Student Support and Success

Student Support and Success — the institutional-side work of helping students stay, progress, and finish: advising, administrative assistance, outreach, referral, and the allocation of scarce support. This is the page for what institutions do to students rather than what happens inside a course. Higher Education is the broader umbrella for the sector and Student Experience covers how AI lands on the student's own experience; Learning Gains measures whether learning happened. The distinctive finding in this knowledge base is that AI support reliably moves task completion — a dated, binary action a student controls and the institution can observe — while leaving persistence, credits, and graduation largely untouched. Separating those outcomes is the point of this page.

Questions to Consider

• Which outcome are you trying to move? A registration reminder and a retention program are different interventions with different evidence, and the research here suggests one does not buy the other.
• If a model flags a student as at risk, what happens next? Who acts, with what capacity, and what would make the recommendation feasible rather than merely accurate?
• Support capacity is finite. When AI ranks or allocates it, what does that do to the students a human advisor would have noticed anyway?
• Who is accountable when an automated message, referral, or risk score is wrong? The student sees the consequence; the institution owns the system.
• Does your support data travel? Cross-office sharing is what makes targeting possible and is also the most common reason targeting quietly stops.

Introduction

Student support sits on the institutional side of the relationship. Its work is advising, outreach, referral, and the allocation of finite human and financial capacity; its evidence base is administrative records, registration events, credits earned, and whether a student returns. Generative AI entered this territory later than it entered instruction, and much of the literature here concerns non-generative systems — text-message chatbots, early warning models, federated risk prediction — with generative tools arriving as advisors, assistants, and knowledge-base responders.

The distinction that organizes this page is between what a nudge can do and what a trajectory requires. Support technologies excel at a dated, binary decision: register by this date, complete this form, start Early Start. They have far more difficulty with cumulative outcomes — persistence, credits, graduation — which instruction, finances, employment, family circumstances, and prior preparation all shape. The knowledge base's strongest evidence on this comes from a four-year randomized evaluation that moved registration sharply and graduation not at all.

The support functions

The research clusters into five functions, and the cluster boundaries matter because they carry different evidence.

Outreach and communication. Institutions message students at scale, and the strongest evaluations test whether it changes anything. A four-year randomized study of CSUNny, a non-generative text-message chatbot at California State University, Northridge, followed two undergraduate cohorts (N = 8,708) through eight semesters (Mata, Russell & Page, 2026). A registration reminder sent July 31, 2018 made treated students 34 percentage points more likely to enroll by August 16 and only 2 points more likely by September 15; Early Start reminders produced 11 points more registration by June 7 and 20 points by June 22. Receptivity held: annual opt-out rates never exceeded 4%. A pre-registered multi-semester trial at Georgia State University nudged students in two large asynchronous courses (N = 1,568 and N = 915) with two to three customized messages a week, raising the odds of earning an A or B by four percentage points against 61% of controls and shifting DFW rates by about three points in each course (Meyer et al., 2026).

Advising and academic planning. Course and grade prediction supplies the planning input — one model jointly predicts which courses a student will take and the grades they will receive (Savala, 2026), and another predicts module-level progress in large online programming courses with an intrinsically interpretable decision tree (Zhang, Jeffries & Koprinska, 2026). Transfer credit is an advising problem in its own right: CourseGraph models course content as knowledge graphs to evaluate external course equivalences for mobile students (Nijdam et al., 2026). At the institutional level, a review of 155 studies of AI and service delivery in higher education found learning analytics the most common application at 46 studies (29.7%), ahead of chatbots and virtual assistants at 31 (20.0%) and predictive analytics at 29 (18.7%) (Nyamboga, 2026).

Referral and support allocation. Prediction is not a plan, and the gap between them is where the field's sharpest critique sits. SC2R formalizes it as the actionability gap: a risk score becomes decision support only when its recommendations are semantically feasible and machine-checkable — constrained by timing, budget, immutability, and availability rather than merely model-valid (Le, Abel & Laforge, 2026). Whether models can allocate support well is contested empirically: asked to recommend support plans for 4,500 synthetic student vignettes, three LLMs showed limited sensitivity to student need and sharp inconsistency across models (López-Pernas et al., 2026).

Academic support access. Course-specific retrieval systems aim at the students least likely to ask a human. Beacon, built from a single programming module's approved materials, drew high relevance ratings (89%) and was judged by most students to support rather than replace their learning (66.7%), in a small evaluation of 15 students and 4 academics (Zhou et al., 2026). Tutoring take-up is its own problem: a two-year randomized trial of a virtual tutoring layer for struggling students had to test take-up separately from learning, because reaching the students who need support is not the same as providing it (Fryer et al., 2026).

Administrative assistance. Leadership and administration are studied as their own application domain, with a ten-domain taxonomy mapping where AI lands in educational leadership (Sposato, 2025). Institution-run services reach past advising into health: an integrated campus well-being framework pairs prevention — improving how feedback is collected — with intervention through mental-health detection (Tang, 2026). Credentialing sits adjacent: when an agent can complete a course on a student's behalf, a credential that says "earned" loses its meaning, which turns completion records into a design problem (Srivastava, 2026).

