Every fall, provosts and boards receive the same grim arithmetic. Roughly one in three undergraduates who start a bachelor’s degree will not finish it within six years, and each one who leaves takes tuition revenue and a graduation-rate percentage point with them. For institutions already squeezed by enrollment declines and shrinking state appropriations, retention is no longer a student affairs talking point. It is a budget line.
That pressure is why campuses from flagship research universities to regional comprehensives have spent the past several years building, buying, or bolting on AI-powered early alert systems. These platforms pull data from learning management systems, student information systems, attendance logs, and financial aid records, then score each student for risk of stopping out. When a name crosses a threshold, an advisor gets a flag, often days or weeks before a professor would have noticed a student sliding.
Keeping Humans in the Loop
Georgia State University’s decade-long experiment with predictive analytics, built initially with EAB and later folded into its own National Institute for Student Success, remains the reference case for the fi eld. But its architect is quick to correct growing misunderstanding.
“AI never responds to alerts,” says Dr. Timothy Renick, the institute’s executive director. Every one of the more than one hundred thousand alerts Georgia State’s system triggers each year is routed first to the individual advisor assigned to that student, who then decides whether the moment calls for a phone call, a sit-down meeting, a text, or nothing more than a routine email. The distinction is not academic. Predictive analytics can tell an advisor where to look, but a human still decides what happens next.
That human-in-the-loop design extends to what the algorithm is allowed to see in the first place. Georgia State’s risk scores draw only on data the university has always collected for academic and financial purposes: majors, credit hours, grades, satisfactory academic progress, and financial aid eligibility.
“Our early alerts do not use any data on the students’ race, ethnicity, mental health issues, economic background, zip codes, and campus behaviors,” says Renick. Additionally, the scores are never used to exclude a student from a competitive major such as nursing or business, he says, and are not punitive by design. Instead, Georgia State uses them to run what Renick calls “campaigns,” — concentrated pushes during the lulls between the start of a term, midterms, and when grades roll — when advisors work systematically through the students most likely to run out of financial aid eligibility before
finishing a degree.
Exploring Autonomous AI Agents
Not every institution draws its boundaries that narrowly. Newer entrants like ibl.ai’s MentorAI and SEAtS ONE promise something more ambitious: agents that do not just flag a struggling student to a human advisor but message the student directly, nudging them toward tutoring, a financial aid appointment, or a mental health resource, around the clock and without waiting for a staff member to log in.
That gap between what an institution says it does and what its own vendor now sells is not unique to Georgia State. EAB, the company behind Navigate360, the platform Georgia State used to build its original predictive analytics program, now markets a newer “Navigate360 AI Assistant” that the company says detects student challenges and “launches the right next step autonomously,” without waiting for an advisor to log in first. Becca Thompson, an associate vice president for enrollment management and student success at Louisiana State University, says the tool helps her team “make our emails more direct.” Dr. Twyler Earl, an associate vice president of student success at Southern Nazarene University, calls the AI-generated templates “a game changer for our team,” given the volume of messages her office sends about absences and emergencies. Lisa Matye Edwards, vice president for student affairs at Arapahoe Community College, notes that the assistant is “like having an EAB consultant in your back pocket.”
For advising offices that have not added headcount in proportion to enrollment growth, the appeal of any of these tools is obvious. A single advisor at a large public university might carry a caseload of four hundred students or more. No amount of dedication closes that gap through manual outreach alone. AI promises to do what an overstretched human staff cannot: notice the student who quietly stopped logging into the learning platform three weeks before finals, or the one whose campus card swipes at the dining hall dropped off entirely.
But noticing is not the same as understanding, and that is where the technology runs into a set of legal and ethical tripwires the field has not fully worked out.
Navigating Legal and Ethical Complexities
The clearest is mandatory reporting. Federal law under Title IX and the Clery Act requires most college employees to report disclosures of sexual assault, harassment, and certain other safety concerns to designated campus officials. When a chatbot is the first point of contact for a student in distress, it is genuinely unclear whether that chatbot, or the company that built it, is obligated to report what it hears.
