ChatGPT
Claude
Perplexity
Gemini
GrokCan AI reduce summer melt? Yes, for specific tasks and specific students
Published trials show that personalized text messages, and in one case a conversational AI assistant, can help admitted students finish the steps between deposit and move-in. The effects are real but uneven: they appear where students lack other help, and several comparisons in the same studies show nothing.
"Summer melt" means students who intend to enroll, often with a deposit paid, but never show up in the fall. Castleman and Page estimated melt rates "ranging from 8 to 40 percent" in their 2014 analysis of college-intending graduates. That is a wide range because it depends on the population: the study centers on low-income graduates and does not describe your admitted class.
A note on method: no paid keyword-volume tool was available for this article, so it was scoped from public search results. Every figure below comes from a source we opened. Nothing comes from Skolbot or its customers.
Why admitted students melt
Most melt is administrative before it is motivational. Between May 1 and the first day of class, an admitted student must complete a list of tasks that no one else is tracking for them.
| Summer task | What goes wrong |
|---|---|
| Financial aid | FAFSA verification requests, award letters nobody understands, a missed loan acceptance |
| Housing | Application deadlines, roommate forms, deposit confusion |
| Health and immunization | Records requested by the health center that families do not have |
| Transcripts and test scores | Final transcript not sent, a placement test never taken |
| Orientation and registration | A hold on the account blocks course registration |
| Billing | A first tuition bill arrives with a balance the family had not planned for |
Each item is small. Together they produce a student who, in August, finds a hold on their account and quietly decides to start at a community college or wait a year. The college sees the loss only when the roster is final.
What the trials found
Three bodies of work matter here. They differ in design, so we describe each separately.
Castleman and Page: texting and peer mentors
Castleman and Page's "Summer Nudging" (Journal of Economic Behavior & Organization, 2015) tested two low-cost interventions: automated, personalized texts about pre-enrollment tasks, and outreach by near-age peer mentors. The authors report that both raised enrollment among students with less access to quality counseling or information during the school year.
The What Works Clearinghouse review gives the detail, and it is more sober than the headline. The study included 6,196 students across sites in Massachusetts, Pennsylvania and Texas. Ten texts were sent about every five days in July and August. Most comparisons were not statistically significant. The clearest gain was in Lawrence and Springfield, Massachusetts (746 students), where enrollment in any college rose from 62.8% to 69.9%. In Boston, texting showed no significant effect.
The WWC page is itself inconsistent on one Dallas comparison, which one section marks significant and another does not. We do not rely on it.
Georgia State: a conversational AI assistant
Page and Gehlbach (AERA Open, 2017) used conversational AI to send personalized text outreach to would-be freshmen at Georgia State University during the summer. Students who had committed to the university and were assigned to the treatment group were, in the authors' words, "3.3 percentage points more likely to enroll on time." The authors say this is comparable to earlier interventions with far less staff effort.
This is the closest published evidence for an "AI agent" in the melt window. It is also a single large public university, from before generative AI. Treat it as proof of the pattern, not a forecast for your campus. Our companion piece on what US studies show about admissions chatbots covers the wider literature and its limits.
What the two studies share
Both interventions were short, timed to a deadline, and tied to a task the student could complete. Neither tried to change a student's mind about the college. That distinction drives the design advice below.
Where an AI agent fits, and where it does not
An AI agent can answer a student's question at 11 p.m. on a Sunday, such as "Why is there a hold on my account?" or "Where do I send my immunization form?" It can also send reminders tied to the student's own checklist. Those two jobs map onto the Georgia State design: personalized outreach plus on-demand answers.
An agent does not decide aid packages, waive a hold, or talk a family through a financial crisis. Those go to a person. The research supports AI as a complement to staff, not a substitute. Our definition of an AI admissions agent lays out the boundary between what an agent answers and what it hands off.
| Job | Agent | Human advisor |
|---|---|---|
| Reminder before a housing deadline | Yes | Not needed |
| "Where is my immunization form?" | Yes, if the answer is in approved content | Backup |
| Explaining a changed aid award | Gives the general process, then hands off | Yes |
| A family considering withdrawing over cost | Flags and routes quickly | Yes |
| Confirming a deposit was received | Yes, if connected to the system of record | Backup |
The last row is a capability question. An agent that cannot see the student's record can only give general answers. Ask any vendor, including us, exactly which systems the agent reads and writes, and test it before signing.
A summer playbook a US college can test
This is an illustrative workflow, not a measured result. It follows the pattern the trials share.
- Build the checklist per student. Start from the tasks that your own data show hold up enrollment: aid, housing, health forms, final transcript, registration holds.
- Send fewer, dated messages. The Castleman and Page design used ten texts over two months. More is not better.
- Make every message answerable. A student who replies should get a useful answer, not a no-reply notice.
- Route emotion to people. Messages about cost, family conflict or doubt should reach an advisor the same day.
- Prioritize the students with the least backing. The trial gains concentrated among students with fewer other sources of help, and the Boston result suggests the lift is small where support already exists.
- Clean the contact data first. A message sent to a wrong or missing number reaches no one, and the dashboard will not tell you.
Measure it properly
The most useful thing the research gives you is a method. Randomly hold back a sample of admitted students from the new outreach, and compare the two groups on one outcome: a completed checklist, or enrollment on census day.
Before-and-after comparisons are unreliable here. Melt varies by class, by aid year and by economy, so a better August may say nothing about your agent. A control group is cheap to create and is what turned the Georgia State result into evidence.
Privacy and consent
FERPA governs education records, and the Department of Education's Student Privacy Policy Office explains how it applies. Applicant data is often handled under the institution's own privacy notice and state law, so confirm with counsel which rules cover admitted students before you connect an agent to your records.
Texting adds consent requirements under federal and state rules. We do not summarize them here because they change; have counsel review your opt-in language and opt-out handling. Tell students plainly that they are talking to an automated assistant, and keep admission decisions with people.
Where this fits in your enrollment plan
Summer melt is the last stretch of a longer funnel. Our guide to yield management and post-offer drop-out covers the earlier stages, and the guide to AI agents for student recruitment shows where the summer agent sits in the full journey. Admitted students start by applying through the Common App or a direct application, and the agent's job is to keep that early commitment from fading.
