Early detection and personalized academic support using a predictive chatbot for student success

Early warning systems and predictive analytics

Early interventions are particularly critical in online and blended learning environments, where timely support can prevent students from falling behind. Integrating machine learning-based predictive analytics helps to identify at-risk students before problems escalate (Arévalo-Cordovilla & Peña, 2024). For instance, a study at a large public university showed that referring struggling economics students to a Student Success Center improved their final exam scores, emphasizing the value of early targeted interventions (Gordanier, Hauk & Sankaran, 2019). Moreover, the advent of smart campus technologies, which leverage AI and data analytics, has enhanced the precision of early warning systems. These tools collect and analyze various data points to generate tailored recommendations that improve learning outcomes (Villegas-Ch, Arias-Navarrete & Palacios-Pacheco, 2020). Similar efforts are underway in engineering education, where analytic frameworks have been developed to identify potential dropout situations, illustrating the promise of predictive models for early risk detection in specific programs (Mussida & Lanzi, 2022).

As access to higher education expands, understanding and supporting diverse student populations has become increasingly urgent. Institutions face new challenges in adapting to various needs while maintaining academic standards (Kocsis & Molnár, 2024). Research has shown that social identity profoundly affects academic outcomes. Students experiencing “disidentification” (i.e., negative perceptions of their role as students) tend to underperform and drop out at higher rates, an effect particularly pronounced among those from lower social class backgrounds (Matschke, de Vreeze & Cress, 2023).

The role of conversational agents in education

Traditional interventions often rely on generic approaches that fail to accommodate learners’ diverse backgrounds and needs. In response, there is growing interest in personalized strategies facilitated by conversational agents, such as chatbots (Gupta et al., 2019; Liu et al., 2023). Chatbots have rapidly evolved from administrative tools that answer questions about courses and deadlines (Okonkwo & Ade-Ibijola, 2021; Al-Abdullatif, Al-Dokhny & Drwish, 2023) to advanced educational support roles. Currently, chatbots provide personalized feedback, tutoring, and assessment by adapting to specific learning styles (Yang & Evans, 2019; Pérez, Daradoumis & Puig, 2020; Puertas, Mariscal-Vivas & Martínez-Requejo, 2023).

Evidence from various educational domains supports the growing impact of these tools on learning. For example, in nursing education, an AI-driven chatbot enhanced students’ interest in learning and independent study skills (Han, Park & Lee, 2022), while a similar tool at a Vietnamese university improved student-centered learning and engagement (Nghi et al., 2024). In engineering, chatbots have been used since 2008 in subjects such as thermodynamics and have been shown to facilitate knowledge acquisition, enhance motivation, and foster self-directed learning (Maladzhi et al., 2023; Bravo & Cruz-Bohorquez, 2024). These pedagogical conversational agents (PCAs) are more effective when a “common ground” is established with the learner (Tolzin et al., 2023). Furthermore, gamified chatbots have been shown to outperform traditional versions in boosting academic performance and motivation (Xu et al., 2024), and efficient designs can yield high accuracy in student interactions (Elragal et al., 2024) and enhance learning outcomes through personalized feedback (Ashok et al., 2021).

The gap: integrating prediction with intervention

The integration of predictive models with AI-based chatbots represents a significant advancement in the field of educational technology. Predictive analytics can identify students at risk of underperformance or dropout well before these issues manifest (Arévalo-Cordovilla & Peña, 2024). When coupled with AI-driven chatbots capable of proactive engagement, these tools facilitate early, individualized intervention (Ma et al., 2024). However, implementation challenges persist, such as privacy concerns and ethical considerations, requiring careful evaluation to ensure that these approaches effectively support diverse student groups (Wang et al., 2023).