At Georgia State, that ambiguity already has at least a partial, deliberate answer built into the system. The university’s AI chatbot fields routine questions about deadlines, processes, and campus resources, but Renick said humans write and vet every answer in its knowledge base “so we know what the bot is saying at all times,” and that AI does not write any of the responses the bot sends.
“I would not support any system that would use AI to write its own responses to such topics, even questions about financial aid deadlines or where to find Chemistry tutoring,” says Renick. “I would be even more strongly opposed to AI generating automatic answers to the range of advising issues posed by students about majors, careers, dropping courses, planning for future semester, and so forth.”
The rule holds even, or especially, when a student brings the bot something it was never built to handle. “If a student mentions mental health issues to our AI chatbot, that text is not responded to by the bot but the message is automatically forwarded to a live person in the Dean of Students Office for a personal intervention,” Renick says. The move amounts to a workaround for a legal question regulator have not yet answered. Rather than decide whether a chatbot counts as a mandated reporter, Georgia State simply never lets the chatbot be the one holding that conversation.
The confidentiality question cuts the other way, too, especially at institutions less quick to route a disclosure to a person. A student who would never voice a disclosure to a professor or a resident advisor might type it into a chat window at two in the morning precisely because it does not feel like talking to a person. If that disclosure triggers an automatic report, colleges risk teaching students that the “safe” late-night outlet was never confidential to begin with, which could push the most vulnerable students toward silence rather than help.
Bias and Boundaries
Bias compounds the problem. Researchers have examined predictive models used across community college systems and have documented calibration bias in risk-scoring tools, meaning that at the same predicted-risk score, Black students in some models have shown higher actual success rates than white students. Small modeling choices — choosing one machine learning method over another — can meaningfully change which students get flagged for extra support and which get missed. New America researchers Manuela Ekowo and Iris Palmer reached a similar conclusion years ago in what remains one of the fi eld’s foundational ethical frameworks: predictive analytics can entrench the very inequities it is marketed to solve if institutions treat the risk score as a verdict rather than a starting point for a human conversation. Georgia State’s decision to keep race, ethnicity, and economic background out of its models entirely sidesteps some of that risk, though it also means the system is not designed to catch the inequities in access to advising, tutoring, or awareness of support resources that a broader model might reveal.
That is the tension sitting underneath every early alert dashboard. The more precisely a system can detect distress, the more it invites questions about what obligations follow from detecting it, and for whom.
Those questions become sharper still when the source of a student’s distress sits outside anything an academic advising model was built to anticipate.
Navigating Geopolitical Realities
Federal immigration enforcement activity has intensified in states including Texas and Florida, and campus mental health staff in both states have described fielding more disclosures tied to family immigration status, even as both states have simultaneously restricted how institutions can collect, discuss, or act on data related to immigration status, DEI-adjacent categories, and related political flashpoints. An AI system trained to flag “behavioral risk indicators” is, by design, agnostic about the source of the behavior. A sudden drop in class attendance looks the same in the data whether the cause is a bout of flu, a mental health crisis, or fear of a parent’s arrest. What differs is what an institution is legally permitted, or willing, to do once a human is looped in, and that answer may vary sharply depending on which state the flag goes off in. A system built the way Georgia State’s is, restricted deliberately to academic and financial data, would likely never register a case like that at all. A system built to have open-ended wellbeing conversations with students, the direction much of the industry is now heading, cannot make that same choice to look away.
No institution contacted for this story had a fully built-out answer to that scenario, and it may be too early for one to exist. But as AI-powered advising tools grow more sophisticated at reading the earliest signs of student distress, the industry’s harder work is just beginning; deciding not just what the algorithm should notice, but what a college is prepared to do, and legally allowed to do, once it does.
This story originally ran in the July issue of The EDU Ledger’s 2026 Summer Supplement.
