An AI Chatbot for Universities Won’t Recruit a Single Student

Half of 7,489 admitted students got a text-message bot. The other half got the usual emails and letters. Among the students who had not yet committed to the university, the bot moved enrollment by nothing at all. That result comes from the only published randomised controlled trial of an AI chatbot for universities, and it is the most useful finding in the entire higher-ed AI literature.

Not because the thing failed. Because of exactly where it worked and where it didn't. Almost every campus procurement conversation we've watched gets this backwards: the bot goes on the admissions page, aimed at prospects. The evidence says that's the one place it does nothing.

The Georgia State experiment, in the numbers that matter

Summer 2016. Georgia State University, working with AdmitHub, built a conversational assistant called Pounce and gave it to a randomly chosen half of its admitted cohort. Control students got GSU's standard communication. Treatment students got a bot that answered questions by text and nudged them through the pre-enrollment gauntlet: FAFSA, final transcripts, immunisation records, loan counselling, orientation.

Lindsay Page and Hunter Gehlbach published the results in AERA Open in 2017. Students who had already committed to GSU and got Pounce were 3.3 percentage points more likely to actually enroll — a 21% cut in summer melt. Applied across a freshman class of roughly 3,500, that's about 116 students who would otherwise have disappeared between May and September.

The task-level results are where it gets concrete:

• 4 points more likely to submit a final transcript
• Roughly 5 points less likely to be sitting on an immunisation hold
• 3 points less likely to have a FAFSA verification hold
• 6 points more likely to have finished loan counselling

Cost was $7–15 per student per year. Individual counsellor outreach, which produces comparable enrollment effects, ran $100–200.

The sentence nobody quotes

Page and Gehlbach split their sample by whether a student had already filed a commitment-to-enroll form. For the ones who hadn't, they report impacts that are “essentially zero.”

Read that again, because it undoes most of the marketing around this category. The bot did not persuade anybody. It did not warm up a cold prospect or rescue a student who was drifting toward another school. What it did was clear obstacles for people who had already decided to come and were quietly getting stuck on paperwork.

Summer melt runs 10–20% of college-intending students nationally, and higher among low-income and first-generation students. GSU was seeing up to 18% among students who had already said yes. Those aren't students with second thoughts. Those are students who couldn't get a verification hold cleared in August.

What the bot was actually doing

This part matters more than the headline number, and it's the part that doesn't transfer to a generic widget.

Pounce was wired into GSU's student information system. It knew, per student, which of the required tasks were done and which weren't, and it only nudged on the unfinished ones. The paper flags that integration as an innovation in its own right — prior summer-melt work ran out of high schools, where counsellors could see that a student intended to enroll but had no visibility into whether that student's transcript had landed.

So the bot wasn't a FAQ. It was a checklist with a mouth.

The caseload arithmetic explains the appeal. A human counsellor can carry 40–60 students through a summer. Plain automated texting gets you to about 200. An AI system with SIS access covers a whole cohort, and the marginal student costs approximately nothing.

One more figure worth sitting with: 13.5% of students sent Pounce something it couldn't handle, and those messages got routed to a human by email. That's not a rounding error. Roughly one student in seven needed a person, in a narrow, well-defined domain, with a system purpose-built for it. Any campus plan that doesn't budget for that path is going to discover it the hard way in week one. We've written separately about designing a human handoff that doesn't frustrate people — on a campus the stakes are higher, because the student on the other end is trying to register for classes before a deadline, not asking about a t-shirt.

An AI chatbot for universities earns its keep at the registrar, not admissions

Here's the practical case for 2026, and it has nothing to do with recruitment.

Universities are cutting. Temple laid off about 40 staff in July against an $85 million deficit, after eliminating 190 positions the year before. Howard offered buyouts to 16% of its faculty and staff. UT Tyler offered them to nearly a quarter of its workforce. Harvard's Faculty of Arts and Sciences laid off dozens in August (Christian Science Monitor). The enrollment cliff, federal funding cuts and the collapse in full-fee international students are all landing at once.

