AI-Based Academic Advising Across the Student Lifecycle: A Systematic Literature Review

Abstract

Academic advising is fundamental to student success, yet the rapid integration of Artificial Intelligence (AI) and Machine Learning (ML) is fundamentally transforming the delivery of academic support. While predictive models and Recommendation Systems (RS) are becoming more accessible, the existing literature remains fragmented across diverse technical architectures and institutional objectives, preventing a clear understanding of the field’s evolution. In view of this, we present a Systematic Literature Review (SLR) of AI-driven academic advising, adhering to the PRISMA 2020 framework. We analyzed 27 peer-reviewed studies published between 2018 and 2025 to synthesize methodological trends and functional applications. Our findings reveal that while most systems prioritize pathway recommendations via classical ML or hybrid architectures, Early-Warning Systems (EWS) remain anchored in predictive classification. Furthermore, a nascent shift toward Generative AI (GenAI) indicates a move toward more interactive advising, though transparency and evaluation standards remain inconsistent. This review identifies a critical tension between algorithmic performance and institutional interpretability. We conclude by proposing a research agenda that emphasizes the need for cross-context validation and the development of socio-technical frameworks that integrate AI into existing higher education management structures.

1. Introduction

Student academic pathways in higher education have become increasingly complex. Students are required to make early and repeated decisions regarding program selection, course enrollment, and academic progression, often within flexible and highly structured curricula. At the same time, student populations have grown larger and more diverse, making individualized academic support increasingly difficult to provide at scale. These conditions expose clear limitations in traditional academic advising models that rely primarily on human expertise and institutional experience. Academic advising plays a critical role in helping students navigate these complexities, as it directly influences academic decision-making, persistence, and timely degree completion [1,2,3,4]. However, as advising demands increase, purely human-centered approaches face challenges related to scalability, consistency, and the ability to intervene at the right moment, particularly for students at risk [5].

Therefore, higher education institutions have experienced substantial changes in how student data are collected and used. The increasing availability of academic records, enrolment trajectories, retention indicators, and interaction data from digital learning environments has reshaped how advising decisions can be supported. Despite this, research in Educational Data Mining (EDM) and Learning Analytics (LA) has shown that such data can be used to identify patterns related to academic performance, progression, and disengagement [6,7,8]. In contrast, these developments gave rise to early data-driven advising applications, ranging from descriptive reporting to predictive models for early warning, course allocation, and targeted intervention, though many remained limited in scope and transparency.

More recent studies have expanded advising systems through the integration of AI techniques. In this context, ML models are widely used to classify students, predict outcomes, and assess academic risk, particularly during the early stages of university enrolment [8,9]. Equally important, Recommender Systems (RS) have been introduced to support advising decisions that involve multiple viable options, such as course selection, pathway planning, and academic orientation, by leveraging similarities between student profiles, preferences, and historical cases [10,11,12]. In addition, advances in Deep Learning (DL), and more recently GenAI, have further extended system capabilities, enabling the modeling of complex student trajectories, sequence-aware recommendations, and interactive advisory interfaces. Despite this expansion, AI-based advising systems do not follow a single design rationale. Similar computational approaches are often applied to different advising functions, while identical advising tasks may be supported by very different AI paradigms.

Importantly, academic advising itself is not a single or isolated activity; it unfolds across multiple stages of the student academic lifecycle, from initial orientation and program selection to continuous academic progression and, finally, transition toward graduation and professional integration [1]. In contrast, many AI-supported advising systems are designed to address a specific decision point rather than to support advising as a longitudinal process. In short, the relationship between system design, advising function, and lifecycle stage is not always clearly articulated in the literature. This makes it difficult to assess the practical suitability of different AI approaches or to compare systems developed for distinct advising contexts.

Although the broader field of AI in education has attracted substantial attention, existing systematic reviews of AI in education and learning analytics often emphasize algorithmic novelty or system-level performance, with limited focus on advising-specific practices and lifecycle considerations [13,14]. Recent reviews have also examined related domains such as AI-driven tutoring and dropout prevention in higher education, as well as recommender systems supporting academic decision-making [12,15]. However, these reviews remain focused on specific application domains rather than on academic advising as a lifecycle-wide institutional practice. In contrast, the present review compares AI paradigms across advising functions, student lifecycle stages, and implementation constraints, thereby offering a more advising-centered and institution-aware synthesis.

In response to these limitations, this study conducts a systematic literature review of 27 peer-reviewed studies published between 2018 and 2025 that examine the use of AI in academic advising. Considering the amount of research in this area, this study adopts an analytical framework that jointly considers computational paradigms, advising functions, and stages of the student academic lifecycle, rather than focusing on individual algorithms in isolation. By integrating these perspectives, the study offers a more structured synthesis of the literature and draws attention to areas where AI-based advising approaches remain fragmented, insufficiently evaluated, or largely conceptual.

2. Background

2.1. The Principles and Goals of Academic Advising

Academic advising is an important element in post-secondary education and sits relatively at the center of the trifecta of academic advising, student development, and institutional function. Advising is not supposed to be only administrative in nature; rather, it is described in the literature as teaching and learning processes that assist students in refining their objectives and developing an understanding of the requirements in order to effectively navigate their educational pathways [4,16]. In this respect, advising is positioned to assist students in the effective management of intricate educational programs.

Academic advising objectives are not limited to just scheduling classes and ensuring that students complete their degree requirements. Empirical studies have established a connection between effective advising and heightened student engagement, retention, persistence in their program, as well as overall academic achievement [1,2]. These studies have caused a paradigm shift in how advising is viewed, streamlining it towards a purposeful interaction that has the potential to affect retention, progression, and timely completion. The advising philosophy has had to evolve to more formalized and structured systems to cope with the increasing complexity and the sheer number of students in the system. Academic advising occupies a central position within higher education institutions, linking student development processes, teaching–learning mechanisms, and institutional performance objectives. It operates at the intersection of academic progression monitoring, decision support, and strategic enrollment management.

2.2. Evolution of Academic Advising Practices

Reflecting changes in higher education systems and student demographics, the practices of academic advising have progressed in multiple identifiable phases. In the beginning, advising was informal and reactive, and took the form of individual conversations with students after the academic problems had already arisen. However, with the growth of institutions and diversification of student populations, advising practices began to incorporate more of the planned developmental models with a focus on early and ongoing engagement [17].