Although early warning systems have shown potential for targeted interventions (Buschetto Macarini et al., 2019; Baneres, Rodríguez-Gonzalez & Serra, 2019), and AI-based personalization holds promise (Maghsudi et al., 2021), a persistent gap exists in the literature. In the era of digital transformation, many studies on AI chatbots have focused on student perceptions rather than integrating early predictive modeling (Monrad Schei, Møgelvang & Ludvigsen, 2024). Similarly, research on tools such as MoodleBot highlights improved engagement but stops short of linking these improvements to tailored interventions based on predictive insights (Neumann et al., 2025). Addressing this gap is critical for realizing the full potential of AI-based tools in comprehensive educational frameworks.

Design of the study

This study was structured as a pilot investigation to assess the technical feasibility and initial user engagement of an integrated early alert and automated intervention system. The research design follows a quantitative, single-group approach focused on operational metrics, such as identification coverage, system latency, and student response rate, rather than pedagogical outcomes. Because the pilot was conducted before the final course results were available, predictive accuracy could not be evaluated. Instead, system reliability and user responsiveness were used as proxies for functional validity.

The primary goal of this phase was to establish proof-of-concept and determine whether the system could operate reliably in a real-world educational setting and successfully engage students, serving as a prerequisite for subsequent studies that will integrate academic outcome data for full predictive validation (Monrad Schei, Møgelvang & Ludvigsen, 2024).

Consequently, this study did not include a control group or a comparison with human tutor interventions. Its scope is intentionally limited to answering foundational questions regarding system performance and student receptiveness. Therefore, a pilot test controlled by the researchers was conducted to collect the necessary operational metrics for preliminary validation.

Objective of the study

To design, implement, and preliminarily validate an automated educational intervention system that combines a previously validated, high-performance predictive model with an AI-driven chatbot to identify at-risk students early and provide them with personalized tutoring.

System functionality details

The developed system consists of two main components: a predictive model and an intervention chatbot.

Tools and technologies used

The development and implementation of the system were performed using the following technologies and dependencies:

• Node.js v20.18.0: JavaScript runtime environment used for the development of the chatbot backend.

• Baileys—TypeScript/Javascript WhatsApp Web API: Library for interacting with the WhatsApp Web API and enabling the sending and receiving of automated messages.

Main Dependencies:

• @hapi/boom: Structured management of HTTP errors.

• @prisma/client and prisma: ORM interacts with PostgreSQL database.

• @whiskeysockets/baileys: Interface for WhatsApp Web API.

• dotenv: Environment variable management.

• pino and pino-pretty: Logging and formatting for system performance monitoring.

• qrcode: Generation of QR codes for WhatsApp authentication.

API: Used to integrate generative AI capabilities using the GPT-4o-mini model, which enables natural and contextual responses.

Syllabus Vectorization: Processing and transformation of syllabi into vectors using semantic search algorithms, which facilitates precise and contextualized responses.

Implementation procedures

Chatbot interaction flow

To provide a comprehensive understanding of the chatbot’s operational structure, the following flowchart illustrates the sequential and interconnected processes that drive its functionality. The diagram represents the end-to-end flow, starting from the student’s risk identification using the predictive model for the delivery of personalized academic support.

Figure 1 shows a flow diagram that encapsulates the operational logic of the chatbot.

Data, code, and schema availability

The full source code for the chatbot implementation, including the Node.js backend, the cron job scripts used for data extraction and prediction, and the complete SQL database schema, is openly available on Zenodo (Arévalo-Cordovilla & Peña, 2025b).

In accordance with institutional data protection policies and the sensitive nature of student academic records, the original student-level dataset used in this study cannot be shared, even in an anonymized form. However, the provided database schema and example configuration files are sufficient for other researchers to adapt the code to their institutional data while preserving student privacy.

Interaction and general use of the chatbot

During the 27-day trial period, the chatbot was successfully integrated into the Digital Competencies and Technological Strategies course (Master’s in Education, Cohort II-2024, Parallel B3: N = 108 students). The system identified and contacted 39 at-risk students, resulting in 742 messages being exchanged (355 sent by the chatbot and 387 by users). This high level of interaction indicates significant student participation. To illustrate the nature of these conversations, a typical anonymized interaction is shown in Fig. 2.