The questions don't drop when the staff do. Bursar, registrar, financial aid, housing, parking, library, IT helpdesk, international student services. And unlike a business, the demand curve is savagely seasonal — move-in, add/drop, registration, finals. Everything is quiet, and then a week arrives where every office is underwater simultaneously.

I'd love to give you a benchmark for university helpdesk ticket volume. I can't, honestly: the public numbers are almost entirely vendor blog posts citing each other, and I'm not going to launder one into a statistic. What isn't in dispute is the shape. It's spiky, it's predictable, and it's the worst possible fit for a fixed headcount — which is precisely the shape an always-on system absorbs well and a hiring plan doesn't.

That's where the deflection actually lives. Not “should I apply here,” but “what's the add/drop deadline,” “how do I get a parking permit,” “my VPN won't connect,” “where do I upload my immunisation form.”

FERPA is an access-control problem, not an AI problem

The version of this that goes wrong is easy to picture. Somebody gives a chatbot a service account with broad read access to the SIS so it can answer “what's my balance.” It answers that beautifully. Then it's one badly-worded prompt away from answering what's Sarah's balance, and now you have a systemic disclosure problem rather than a support tool.

Split the surface in two and most of this evaporates:

Anonymous, public questions. Deadlines, policies, procedures, locations, how-do-I. No student record involved, no authentication needed, no FERPA exposure whatsoever. This is the overwhelming majority of the volume, and it's the part that should be grounded strictly in your published content, with a citation in every answer so the student can click through and check.

Authenticated, account-specific questions. Balances, holds, grades, aid packages. These sit behind login, scoped to the person who logged in, with the record fetched at answer time and never cached into a model.

The other rule is unglamorous: don't put student records into training data. Retrieval, not fine-tuning. A model that has read a record to answer one question forgets it; a model trained on records has absorbed them, and you will struggle to argue that what came out the other side isn't an education record. For institutions where this is the deciding question, self-hosting the whole stack puts the data and the jurisdiction on your side of the line — though it buys you less automatic privacy than people assume, and we've been blunt about the tradeoffs there.

Worth checking too: FERPA's school-official exception only covers a vendor when there's a written agreement, a legitimate educational interest and genuine institutional control. Plenty of AI vendor contracts were not drafted with those three conditions in mind.

Your real constraint is your own content

Every campus we've talked to underestimates this one.

A university website is a decade of accreted PDFs, three competing versions of the academic calendar, and a policy page last touched by someone who left in 2019. Point a retrieval system at that and it will answer confidently and wrongly, because it has no way to know which of your four withdrawal deadlines is the current one.

The fix isn't a better model. It's owning a maintained source of truth for each domain, having the bot cite which document an answer came from, and treating every “I don't know” in the logs as a content bug with a named owner rather than an AI failure. We made this argument at length in why your help centre content caps your AI support, and it applies double on a campus, where nobody owns the website as a whole and every office owns a corner of it.

Measure task completion, not conversations

Deflection rate is the metric vendors like because it's flattering and unfalsifiable. Skip it. The Georgia State study is a template for something better, because it measured things that either happened or didn't: did the transcript arrive, did the hold clear, did the student show up in September.

Pick the three or four holds that actually block registration on your campus and track the clear rate before and after, by cohort. If the number moves, you have a result nobody can argue with in a budget meeting. If it doesn't, no amount of chat volume will save the business case. Our piece on what containment and resolution rates really measure covers why the softer numbers diverge so wildly between vendors — the same trap sits waiting in higher ed.

One last thing to keep in the back of your mind. The Digital Education Council's 2026 global survey, across 45,398 responses in 35 countries, found 88% of students already using AI in their studies, and 57% saying their institution gives them inadequate guidance on it. Your students are not going to be impressed that you have a chatbot. They have better ones on their phones.

Which means the only thing your assistant can offer that ChatGPT can't is your data: this student's hold, this campus's deadline, this office's actual policy. Generic answers are a commodity now. Specific ones aren't. Build for the specific ones, and put them in front of the students who have already chosen you — because that, on the evidence, is the only group where any of this has ever been shown to work.

Written by the ChatterMate team — we build an open-source AI support agent that answers from your own documentation and cites its sources. It's free to start (first 300 chats on us) and self-hostable if your data can't leave campus.