbers were missing. Contact data quality limits a study just as it limits a rollout.

Beyond admissions: course chatbots show a similar pattern

Meyer, Page and co-authors tested a non-generative chatbot in undergraduate courses through pre-registered randomised trials. Messages raised final grades and use of academic support. Effects were broadly similar across demographic groups, apart from women in one microeconomics course, whose grades rose by about seven percentage points.

This concerns enrolled students. It shows the "short message, specific task, right moment" pattern recurring at several stages, but it measures nothing about converting a prospect.

The counter-example: not every nudge works

Oreopoulos and Petronijevic followed nearly 25,000 students across three campuses. Online coaching and text messages improved well-being and study time slightly, but no intervention significantly changed academic outcomes.

That tempers the other findings. Messages seem to work best on a precise, dated administrative task such as submitting a form, less well on diffuse behaviour change. Page and colleagues reach a similar conclusion in their analysis of when nudging works.

What the research does not say for Ireland

It says nothing about a generative chatbot on an Irish college or university website answering prospects' questions. We found no published randomised trial, in Ireland or elsewhere, measuring its effect on applications or enrolments.

The Irish route is different from the US one. Most undergraduates apply through the CAO, where Leaving Certificate points and a rounds-based offer system decide placement. Between a CAO offer and registration there are acceptance deadlines, grant applications and accommodation to sort out. Those are the dated steps where the research suggests a message may help, though no Irish trial has tested it.

Finally, the studies describe assistants designed and monitored by research teams. A product deployed without that rigour has no guaranteed result.

What it means for an Irish institution

The legal frame matters more than the measured effects. The GDPR and the Data Protection Act 2018 apply to every message sent to an applicant, and the Data Protection Commission expects transparency, data minimisation and a lawful basis. Many applicants are under 18, and Ireland sets the digital age of consent at 16, so the information must be clear and consent handled with care.

The EU AI Act applies directly in Ireland and classes systems that determine access to education as high risk. An assistant that informs is not a system that decides, but a tool that ranks applications comes close. Keep admission decisions with people. For programme quality, QQI is the reference body.

Four practical points follow:

• Aim at dated tasks: missing documents, open day dates, acceptance deadlines, interview slots. That is where trials find effects.
• Target the least-supported applicants rather than sending everyone the same message.
• Fix contact data and consent before any rollout.
• Measure with a control group: randomly hold back a sample of applicants from the nudge, then compare completion and enrolment.

For next steps, read our guide to AI agents for student recruitment, the definition of an AI admissions agent, the piece on yield management and post-offer drop-out and the AI chatbot guide for student recruitment.