The chatbot demonstrated a fast response time with an average generation time of 12.9 s per response, allowing smooth conversations without delays. The chatbot processed more than 1.36 million tokens during the pilot test, without any issues. Notably, 100% of at-risk students were contacted by the chatbot, achieving full intervention.

Overall, these results suggest that integrating the chatbot led to a high level of student engagement and that the system could sustain frequent interactions with short response times, which is a critical requirement for an effective educational conversational agent in real-world settings. This is further supported by the data presented in Table 1, which provides a summary of the key performance metrics.

Preventive intervention time

The chatbot demonstrated excellent performance in terms of the speed of intervention after the first alert signal. The average time between the automatic detection of an at-risk student and the chatbot’s first support message was 0.11 min (~6.6 s), with a median of 0.09 min (~5.4 s). This near-instantaneous response means that, upon identifying a student with potential difficulty, the system acts within seconds, providing timely assistance, as shown in Fig. 3. The small difference between the mean and median suggests that there were virtually no atypical delays in the interventions, which were consistently rapid in all cases.

In our study, the chatbot achieved reaction times well below one minute, far exceeding the recommendation for immediate intervention. Therefore, our results confirm that a rapid preventive intervention through chatbots is feasible and potentially valuable in addressing academic issues before they escalate, aligning with best practices for early warning systems in education.

Response rate to initial intervention

The response rate of the students to the first contact with the chatbot was remarkably high. Of the 39 students who were initially contacted, 35 responded to the initial message, resulting in an 89.74% response rate for the first stage. In other words, nearly nine out of ten students engaged in conversations with the agent following the initial intervention, demonstrating a high level of receptiveness and commitment from the users. The average response time for students to respond to the first message was approximately 5.9 h. This indicates that while most did not respond immediately, many did so on the same day they received the initial notification, which is reasonable given that messages could have arrived outside class hours or at unexpected times.

Notably, only 10.26% of the students did not respond to the chatbot on their first attempt, which is a considerably low percentage compared with other automated interventions. The lack of an initiall response is often an indication that something is failing in the design or relevance of the contact message. In this case, the low level of non-response suggests that the chatbot’s initial message successfully captured students’ attention and overcame potential adoption barriers. This indicates that the initial communication strategy was effective in motivating students to engage.

Establishing this communication channel is a crucial step, as a system that fails to attract student participation is unlikely to have a positive effect on student performance. The fact that a vast majority of students voluntarily interacted with the chatbot from their first contact demonstrates that chatbots can be viable mediators in educational interventions, aligning with studies that have shown their high acceptance in providing personalized and relevant support to at-risk students.

Chatbot efficiency

The collected efficiency data showed that the chatbot exhibited strong technical performance throughout the pilot test. On average, the student queries contained approximately 3,495 tokens (model-specific tokens as processed by the OpenAI API), whereas the chatbot responses averaged approximately 171 tokens. This indicates that chatbots can process complex queries and provide concise, focused answers. The average response time was approximately 12.9 s per query, even with variations in the daily volume. Throughout the trial period, the chatbot handled 355 interactions without any noticeable decrease in response speed and processed more than 1.36 million token. No significant increases in latency were observed, even on high-usage days, suggesting that the system scaled effectively during demand peaks. The technical performance indicators demonstrated that the chatbot was robust and reliable in real-world settings. It can maintain low response times even when handling large amounts of information and complex queries, thereby demonstrating the efficiency of its design. This finding is crucial because the technical challenge of maintaining AI performance under real-world conditions is often underestimated, which is essential for intervention scalability. In our case, the chatbot operated in a stable and effective manner, consistent with previous reports of well-designed conversational systems that achieved high-quality interactions with students.