From prediction to support

Risk prediction — early warning systems, dropout models, at-risk classifiers — is Learning Analytics territory in this knowledge base, and this page treats predictive analytics as the same thing. The modeling work is substantial: supervised classifiers identify students before withdrawal from academic performance, demographic, and enrollment records (Gheisari & Salarian, 2026); a dual-layer framework combines Codeforces behavioral logs (n = 1,816) with psychographic survey data from ten universities to predict attrition in competitive programming (Alam et al., 2026); and a federated architecture predicts performance and dropout across institutions without sharing raw student data, reaching AUC = 0.918 on OULAD against 0.925 centralized (Villegas-Ch et al., 2026). A precision-education vision extends the logic to student digital twins and "preventive student success" (Han et al., 2026).

What belongs here is the step after the score: support allocation. Two findings set its limits. First, the ranking-versus-calibration split — federated risk models kept their AUC under distributional shift while their calibration degraded markedly, so a model that still ranks students correctly can be wrong about how likely each one is to need help. Second, the enabler study: an international Delphi and AHP/SNAP analysis of learning-analytics-to-intervention identified seven enablers and ranked institutional strategic orientation highest (priority 0.2072) and most influential on the others (PageRank 0.2430), locating the bottleneck in institutional planning rather than in the models (Svetec, Divjak & Kadoić, 2026). Behavioral clustering of 14,003 student records into six profiles, mapped to recommended learning objects, is the recommendation layer this points toward (Najem et al., 2026).

The outcome ladder: what each measure actually means

The word "success" hides at least five different measurements, and AI support does not move them equally. Distinguishing them is the most useful thing this page can do.

• Task completion is a single, dated, binary action the student controls and the institution observes within days — registering by a deadline, filing a form, enrolling in Early Start. It is where nudges work, and the effects can be large and immediate (34 percentage points, then 2, in the CSUN registration reminder).
• Credits (units enrolled and earned) are cumulative and depend on course availability, sequencing, and how many terms a student can afford. In the CSUN evaluation no significant treatment effect appeared on units enrolled, earned, or cumulatively earned.
• Persistence is continuation across terms — enrollment by semester sequence — and it did not move either, at N = 8,708 with power to detect effects of 0.05 standard deviations or larger.
• Retention is the institutional rate that persistence produces, and it is usually reported at the program or cohort level. Course-level DFW rates are the closest thing to a leading indicator in this literature, and they moved modestly (−3 percentage points in each course of the Georgia State trial).
• Graduation is the terminal, multi-year outcome. Control mean graduation by the fourth year was 0.190 in the CSUN study, and the treatment effect on it was not statistically significant.

The authors' explanation for the pattern is the page's central claim: a reminder acts on a decision that is discrete and near-term, while persistence and GPA are cumulative and shaped by instruction, finances, employment, family circumstances, and prior preparation. They conclude that communication through a tool like this, on its own, may be insufficient to change them — and that the nulls are precise rather than underpowered. Early registration still carries institutional value for staffing and space planning even where learning outcomes do not move.

Equity and the risks of acting on a score

Support systems act on students, which makes their failure modes different from a tutor's. A stress test of six post-hoc fairness interventions on a replicated vendor-controlled early warning system built from 168,550 student records found the interventions failing to deliver the fairness they promised (McConvey et al., 2026). A student-advocacy review organizes the risk surface around admissions, recruitment and financial aid, and student-success services — the areas where institutional AI bears most directly on students (Student Defense, 2026). Privacy and equity are structural here rather than incidental: federated learning exists because sharing student records across institutions is unacceptable, and the human-in-the-loop component in the CSUN program — administrators answering what the chatbot could not, then folding the answer back into its knowledge base — is what the authors argue reaches students who ignore email and phone.

The institutional conditions

Durability in this literature is an organizational property, not a technical one. The CSUN program survived four years because the channel was centrally owned, jointly overseen by the Office of Undergraduate Studies and the Office of the Registrar, and written by a single communication specialist for a consistent voice. Its targeting decayed for an equally organizational reason: because student data was not centralized, a financial-aid reminder required one office to identify non-filers and another to pass the subset along, and that friction was burdensome enough that targeted campaigns fell from 36% of all campaigns in AY2018-19 to 6% by AY2022-23. Where the work sits, who owns the data, and whether cross-unit coordination is sustainable decide what a support system can actually do — which is why this page carries Change Management, AI Governance, and Educational AI Policy as facets rather than treating deployment as an IT decision.