The practice of academic advising has gradually shifted away from being compliance-based to being more developmental and proactive, focusing on student development, student autonomy, and long-term planning. Initially, academic advising was more of an administrative process, focusing on institutional policies, course registration, and record management. Gradually, it shifted to being more collaborative and student-centered, focusing on dialog, reflection, and informed academic decision-making. This evolution of academic advising can be seen to follow the path suggested by the National Academic Advising Association (2020) [4]. Four phases of the evolution of academic advising can be identified. In the 1970s and early 1980s, academic advising was primarily compliance-based. In the 1990s, there was the formalization of degree planning through prescriptive models. In the 2000s, there was the emergence of more developmental models, which focused on student engagement and student responsibility in academic decision-making. In recent times, it has become more decision-oriented, using data analysis to inform the advising process. While these models have differed in focus, the common and growing challenge of these systems has been scaling of personalized support without excess burdening of human advisors. This has become a more evident challenge, as institutions have tried to support and advise large numbers of students in a consistent and structured manner.

2.3. The Development of Data-Driven Analytical Advising

The increased accessibility of institutional data changed the course of academic advising. The introduction of academic and learning analytics allowed institutions to examine student data, Learning Management System (LMS) activity, and other performance data to inform decision-making [6,18]. Early on, these initiatives focused on automated decision-making and were directed at descriptive analytics reporting, where dashboards and summaries were provided to inform rather than automate the advising and administration decision-making process.

Predictive analysis became the core of data-driven advising as the learning analytics field matured. Research demonstrated the ability of predictive models to identify students at risk of underperforming and dropping out as early as the beginning of their academic careers [7]. This development enabled the implementation of EWS to aid in targeted, preventive intervention, particularly in large educational systems. However, these systems were often limited to a narrow range of performance criteria and provided little insight into the rationale for their predictions. At the same time, this data-driven foundation helped enable the broader adoption of AI methods in higher education, including predictive, adaptive, and more recently generative approaches to student support [15,19].

3. Materials and Methods

To clarify the scope of the literature on AI in academic advising within higher education, this review was conducted following the Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) (Supplementary Materials), ensuring transparency in every methodological step [20]. In line with these transparency principles, the study protocol was prospectively registered on the Open Science Framework (OSF). The registered protocol specifies the research objectives, eligibility criteria, search strategy, and methodological procedures adopted in this review. Therefore, to identify the relevant studies on this matter, the review follows an overall process that consists of five stages as shown in Figure 1 below.

3.1. Planning and Review Design

Before conducting this review, it is necessary to determine the key elements that organize the process and guide the identification of the best evidence from the existing literature on this predefined scope. Therefore, these elements were defined in advance to support consistent decisions during the search, screening, and analysis stages.

3.2. Research Questions

To better understand how AI is currently being used in academic advising, a set of research questions was defined to guide the review. These questions were formulated to reflect the main aspects that repeatedly appear in the literature, while keeping the scope focused on advising-related applications rather than broader uses of AI in education. In particular, attention was given to the types of AI approaches adopted, the advising functions these systems are intended to support, and how their performance or usefulness is assessed. Moreover, defining the research questions at this stage helped narrow the field of investigation and supported consistent decisions during the search and screening process. They also served as a reference point during analysis, allowing studies with different designs and objectives to be examined within a common frame. The full set of research questions used in this review is presented in

Table 1. Research questions.

Research Question	Aim
RQ1: What types of AI approaches are used in academic advising systems within higher education?	To examine the range of AI paradigms adopted in the literature, and to understand how advising problems are computationally framed.
RQ2: Which academic advising functions are supported by AI-based systems across the student lifecycle?	To investigate how AI is applied to different advising activities and identify which stages of the student lifecycle receive the greatest attention in research.
RQ3: How do different AI paradigms perform across advising functions and stages of the student lifecycle?	To compare AI approaches in terms of their practical suitability, reported strengths, and observed limitations when applied to different advising tasks and lifecycle stages, highlighting where certain paradigms demonstrate greater or lower practical effectiveness.

.

3.3. Search Strategy and Keyword Identification

After defining the relevant research questions in the section above, it becomes clear to identify the adequate keywords that frame the scope of our review by combining the most used keywords in the field of Academic Advising and AI. This is represented as a query; the keywords’ identification was carefully assessed to gather most of the relevant AI academic advising studies. However, the queries were inserted into reputable research databases and search engines, such as Scopus, Springer, Web of Science, Google Scholar, and Springer Link. As depicted in Figure 2, we used two combinations of keywords intersected semantically, where one is related to the method (AI) and the other to the application domain (academic advising). Moreover, the search queries are presented in

Table 2. Search Queries for each search database.

Database	Search Query
Scopus	TITLE-ABS-KEY (“academic advising” OR “student advising” OR “student orientation” OR “university orientation” OR “orientation program*” OR “career guidance” OR “academic guidance” OR “student support”) AND TITLE-ABS-KEY (“machine learning” OR “educational data mining” OR “learning analytics” OR “early warning system*” OR “student risk prediction”)
Web of Science	TS = (“academic advising” OR “student advising” OR “student orientation” OR “university orientation” OR “orientation program*” OR “career guidance” OR “academic guidance” OR “student support”) AND TS = (“machine learning” OR “educational data mining” OR “learning analytics” OR “Artificial Intelligence”)
ScienceDirect	(“academic advising” OR “student advising” OR “student orientation” OR “university orientation” OR “orientation program*” OR “career guidance” OR “academic guidance” OR “student support”) AND (“machine learning” OR “educational data mining” OR “learning analytics” OR “Artificial Intelligence”)
SpringerLink	(“academic advising” OR “student advising” OR “student orientation” OR “university orientation” OR “orientation program*” OR “career guidance” OR “academic guidance” OR “student support”) AND (“machine learning” OR “educational data mining” OR “Learning Analytics” OR “Artificial Intelligence”)
Google Scholar	(intitle: “academic advising” OR intitle: “student advising” OR intitle: “student orientation” OR intitle: “university orientation” OR intitle: “career guidance”) (“machine learning” OR “educational data mining” OR “learning analytics” OR “Artificial Intelligence”)

* Wildcard operator: retrieves all word variants of the same root (e.g., “program*” matches “program” and “programs”).