In summary, the results indicate that the chatbot is not only theoretically viable for providing personalized support but also meets the practical operational efficiency requirements necessary for sustainable integration into real educational environments (Fig. 4).

Usage patterns and communication modality

Regarding communication methods, students showed a clear preference for text-based over audio communication. Of the total recorded interactions, 95.6% were text messages (370 messages), compared with 4.4% for voice messages (17 messages). This indicates that written communication is the primary channel for most users, likely because of the familiarity and convenience of text messaging during online learning. Although the option of sending voice messages was available, its use was limited to specific cases, suggesting that students predominantly chose to type their queries and read the chatbot responses. This result aligns with expectations in educational chatbot environments, where interactions typically occur primarily throughh text, highlighting that written language remains the preferred medium for immediate academic support.

In addition, significant temporal patterns emerged in chatbot usage, highlighting its role in supporting asynchronous study. The distribution of messages was not uniform, with substantial activity during non-traditional working hours. Specifically, 157 messages were exchanged during the early morning hours (between 12:00 and 5:59 AM). The highest number of interactions occurred at 3:00 AM, with a peak of 52 messages recorded in a single hour. Crucially, this activity was not driven by a single individual but was generated by 12 distinct students, representing nearly a third (30.8%) of the entire cohort. Furthermore, this late-night usage was a recurring pattern, with interactions at this specific hour recorded on seven different days during the 27-day trial. This finding strongly suggests that the chatbot fills a critical support gap by providing assistance during key independent study times when traditional academic resources are unavailable.

This suggests that many students took advantage of the chatbot’s 24/7 availability to seek assistance outside regular academic hours, possibly while studying independently or working on assignments late at night. The ability to access online support at flexible times appears to reinforce students’ self-study strategies, aligning with previous findings that personalized feedback through chatbots fosters student autonomy and independent learning.

Overall, the detected usage patterns highlight one of the key advantages of educational chatbots: providing immediate assistance at any time and adapting to students’ study needs and habits, whether through quick text responses for discreet inquiries or audio options for cases in which users prefer verbal communication. This underscores the chatbot’s flexibility in integrating into students’ diverse routines, offering personalized support whenever needed, without the time constraints of traditional tutoring resources, as shown in Fig. 5.

Ethical considerations

This study was conducted in accordance with the institutional regulations for research involving academic records at the Universidad Estatal de Milagro (UNEMI). Such research is overseen through formal institutional authorizations rather than numerical IRB-style protocol codes.

The use of student data for this doctoral research project was formally authorized in 2023 through a series of official memoranda: UNEMI-FACI-2023-0423-MEM (Faculty of Science and Engineering authorization), UNEMI-VICEACADFYG-2023-0586-MEM (Academic Vice-Rectory authorization), and UNEMI-R-2023-2595-MEM (Rector’s Office authorization). These authorizations cover the broader, multi-year doctoral project on predictors of academic performance in higher education, under which the present pilot study with the 2024 cohort is conducted. The 2024 cohort analyzed in this study corresponds to a later phase of the same approved project, and no changes were made to the scope, type of data, or risk level originally authorized.

UNEMI does not issue numerical IRB-style approval codes; instead, ethical oversight is executed through formal institutional approvals, which are archived and available upon request. All student data used in the predictive modelling and interaction log analyses were fully anonymized prior to analysis, and no personally identifiable information (PII) was retained in the research database. In addition, informed consent was obtained from all the student participants involved in the chatbot pilot intervention.