Connections to related concepts

Student support connects to Student Experience as the student-facing counterpart — the same technologies seen from the student's side rather than the institution's — and to Help-Seeking for the mechanism by which students who need support actually obtain it; outreach and course-specific assistants are both attempts to lower the cost of asking. Learning Analytics owns the prediction that feeds allocation, Learner Modeling and Adaptive Instruction and Knowledge Tracing the models underneath it, and Recommender Systems and Learning Paths the recommendation layer. It connects to Well-Being through campus mental-health and prevention systems, to Career Development and Readiness as the outcome that follows completion, to Equity and Privacy through the risks of acting on scores, and to Administrators, Stakeholders, and Change Management as the roles and processes that decide whether any of it is sustained. Learning Gains is the adjacent measurement node: this page's outcomes are administrative rather than instructional, and the two do not move together.

Connected Concepts

• Higher Education — the sector umbrella this page sits under
• Student Experience — how AI lands on the student's own experience
• Learning Gains — whether learning happened, the instructional outcome this page does not measure
• Learning Analytics — prediction, early warning, and predictive analytics
• Learner Modeling and Adaptive Instruction — the modeling layer beneath risk prediction
• Knowledge Tracing — fine-grained skill and mastery estimation
• Recommender Systems and Learning Paths — recommending courses, resources, and paths
• Help-Seeking — how students come to ask for support
• Well-Being — campus mental health, prevention, and intervention
• Career Development and Readiness — the outcome after completion
• Administrators — the role that owns and runs support systems
• Stakeholders — who has a claim on institutional AI decisions
• Change Management — sustaining the program after the pilot
• AI Governance — policy and oversight for institutional AI
• Educational AI Policy — the policy context for institutional deployment
• Equity — who the system reaches and who it misses
• Privacy — student records, data sharing, and federated approaches
• Human-in-the-Loop — human judgment inside an automated support flow
• RCT — the design behind the strongest evidence here

Connected Articles

• Sustaining AI-enabled student support: A four-year implementation and impact study — four-year randomized evaluation of a university support chatbot: task completion moved, graduation and GPA did not
• Let's Chat: Leveraging Chatbot Outreach for Improved Course Performance — pre-registered multi-semester outreach trial: higher A/B rates, modest DFW shifts, one demographic exception
• Can AI deliver appropriate support for diverse student profiles? A large-scale evaluation — 4,500 synthetic vignettes: limited sensitivity to student need and cross-model inconsistency in support recommendations
• From Student Risk Prediction to SC2R: Semantics-Constrained Counterfactual Recourse for Educational Decision Support — the actionability gap: recourse must be feasible and checkable, not merely model-valid
• Fairness Theatre: Evaluating Post-Hoc Fairness Interventions in Vendor-Controlled Early Warning Systems — six post-hoc fairness interventions on a vendor early warning system built from 168,550 records
• Analysis and Prediction of At-Risk Students Using Machine Learning Algorithms — supervised classifiers identifying students before withdrawal
• Predicting Student Attrition in Competitive Programming: A Large-Scale Study Integrating Survey Insights and Global Behavioral Logs — behavioral logs plus psychographic survey predicting attrition
• Federated and Explainable Learning Analytics for Privacy-Preserving Academic Risk Modeling Across Heterogeneous Educational Institutions — federated risk modeling across institutions without sharing raw student data
• From Precision Medicine to Precision Education: A Vision for AI-Powered Student Digital Twins, Preventive Student Success, and Career-Aligned Academic Pathways — digital twins and "preventive student success" as a vision
• From Learning Analytics to Educational Interventions: Enhancing Decision-Making and Learning Design — seven enablers for closing the loop from analytics to intervention
• Reducing Barriers to Academic Support: Evaluating a Course-Specific RAG System for Addressing Help-Seeking Disparities in Higher Education — a course-specific assistant aimed at lowering the cost of asking for help
• Virtual Tutoring with Computer-Assisted Learning: An Experiment in Take-Up and Learning — tutoring take-up tested separately from learning
• Behavioral clustering for adaptive learning object recommendation in higher education using learning analytics for personalized student success — six behavioral profiles mapped to recommended learning objects
• Jointly Predicting Courses and Grades Using a Transformer-Based Model (TRACE) — joint prediction of courses and grades for planning
• A Machine Learning Approach for Predicting Student Progress in Online Programming Education — interpretable module-level progress prediction
• From digital transformation to intelligent classrooms: artificial intelligence, adaptive leadership, and service delivery in global higher education—a systematic literature review — 155 studies of AI, leadership, and service delivery in higher education
• Artificial intelligence in educational leadership: a comprehensive taxonomy and future directions — ten-domain taxonomy of AI in educational leadership
• Students at Stake: Risks of AI Deployment in Higher Education — student-side risk surface: admissions and aid, student-success services, and instruction
• New AI-Driven Tools for Enhancing Campus Well-being: A Prevention and Intervention Approach — campus well-being support spanning prevention and intervention
• CourseGraph: Finding overlaps and differences in Computer Science courses across universities — course equivalence for transfer credit and mobility
• Credentials That Carry Their Evidence: Credential design, dynamic tiering, and approved use when AI agents can do the work — what a completion record means when an agent can do the work