.

3.4. Selection Criteria and Eligibility

At this stage, after the formulation of the search query, a total of 1604 records were retrieved from the selected databases and search engines. In accordance with the PRISMA 2020 guidelines, a structured three-phase screening procedure was applied, as shown in Figure 3.

First, all records were compiled in Rayyan, a web-based tool designed to support collaborative screening in systematic reviews [21], using which duplicate entries were identified and removed (n = 523), resulting in 1081 unique records.

Second, titles and abstracts were screened independently by two researchers within Rayyan. Each record was classified as “Yes,” “No,” or “Maybe” based on the predefined inclusion and exclusion criteria provided in

Table 3. Inclusion and exclusion criteria.

	Inclusion Criteria	Exclusion Criteria
Population	The target population includes students enrolled in higher education institutions.	The target population is not students.
Article Type	Peer-reviewed journal articles and conference papers.	Theses, reports, short abstracts, Review.
Year of Publication	Studies published between 2018 and 2025.	Studies published before 2018.
Language	Publications in English.	Publications not written in English.
Scope/Domain	Academic advising, guidance, orientation, or decision support in higher education.	General learning analytics, administrative systems, or educational prediction without an advising component.
Methods/Techniques	The study applies ML or AI techniques to support student academic orientation.	The study does not apply ML or AI techniques.
Accessibility	The Article is accessible.	The Article is not accessible.

. Records marked as “No” were excluded at this stage, leading to the removal of 914 studies. Records categorized as “Maybe” were retained for further evaluation, and their full texts were consulted when necessary to resolve uncertainties.

Third, 167 reports were sought for full-text retrieval. Among these, 58 reports could not be accessed and were thus excluded. The remaining 109 full-text articles were assessed against the eligibility criteria, resulting in the exclusion of 82 studies due to scope misalignment or insufficient relevance. Ultimately, 27 studies satisfied all criteria and were included in the final synthesis.

Although snowballing is a common complementary strategy, it was not formally applied, as the initial search results were sufficiently comprehensive to cover the core literature within the scope of AI-enabled advising.

3.5. Analytical Framework

To systematically analyze the selected studies, an analytical framework was developed to structure the synthesis of the literature, as shown in Figure 4. The framework organizes the reviewed studies according to four complementary analytical dimensions. First, it considers the AI frameworks and computational paradigms adopted in academic advising systems. Second, it examines the academic advising functions addressed by these systems across the student lifecycle. Third, it analyzes system design and integration aspects, highlighting how different AI components are combined within advising architectures. Finally, the framework captures comparative and evaluative insights reported in the literature. This structure provides a consistent lens for examining how AI techniques are applied in academic advising and enables the classification and comparison of the reviewed studies.

4. Results

This section presents the results obtained from the analysis of the selected studies. The findings are organized according to the following categories:

•

Descriptive statistics.

•

Identified AI paradigms and frameworks.

4.1. Descriptive Statistics

Following the identification of eligible studies based on the PRISMA inclusion criteria [20]. This section presents an overview of the quantitative characteristics of the studies included in this review. Figure 5 illustrates the number of articles published per year, showing a significant evolution in interest in AI-based academic advising research, with over 50% of the selected studies published between 2024 and 2025. Figure 6 illustrates the distribution of the selected studies by country, showing that Morocco and the United Arab Emirates are the most represented in the reviewed literature. Figure 7 presents the distribution of publication types among the selected studies, where journal articles are predominant, while conference papers account for a smaller proportion of the reviewed contributions. Figure 8 shows the distribution of articles according to journal quartiles, offering insight into the quality levels of the publication venues represented, and indicating that most articles are published in Q1 and Q2 journals.

4.2. Identified AI Frameworks and Approaches

This section examines the methodological foundations of AI-driven academic advising by focusing on the computational paradigms adopted across the reviewed studies. The aim is to clarify how advising problems are framed from a technical perspective and to identify the dominant families of approaches used to generate advising decisions. Accordingly, the studies are grouped according to the primary logic through which advising decisions are generated. Based on this classification, five major paradigms emerge from the literature, as shown in Figure 9: Classical Machine Learning (CML), Recommender Systems (RS), Hybrid AI, Deep Learning (DL), and Generative AI (GenAI) approaches, with CML and Hybrid AI emerging as the dominant paradigms in the reviewed studies, as illustrated in Figure 10. These paradigms reflect different assumptions about how academic advising problems should be modeled and how student data should be interpreted. Moreover, 

Table 4. Summary of the reviewed studies.