Comparison with previous studies

The findings of this study largely align with the trends identified in the literature on educational chatbots and learning analytics. First, the high student participation rate observed (89.74% of at-risk students responded to the automated intervention) is consistent with previous reports highlighting students’ positive attitudes toward chatbots as an academic support tool (Shoufan, 2023; Hind et al., 2023; Monrad Schei, Møgelvang & Ludvigsen, 2024). Various studies have documented that well-designed chatbots can enhance motivation, engagement, and student performance by providing personalized feedback and assistance. For example, Xu et al. (2024) gamified chatbots significantly increased student motivation and participation compared to traditional approaches, which aligns with the positive responses and recurring chatbot use observed in our pilot study. Similarly, (Al-Abdullatif, Al-Dokhny & Drwish, 2023; Huang, Lee & Kwon, 2023) it was found that integrating an educational chatbot had favorable effects on learning strategies and student autonomy, which is reflected in our context, where the conversational agent provided on-demand personalized tutoring.

However, our approach of combining predictive early warning models with automated chatbot interventions addresses the gap identified in previous studies. Recent studies have emphasized the potential of predictive analytics to anticipate academic risks and chatbots to offer personalized support; however, these two areas have often operated independently of each other.

Villegas-Ch, Arias-Navarrete & Palacios-Pacheco (2020) proposed theoretical architectures for integrating AI-powered chatbots into university settings; however, empirical evidence of their effectiveness is still lacking. In this regard, our results provide concrete data that complement the literature by achieving near-instant proactive intervention following risk detection (average time of ~0.1 min) and successfully engaging with most of the contacted students. We demonstrate how the synchronization of prediction and intervention can lead to tangible improvements.

This empirical finding validates the recommendations of previous studies that advocate a shift from reactive approaches to integrated proactive strategies, indicating that an effective combination of analytics and chatbot interventions is viable and beneficial in practice. However, unlike some early studies that focused primarily on usability perceptions without linking them to educational outcomes, this study provides initial evidence of operational impact (e.g., response times and student response rates), thus bridging the gap between the theory of automated intervention and its real-world classroom application.

Practical implications

The results of this pilot study have significant implications for higher education, particularly in terms of early intervention strategies, personalized learning, and dropout reduction. The system’s abilityy to react within seconds of identifying at-risk students suggests a paradigm shift toward proactive prevention. Thus, institutions can implement intelligent agents that continuously monitor student progress and provide immediate assistance as soon as the warning signs emerge. The literature has highlighted that early and timely intervention can disrupt patterns of poor performance before they worsen, thus maximizing their positive impact. Consistently, our chatbot responded with immediacy that would be unfeasible through traditional human intervention, establishing a help channel at a critical moment when the student was still receptive to help. This immediacy is particularly valuable in preventing minor academic setbacks from escalating into major problems. For example, in our pilot study, chatbot activity was observed even during unconventional hours (peak usage at 3:00 AM) when students typically lack access to available instructors. Having an automated tutor available 24/7 not only facilitates addressing questions outside regular hours but also helps reduce academic anxiety by providing continuous support, which is recognized as beneficial in research on automated learning assistance.

It is important to properly interpret the value of a system’s immediacy in this context. While the chatbot initiated contact in seconds, the average student response time was approximately 5.9 h. This does not diminish the system’s value but rather reframes it: the primary benefit is not forcing a synchronous, real-time conversation but ensuring 24/7 availability and eliminating institutional delays. The system’s speed ensures that a support channel is opened the moment a risk is detected, making help available for the student to access whenever they are ready—whether immediately or hours later, as evidenced by the peak usage at 3:00 AM. This asynchronous support model respects student autonomy and schedules, providing assistance at their point of need, which is a distinct advantage over human-only interventions that are limited by traditional work hours.

In summary, integrating a chatbot with predictive analytics can transform student support efforts by shifting from delayed reactive interventions to personalized real-time guidance, with the potential to improve student retention and academic success on a large scale.