Study	Techniques	Category	Dataset	Summary
[11]	Collaborative Filtering (CF) (user-based and item-based), Demographic Filtering (DF), Case-Based Reasoning (CBR), Ontology	Hybrid AI	Lebanon university graduate profiles (n = 869) Courses rating (n = 20,000)	Compares multiple RS and demonstrates that the hybrid approach (CF, CBR, and ontology-based methods) provides the most effective and personalized recommendations.
[22]	Reinforcement Learning (RL), Neural Network (NN), K-Means, Association Rule Mining (ARM), Support Vector Regression (SVR)	Hybrid AI	Student records & preferences	Introduces a Multi-agent advising framework (AMASIA) enabling adaptive advising, monitoring, and course suggestions.
[23]	Decision Tree (DT), Artificial Neural Network (ANN), Random Forest (RF)	CML	UAE HEI students (n = 428)	RF achieves the highest accuracy in predicting need for extra advising.
[24]	Neural Network (NN)	DL	Morocco high-school cohort (n = 72,010)	Predicts baccalaureate average to support orientation toward suitable university pathways.
[25]	Support Vector Machine (SVM), k-Nearest Neighbors (KNN), Logistic Regression (LR), DT, RF, Naïve Bayes (NB)	RS	High-school dataset (n ≈ 6000)	Predicts suitable specialties to improve matching and reduce dropout, RF achieves the highest accuracy.
[26]	DT, RF, SVM, CF	Hybrid AI	Algeria first-year students (n ≈ 520,000)	Introduces a hybrid RS to help students in Algerian universities select optimal academic choices by integrating ML and CF.
[27]	RF, Bayesian Belief Network (BBN), NN, DT, AdaBoost, curriculum graph	RS	Dubai university’s Student Enrollment and grades record data (n = 200)	Proposes an automated advising system combining curriculum graphs and performance prediction to generate personalized study plan. RF achieves the highest accuracy (86%).
[28]	RF, KNN, SVM, Gradient Boosting (GB), NB	CML	Kazakhstan IT students (n = 692)	Predicts IT specialty for IT university students in Kazakhstan, Gradient Boosting achieves the highest accuracy (~92%).
[29]	Rule-based Expert System, ML, chatbot	Hybrid AI	Student transcripts and university Rulebook data	Provides a hybrid advising framework combining policy rules, at-risk flagging, and FAQ chatbot to support academic progression.
[30]	Multi-view Genetic Programming (interpretable rules)	Hybrid AI	Virginia Common University Engineering data (multi-year historical data + LMS logs)	Multi-view Genetic Programming performs better on imbalanced data and produces clear, interpretable rules for underrepresented students.
[31]	Graph Neural Network (GNN), Recurrent Neural Network (RNN), Retrieval-Augmented Generation (RAG)	Hybrid AI	UGC university student dataset and course dataset	Introduces an integrated platform that delivers personalized, context-aware guidance across academic and career decisions, but also highlights emerging concerns around bias, transparency, and data privacy.
[32]	Large Language Model (LLM), RAG, Few-Shot Learning (FSL), Contextual embeddings	GenAI	University student’s dataset (performance + socio-demography)	Introduces LLM-based framework that delivers more personalized and context-aware major recommendations than traditional methods, while also surfacing important concerns around bias, transparency, and ethical deployment in academic advising.
[33]	ARM	CML	Institutional academic advising data (private dataset) student course histories, grades, advising records (exact size not specified)	Proposes an ARM-based approach that enables personalized and explainable course recommendations aligned with students’ academic histories, improving advising efficiency and supporting data-driven decision-making.
[34]	RF, XGBoost, GLMNET, KNN	RS	Colombia ICFES undergraduate students (n = 921,041)	XGBoost achieves the best performance (RMSE ≈ 30), while the inclusion of socioeconomic variables improves degree guidance.
[35]	Ontology-Based Model (EDUC8), K-Means, SWRL Rule-Based Reasoning	Hybrid AI	Greek HEI students (n = 200)	Proposes a hybrid semantic and ML framework that personalizes academic pathways while supporting early risk detection and academic planning.
[36]	Gradient Boosting Regression (GBR), CatBoost, LightGBM, Hierarchical Multi-Model Aggregation	RS	Student academic data (n = 101,330) Courses dataset (n = 600)	Proposes a hybrid framework that combines multiple specialized models through a hierarchical aggregation mechanism to generate accurate, constraint-aware course recommendations.
[37]	GPT-4, embedding-based semantic similarity	GenAI	Survey data of 18 students and 18 advisors	Evaluates GPT-4 recommendations in academic advising, with advisors rating outputs as helpful (≈3.9/5) and reporting partial agreement (39%) with human decisions, suggesting a complementary role rather than full automation.
[38]	Long Short-Term Memory (LSTM)	DL	FIU University students (n = 3328) and courses data (n = 647)	Proposes a session-based recommendation approach that generates semester-level course sets based on sequential enrollment patterns.
[39]	DT, LR, RF, KNN, SVM, AdaBoost, GB, XGBoost, CatBoost, Lasso, Ridge, Bayesian Ridge, Huber Regressor, RS	Hybrid AI	Public Moroccan high school graduates (n = 500) from 12 institutions, combined with colleges’ eligibility criteria.	Predicts admission probability and recommends alternative universities using RS strategies, with Huber Regressor outperforming other evaluated models.
[40]	RF	CML	UAE private university IT graduates (n = 105)	Predicts Cumulative Grade Point Average (CGPA) with over 92% accuracy and enables early identification of at-risk students from the second year.
[41]	ARM	CML	Institutional student dataset from Universitat Oberta de Catalunya (UOC) (n = 6559)	Introduces an ARM-based approach that identifies course enrollment patterns associated with changes in Grade Point Average (GPA), enabling early detection of at-risk students and more informed course sequence recommendations.
[42]	Formal Concept Analysis (FCA), ARM	CML	Lebanese University graduate profiles via survey (n = 448)	Identifies interpretable clusters linking interests, majors, and career outcomes, supporting academic orientation while highlighting scalability limits.
[43]	CF, Knowledge-Based (KB), CBR, Ontology	Hybrid AI	University graduates (n = 1000), High school course rating (n = 20,000)	Proposes the COHRS hybrid architecture to address cold-start and data sparsity issues, supporting accurate university major and career recommendations.
[44]	NB, DT, RF implemented using MapReduce (Hadoop framework)	CML	Not specified	RF implemented using MapReduce outperforms Naïve Bayes and Decision Tree models by balancing accuracy and scalability in large-scale orientation settings.
[45]	NB, SVM, RF, NN	CML	Student grades, absences, and attendance	Compares multiple classical ML models, with NB achieving the highest reported accuracy, supporting scalable automated academic orientation.
[46]	Technique for Order Preference by Similarity to Ideal Solution (TOPSIS), Analytic Hierarchy Process (AHP), Information Gain, Synthetic Minority Over-sampling Technique (SMOTE)	Hybrid AI	OULAD dataset (n = 32,594)	Compares multi-criteria decision-making approaches, showing that TOPSIS outperforms AHP, while SMOTE improves the performance of both methods.
[47]	KNN, DT, RF, SVM	CML	Oman MOHERI student records (n = 315,000)	Compares CML classifiers, with SVM outperforming other models and demonstrating the feasibility of large-scale, data-driven academic orientation.

summarizes the main characteristics of the studies included in the analysis.

5. Discussion

This section interprets the evidence synthesized from the 27 included studies and derives implications for AI-based academic advising in higher education. It examines how advising functions are distributed across the student lifecycle and how AI paradigms are positioned with respect to these functions, highlighting recurring design choices and evaluation tendencies reported in the literature. The section then consolidates the main research gaps and concludes with limitations of the current evidence base and directions toward more integrated, reliable, and institution-ready advising systems.

5.1. AI Paradigms Within Academic Advising

The reviewed studies reflect five broad computational paradigms, each associated with distinct advising contexts and design assumptions, as discussed in the following subsections.

5.1.1. Classical Machine Learning

CML remains a dominant methodological family in the reviewed literature, representing 33% of the included studies, as shown in Figure 10. CML is particularly prevalent for prediction and classification tasks where institutional data can be represented in a structured form. Analytically, a substantial subset of these studies—especially those grounded in CML—aligns with the CRISP-DM framework [48], which relies on a step-by-step analytical process combining data preparation, feature extraction, model training, validation, and the final reporting of performance metrics (Figure 11).

Within this paradigm, most studies rely on supervised learning methods, meaning that models are trained on historical student data where the outcome of interest is already known. This reflects both the availability of labeled institutional datasets and the need for decision processes that remain traceable and reproducible. Ensemble-based approaches are particularly frequent; for example, RF models are frequently adopted due to their robustness with heterogeneous features and their balance between predictive performance and interpretability. Several contributions report strong results when using RF for structured advising-related classification tasks, often achieving higher accuracy than other classical models [23,40]. At the same time, comparative evaluations suggest that performance is highly context dependent, as a single classifier consistently outperforms others across all datasets. In certain settings, GB, SVM, or even NB achieve comparable or better results, depending on the data characteristics and evaluation setup [28,44].

Beyond simple prediction tasks, CML is also applied to advising scenarios related to specialization choice or outcome estimation. The contribution of [28] compares several supervised classifiers to model specialization decisions among IT students, where the target label corresponds to the selected academic track; moreover, they indicate that GB performs best overall, while Random Forest and SVM remain competitive baselines. Studies of this kind underline that classical approaches remain effective when advising problems can be framed as structured classification tasks with clearly defined labels.

A smaller group of studies moves away from supervised prediction and instead adopts exploratory techniques, commonly referred to as unsupervised methods, where no predefined outcome variable is specified. Formal Concept Analysis (FCA) is used to reveal interpretable groupings of student trajectories and attribute combinations, allowing relationships between academic paths and outcomes to be examined from a descriptive perspective [42]. The rule-based ML method ARM is also used for a similar logic. The contributions of [33,41] show that this unsupervised technique can uncover frequent patterns and conditional relationships in student enrollment and performance data, emphasizing interpretability and descriptive insight rather than predictive accuracy.

In contrast, another set of contributions addresses scalability concerns, shifting the focus away from methodological novelty toward practical implementation constraints. In this context, several studies propose distributed solutions by deploying classification models in environments such as MapReduce. These studies show that classical classification techniques can be applied to large institutional datasets while maintaining acceptable execution times and stable performance, as demonstrated by [44,45].

Taken together, these contributions indicate that CML approaches are benefiting from methodological maturity and operational simplicity, performing well when advising concerns are narrowly defined. At the same time, their reliance on fixed prediction formulations limits their ability to accommodate evolving student trajectories or adaptive decision-making.

5.1.2. Recommender Systems

RS represent another application of AI that aim to assist users in making decisions from available options. These systems do not provide a single predicted answer but rather, they provide a set of recommendations based on user profile information, preferences, ratings, feedback, and previous interactions with the system [11]. In academic advising, this technique is most applicable when students are faced with several viable choices, such as selecting courses, academic programs or various academic pathways to follow.

In this reviewed literature, RS are most frequently used to assist in the advising task related to course selection, program matching, and pathway planning. The prevailing assumption is that students with the same profile or having a similar academic background will benefit from similar recommendations. Many of these systems rely on CF, a recommendation technique that exploits similarities between students or between historical advising cases to generate personalized suggestions. In this context, CF leverages student profiles, academic records, and curriculum structures to infer suitable academic choices. However, multiple studies indicate that RS in academic advising faces well-known challenges from the broader RS domain, notably cold-start situations and data sparsity, especially when historical data are limited or incomplete [11,26,35].

In an attempt to address these challenges, some contributions go beyond simply data-driven techniques by incorporating semantic structures and knowledge-based components into the recommendation process. In the article [43], the authors present a hybrid advising architecture, which integrates CF with ontology-based modeling and CBR, designed around historical student cases coupled with domain knowledge. The model creates nuanced recommendations that are explainable and personalized. A similar path is taken by [35] where ontologies are used to model adaptive learning pathways, which remain valid despite subsequent changes to the curriculum and the overall fluidity of the institutional context.

Some studies attempt to enhance the degree of adaptability by the integration of multiple recommendation techniques within a single framework. Ref. [36] implements a multi-model technique that synthesizes various decision-making processes to assist students in selecting appropriate study disciplines, while [26] adopts the same approach by merging multiple recommendation strategies to address various dimensions of student profiles and academic limitations.

RS, in some instances, are built to factor in some aspects other than individual preferences or academic performance. Ref. [46] merges advising directions with institutional and contextual restraints by describing them as multi-criteria rankings through the AHP and TOPSIS. In this way, a more practical alignment between quantitative and qualitative indicators can be achieved. From this methodological perspective, the reviewed studies suggest that recommender systems are mainly applied in advising contexts where choice is more prominent than guidance and where predictive outcomes are not the primary objective. Despite these advantages, most recommender-based advising systems still lack evidence of consistent real-world implementation and are evaluated in offline or simulated environments, a limitation also reported in the broader higher-education recommender-systems literature [12], where evaluation remains largely based on offline experiments, case studies, and lab-based settings rather than authentic institutional deployment. Moreover, few studies clarify how recommendations are assessed by human advisors integrated into advisor-facing decision processes, which is important for trust, adoption, and operational sustainability.

5.1.3. Hybrid AI Frameworks

In addition to the literature discussed above, another group of studies examines hybrid AI frameworks to handle the structural and contextual complexity of academic advising. Unlike most CML approaches that focus on improving a single predictive model at a time, hybrid frameworks rely on multiple computational paradigms—such as ML, rule-based reasoning, multi-agent coordination, and knowledge-oriented components—to provide more context-aware decision support. Rather than focusing solely on generating a single prediction for a student’s advising needs, these systems are designed as integrated architectures capable of handling heterogeneous datasets, institutional restrictions, and evolving student needs. Within this category, a smaller subset of studies incorporates RL to support adaptive and sequential advising processes. RL-based approaches learn advising strategies through repeated interactions and feedback, making them particularly relevant for longitudinal advising scenarios in which earlier interventions influence later academic outcomes.