Another key implication is the level of learning personalization enabled by AI-powered chatbots. Unlike traditional early alerts, which typically only notify tutors or generate generic warnings, the system presented herein provides feedback and resources tailored to the specific needs of at-risk students (Davies et al., 2021). This individualized approach, such as offering study tips or review materials related to the difficulties detected by the model, addresses the growing demand to accommodate student diversity. Studies on educational technology emphasize that personalization and immediate feedback are crucial for improving student self-regulation and engagement (McWilliams, 2023; Chang et al., 2023). Our findings reinforce this idea, as the high student response rate suggests that students found the information provided by the chatbot useful and relevant, which likely increased their motivation to resume their coursework with greater confidence.

The ultimate goal of such a system is to reduce academic dropout rates. The underlying premise, supported by the literature, is that if at-risk students—identified with high accuracy by a predictive model (Arévalo-Cordovilla & Peña, 2025a)—receive early and personalized support, they are more likely to overcome challenges that typically lead to attrition. Although this pilot study was not designed to measure longitudinal outcomes, such as dropout rates, the high student engagement observed (89.7% response rate) serves as a crucial validation of the intervention’s first step: successfully establishing a support channel with the target population. This initial success is a necessary precondition for achieving any downstream impact on student persistence and aligns with evidence suggesting that proactive support enhances academic performance and retention.

In summary, adopting systems such as ours in higher education settings could lead to substantial improvements in academic success, early identification and support for vulnerable students, more personalized learning experiences, and potentially lower failure and dropout rates in the long term.

Study limitations

Although the initial results are promising, this study has several limitations. First, the sample size and context were limited; the pilot study involved only 108 students from a single program over 4 weeks. This restricted scope does not guarantee that the effects will be reproduced in larger populations or across different disciplines, as institutional or cultural factors may influence the outcomes differently. Second, the short evaluation period meant that we could only measure short-term impacts and immediate operational metrics. We did not directly assess the effects on final grades, pass rates, or semester retention; therefore, it remains unknown whether the observed improvements would persist beyond the pilot study. Previous studies have noted that there is still limited evidence on how the prolonged use of educational chatbots influences deep learning and knowledge retention (Monrad Schei, Møgelvang & Ludvigsen, 2024), a question that our study could not address because of its exploratory nature.

Furthermore, the intervention’s design exclusively focused on students identified as at-risk. The potential benefits of using the system for broader pedagogical strategies, such as providing positive reinforcement or proactive engagement for all students were not assessed. While this targeted approach was intentional for this pilot, it represents a design limitation and a valuable avenue for future research to explore the chatbot’s role in fostering a positive learning environment for the entire cohort.

Another limitation is the varying acceptance of chatbots, depending on the type of student. While most students used it, some subgroups may have been more hesitant or exhibited different usage patterns than others. For example, those who are less familiar with AI may distrust automated recommendations, whereas students who are more digitally adept may perceive them as more useful and reliable (Shoufan, 2023). Our relatively homogeneous sample may not reflect the diversity of a full campus, which includes students of different ages and academic fields. Therefore, the results on response rates and satisfaction should be generalized with caution; the system may require adjustments for audiences with diverse demographic characteristics and digital skills.

Finally, scalability and sustainability were considered. While the small-scale pilot made technical support manageable, expanding the system to multiple courses or institutions would likely present several challenges, including integration with diverse learning management systems, handling simultaneous queries, and managing the computational costs associated with large-scale natural language processing. Furthermore, ensuring data privacy and regulatory compliance will become increasingly critical for broader deployments. Although the system’s design appears replicable, its effectiveness across various institutional and cultural contexts remains to be validated. Consequently, future research should explore logistical, technical, and adoption factors to determine whether the solution can maintain its efficacy beyond controlled environments.

A significant limitation of this pilot study was its focus on operational metrics without a corresponding qualitative or perception-based evaluation. This study did not capture students’ experiences, perceived value, or attitudes toward the chatbot intervention. Incorporating a theoretical framework, such as the Technology Acceptance Model (TAM), through surveys or interviews would have provided invaluable insights into whether students found the system helpful, intrusive, or motivating. This represents a crucial step for future research to complement the technical validation presented here.