This broader orientation is reflected in several implementation strategies across the reviewed studies. For example, ref. [22] employed the use of this concept in their Adaptive Multi-Agent System (AMASIA); the authors used Knowledge Query and Manipulation Language (KQML) to enable the exchange of advising knowledge among autonomous agents. Through this cooperative mechanism, personalized course recommendations and early interventions can be generated as students’ goals and academic trajectories evolve. Other studies emphasize hybridization for data integration and interpretability. In this vein, ref. [30] presented an interpretable multi-view DL architecture that integrates LMS and institutional data to support academic advising by identifying at-risk students requiring additional advising. Hybridization is also visible in frameworks that combine predictive analytics with institutional knowledge and counseling logic. For instance, ref. [29] conceptualized academic advising as a knowledge management problem, leveraging institutional expertise and academic policies to ensure that computer-generated recommendations remain compliant with institutional rules, while [39] employs a modular hybrid architecture that combines machine-learning–based admission prediction with recommendation components to deliver personalized guidance without tightly coupling data-driven models and symbolic reasoning. A related contribution is provided by [27], which combines student performance modeling and curriculum structure analysis to produce advising actions aligned with formal degree requirements. By integrating both student progress and rule-based curricular dependencies, this framework addresses practical constraints that are central to real-world academic program advising.

Together, these studies reflect the emergence of layered hybrid advising architectures that integrate intelligence and governance to support adaptive decision-making.

5.1.4. Deep Learning

While hybrid frameworks integrate multiple decision components, another line of research focuses specifically on representation learning through deep neural models, which motivates the emergence of DL-based advising systems. In the reviewed literature, DL techniques often appear as components within hybrid architectures, where neural models are combined with other paradigms such as rule-based reasoning, RL, or retrieval systems. DL has gained traction in academic advising research, especially in cases where traditional, indicator-based and structured methods become insufficient. Student data, in practice, are often heterogeneous and include disparate logs, interactions, partial academic records, and varying levels of engagement. However, DL is valuable not only for predictive accuracy but also, more importantly, for representation learning. This allows models to learn relevant features directly from data without relying on prescriptive indicators. This is particularly evident in research that integrates multiple forms of student data.

One of the most established applications of DL in academic advising concerns early intervention and student retention, where most studies aim to understand patterns of academic progression—how students move through the curriculum, when disengagement begins, and which patterns tend to precede negative outcomes. Although literature remains largely focused on risk prediction, neural models have begun to appear in several advising studies, particularly in hybrid architectures where representation learning complements traditional predictive approaches, as they can capture relationships that are difficult to model using classical approaches, particularly when relevant signals are distributed across multiple dimensions [24,31]. In this context, we have noticed that some of the related contributions employ multi-view DL methods that jointly learn behavioral engagement, demographic attributes, and academic performance data to strengthen early intervention systems. Despite these advances, the issue of explainability remains central in academic advising contexts, where transparency in decision-making is critical. This tension between predictive performance and interpretability is highlighted by [30] and appears throughout much of the literature.

DL techniques are also applied in recommendation-based advising systems, although their role differs from that in early warning applications. In these settings, the focus shifts from identifying at-risk students to determining which academic options should be recommended under specific institutional or curricular constraints. For example, ref. [38] uses session-based and sequence-aware DL models, where recommendations are shaped by the order and timing of student activities rather than by a fixed profile. This allows advising to adapt at the course level as engagement patterns evolve. In summary, these studies differ in how personalization is defined; they share the use of deep representations to encode student histories in a form suitable for advising.

5.1.5. Generative Artificial Intelligence

More recently, attention has shifted toward generative models, which are discussed primarily in terms of interaction and explanation rather than prediction alone. In this regard, LLMs are examined as advisory agents capable of producing natural language recommendations and providing accompanying explanations.

Study [37] exemplifies this trend through a comparative analysis combining GPT-4 and conventional major recommendation systems. Its findings suggest that the primary contribution of generative models lies in the advisory dialog, including how recommendations are expressed, how systems respond to follow-up questions, and how guidance is scaffolded. At a broader level, ref. [32] examines recommendations for academic fields of study by mapping learned representations of student profiles to corresponding disciplines.

Overall, GenAI approaches are becoming increasingly visible in academic advising research, particularly in studies that emphasize explainability, interaction, and alignment with the advising process. Unlike RS, which typically rely on historical patterns in student preferences and behavioral data to produce structured recommendations, GenAI uses LLMs to produce contextualized natural-language responses and conversational guidance. Many of the proposed approaches, however, remain modular advisory components rather than fully integrated advising systems, with institutional constraints still only weakly reflected in current implementations.

5.2. Functional Analysis: The Advising Lifecycle

This subsection shifts the focus from the AI components examined in the reviewed literature to the functional roles that AI systems play across the student lifecycle, as depicted in Figure 12. The literature describes a clear progression of interventions: first, to ensure that the right student is placed in the right academic program (Academic Career Advising); second, to identify students at risk of academic difficulty at an early stage (Early Warning and Risk Detection); and sustaining and optimizing student performance throughout the program beyond risk identification alone. (Academic Progress Advising).

Table 5. Classification of reviewed studies by academic advising function.

Study	Academic Function	Primary Objective	Description
[11,24,25,26,28,31,32,34,37,39,42,43,44,45,47]	Academic Career Advising	Major selection, university pathway orientation, and career recommendation.	Primarily rely on CML and hybrid RS, with recent studies beginning to explore GenAI approaches.
[23,29,30,40,41,46]	Early Warning and Risk Detection	Identification of at-risk students and early academic risk prediction.	Primarily rely on supervised classification approaches for early academic risk detection.
[22,27,33,35,36,38]	Academic Progress Advising	Course sequencing, curriculum planning, and progression monitoring.	Focus on supporting course planning and academic progression through structured recommendation approaches.

summarizes the distribution of the reviewed studies according to these three academic advising functions. Each study is assigned to a single category based on its primary decision-support objective, even when secondary functions are present.

As shown in Figure 13, the reviewed literature is dominated by Academic Career Advising (56%, n = 15), while Risk Advising remains comparatively underrepresented (22%, n = 6), and Academic Progress Advising accounts for a quarter of the studies (22%, n = 6).

At this point in the literature, the student advising life cycle begins with initial orientation, whose purpose is to align student capabilities with program demands in order to prevent early attrition. The studies conducted by [24,46] anchor this phase by using historical indicators to support STEM choice and pathway alignment, while the studies conducted by [25,47], show how such systems can be scaled to handle large cohorts at a national level. Most importantly, the literature suggests that efficiency should be considered alongside equity. In particular, ref. [46] emphasizes that the automated advising process should address historical asymmetry issues in the datasets to ensure that student advising processes increase diversity and do not perpetuate existing biases.

Once students are enrolled, the focus turns to early warning and risk detection. To limit dropout rates, refs. [23,40] introduce ‘early warning windows,’ targeting students in need while intervention is still possible.

Beyond risk detection, another set of studies addresses academic progress advising more directly. In this context, refs. [36,38] move beyond simple course lists; they map the entire degree as a sequence, ensuring that plans are logical and fit within university constraints. Ref. [41] extends this logic through ‘GPA Optimization,’ recommending specific course combinations to statistically maximize cumulative grades.

The maturation of AI technologies has enabled a shift toward holistic support systems that address needs extending beyond the classroom. Ref. [31] exemplifies this by using conversational agents to demystify complex financial aid policies, effectively removing administrative hurdles that frequently drive attrition. On the professional front, refs. [11,42] bridge the gap between curriculum and career. Instead of theoretical guidance, their systems utilize historical alumni data to map viable professional routes, ensuring that academic decisions align with current market opportunities.

5.3. Comparative Assessment of AI Paradigms

This subsection provides a comparative assessment of the AI paradigms reviewed, examining their distribution across advising functions and evaluating their respective strengths, limitations, and institutional suitability.

5.3.1. Paradigm Distribution Across Advising Functions

This subsection examines how different AI paradigms are distributed across academic advising functions and what this distribution reveals about their comparative strengths, limitations, and institutional relevance. Rather than comparing isolated algorithms, the discussion focuses on broader methodological tendencies, including where particular approaches tend to cluster, what kinds of advising tasks they are most often used to support, and how their usefulness changes across the student lifecycle. As shown in Figure 14, academic career advising is the most methodologically populated area and is dominated by classical supervised learning techniques, particularly RF, SVM, DT, and KNN. This pattern suggests that orientation and specialization decisions are still largely framed as structured classification or ranking tasks based on student records and measurable academic indicators.

Some techniques also appear across more than one advising function, especially RF and DT, which are used in both Academic career advising and early warning and risk detection, indicating continued reliance on methods that are comparatively robust and practical for institutional use. In the case of early warning and risk detection, the range of techniques is more limited, with the literature concentrating on a smaller set of established approaches designed to support timely and actionable intervention.

Academic progress advising shows a different pattern. Although it includes fewer studies overall, it reflects a relatively broader mix of techniques, including Bayesian models, LSTM, ontology-based methods, ARM, and LLM/RAG-based systems. This suggests that progression-related advising often requires more flexible forms of support, such as sequencing, planning, and personalized pathway guidance, rather than a single prediction outcome.

To extend these tool-level observations, Figure 15 summarizes the findings at the paradigm level and highlights how broader AI families tend to align with specific academic advising functions.

5.3.2. Strengths and Limitations of AI Paradigms

A comparative assessment highlights that no single AI paradigm is universally optimal across all advising functions. Instead, each paradigm offers distinct strengths and limitations depending on the advising task, data context, and level of institutional constraint.

Table 6. Comparative analysis of AI paradigms for academic advising.

AI Paradigm	Advantages	Limitations and Challenges	Most Suitable Advising Functions
CML	• High predictive accuracy on structured data. • Relatively low computational cost. • Mature evaluation practices and easier deployment in institutional settings. • Strong performance for classification and risk prediction	• Limited personalization • Heavy reliance on feature engineering • Limited support for sequential and long-term planning tasks • Often insensitive to contextual or preference-based factors	• Academic Career Advising • Early warning and risk Advising
RS	• Strong personalization capabilities • Effective for preference-aware and similarity-based guidance • Well suited for course selection and pathway optimization • Intuitive outputs for end users	• Cold-start issues and dependence on data density • Limited effectiveness for early-stage risk detection • May ignore institutional constraints without hybridization	• Academic Progress Advising
Hybrid AI	• Combine data-driven prediction with domain knowledge • Improved explainability and policy compliance • Flexible across multiple advising objectives • Robust to institutional constraints	• Higher system complexity—Increased design and maintenance effort • More difficult to evaluate consistently • Require expert knowledge for rule or ontology design	• Academic Career Advising • Academic Progress Advising • Early warning and risk Advising
DL	• Strong capacity to model complex, non-linear patterns • Effective for large-scale and multi-source educational data • Improved performance in early warning and dropout prediction	• Low interpretability • High computational cost and data requirements • Limited transparency for advisors and decision-makers; • Reproducibility concerns	• Early warning and risk Advising
GenAI	• Natural language interaction • Strong explanatory and conversational capabilities • Improved accessibility and user engagement • Effective for guidance synthesis and administrative support	• Reliability and hallucination risks • Limited evaluation standards • Weak grounding without retrieval or rule-based mechanisms • Unsuitable as a sole decision engine in high-stakes advising	• Academic Career Advising • Academic Progress Advising

summarizes the key strengths, limitations, and suitability of each paradigm across different academic advising contexts.

As summarized in

Table 7. Comparative evaluation of AI paradigms used in academic advising across interpretability, personalization, implementation complexity, and policy/governance compatibility (scores range from 1 = low to 5 = high).

AI Paradigm	Interpretability	Personalization	Implementation Complexity	Policy and Governance
CML	4	2	2	4
RS	3	5	3	3
Hybrid AI	4	4	5	4
DL	1	3	4	2
GenAI	2	5	4	2

, CML remains the most operationally mature approach, particularly in structured prediction settings where implementation simplicity, lower computational cost, and established evaluation practices are important. RS contribute stronger personalization and are especially relevant when advising is framed as a choice-support problem. A similar conclusion is reported in [12], where recommendation tasks are centered mainly on academic choice and course-related decision support. Hybrid AI stands out for its ability to combine prediction with institutional knowledge, making it particularly suitable for contexts where policy alignment, explainability, and contextual adaptation are essential. By contrast, DL offers stronger capacity for modeling complex and multi-source data, but remains constrained by interpretability, computational burden, and reproducibility concerns. GenAI introduces important advances in interaction and explanation, yet its current role remains more supportive than decision-critical because of hallucination risks, limited grounding, and weak evaluation maturity. Taken together, these patterns suggest that comparative value in academic advising is best understood not in terms of overall superiority, but in terms of functional fit, governance compatibility, and institutional readiness.

5.3.3. Implementation, Governance, and Ethical Challenges of AI in Academic Advising

The reviewed literature reveals a clear tension between technical performance and institutional usability. As summarized in

Table 8. Cross-cutting challenges identified in the literature and their implications for the deployment of AI-driven academic advising systems.

Cross-Cutting Challenge	Main Implication for Academic Advising
Offline evaluation dominance	Most systems are validated retrospectively, limiting evidence of real-world effectiveness and sustained institutional impact.
Bias and limited transparency	AI recommendations may reproduce historical inequities and can be difficult for advisors or students to interpret, justify, or contest.
Privacy and data governance	The use of academic and behavioral data requires strong safeguards to ensure confidentiality, responsible access, and regulatory compliance.
Weak institutional integration	Many systems remain insufficiently integrated with advising workflows, curriculum structures, and formal academic regulations.
Model maintenance and change sensitivity	AI systems require continuous updating to remain aligned with evolving curricula, program requirements, and student populations.
GenAI hallucination and weak grounding	LLM-based systems may generate fluent but inaccurate or institutionally inconsistent advice unless grounded in validated institutional data and rules.

, many of the main barriers to AI-enabled academic advising extend beyond model accuracy and instead concern implementation, governance, and institutional fit. This observation is reinforced by the fact that 81.5% of the studies evaluated their systems only in offline settings, showing that the field remains dominated by retrospective validation rather than real-world deployment. Related review evidence also highlights limited practical implementation, weak external validation, and persistent ethical and institutional integration challenges beyond predictive performance alone [15].

In academic advising, these challenges are especially important because AI operates in a high-stakes environment shaped by formal rules, curriculum structures, and institutional responsibilities toward students. The main cross-cutting barriers include:

•

Offline evaluation dominance.

•

Bias and limited transparency.

•

Privacy and data governance constraints.

•

Weak institutional integration.

•

Model maintenance under changing academic conditions.

•

GenAI hallucination and weak grounding.

Taken together, these barriers suggest that progress in AI-enabled academic advising should be assessed not only in terms of model performance, but also in relation to governance readiness, transparency, and institutional reliability.

5.3.4. Limitations and Future Directions in AI-Academic Advising

The reviewed literature highlights several limitations that continue to constrain the development of AI-enabled academic advising systems.

•

Context instability and data drift: Advising systems often rely on institutional data that do not remain stable over time. Changes in curricula, grading regimes, academic regulations, and advising policies may reduce the validity of learned patterns and weaken system reliability.

•

Fragmented task coverage: Many studies address only one advising task, such as major selection, risk detection, or course recommendation, without linking these functions across the student lifecycle. This limits continuity and reduces the ability of current systems to support coherent advising.

•

Weak institutional grounding: A large part of the literature emphasizes predictive performance more than institutional rules, academic structures, and advising workflows. As a result, some systems remain technically effective but operationally difficult to embed in real practice.

•

Limited adaptability: Most systems are developed for specific datasets or institutional settings, which restricts transferability. Their effectiveness may decline when student populations, academic structures, or advising requirements change.

•

Limited real-world validation: Much of the evidence is still based on offline evaluation rather than live deployment, leaving uncertainty about how these systems perform under real institutional conditions.

To address these limitations, this study proposes a conceptual framework for AI-enabled academic advising that emphasizes contextual grounding, lifecycle-wide support, and stronger institutional alignment. The framework is organized into four layers:

•

Institutional knowledge and data: integrates academic data, institutional policies, and formal knowledge resources to provide context-aware foundations for advising.

•

Lifecycle agents modules: structures advising support around major lifecycle functions, including orientation, early warning, and academic progress.

•

Multi-Paradigm Orchestrator: coordinates the use of different AI paradigms according to the advising context and task requirements.

•

Human–AI decision support: ensures that recommendations remain interpretable, reviewable, and subject to human oversight.

Future research should focus on validating this framework in real academic settings, with particular attention to contextual adaptation, human oversight, transparency, and long-term maintainability. Further work is also needed to examine how integrated, institution-aware architecture can remain reliable under changing curricula, policies, and student needs (Figure 16).

6. Conclusions

This systematic literature review examined 27 peer-reviewed studies published between 2018 and 2025 to clarify how AI is being used in academic advising across higher education. The findings show that AI in this field is shaped by five main paradigms—CML, RS, Hybrid AI, DL, and, more recently, GenAI—and that their value depends less on overall technical superiority than on their fit with specific advising functions and institutional contexts.

In response to RQ1, the review shows that CML and Hybrid AI remain the most established approaches, while RS are more relevant for decision-support tasks, DL is useful for modeling complex and multi-source student data, and GenAI is emerging mainly as a conversational and explanatory support layer.

With respect to RQ2, our review reveals that AI applications are unevenly distributed across the student lifecycle. Most studies focus on early advising functions, particularly academic orientation and program selection, whereas fewer studies address continuous academic progression or long-term risk detection in an integrated way.

Regarding RQ3, the comparative analysis indicates that no single paradigm is universally optimal; rather, each approach offers distinct advantages and limitations depending on the advising task, available data, and degree of institutional alignment.

At the same time, the literature remains constrained by significant limitations, including the dominance of offline evaluation, fragmented task coverage, weak institutional grounding, limited adaptability, and insufficient real-world validation.

These gaps suggest that future research should move beyond isolated task-specific models toward integrated, institution-aware advising systems. In this respect, the conceptual framework proposed in this study offers a practical direction by combining institutional knowledge, lifecycle-based advising modules, multi-paradigm orchestration, and human oversight. Future work should focus on validating such frameworks in real academic settings and on ensuring that AI-enabled advising systems remain transparent, adaptable, and reliable under changing academic conditions.