AIEd Knowledge BaseAIEd KB
On this page

Research Article

Agentic Generative AI in Higher Education: Perceived Benefits, Risks, and Implications for Learning

added 2026-09-11 · updated 2026-09-19

Synthesis: The AGAI-HE framework — Ilieva et al. (2026) propose the Agentic GAI-Supported Learning Framework for Higher Education, which positions agentic AI as a bounded, human-supervised learning partner rather than a substitute for students or instructors. The framework separates three conditions — traditional e-learning, GAI-chatbot-supported learning, and GAI-agent-supported learning — and specifies agentic support as task contracting and interpretation, decomposition and planning, evidence organization, alternative generation, comparison and decision, feedback and refinement, and learner verification and reflection, all wrapped in human supervision, academic integrity, privacy safeguards, and instructor validation. An exploratory perception study with 130 students in an e-commerce course found both AI conditions rated above traditional e-learning on learning enhancement, personalization, decision-making support, and workflow organization — but no statistically significant difference between GAI agents and GAI chatbots, so the paper offers initial perception-based support for agentic GAI without claiming superior learning outcomes. (Preprint; not peer-reviewed.)

Overview

The premise is a distinction between generative AI as a productivity tool and as a learning-support technology. A chatbot can help a student produce a better answer without improving their reasoning; learning value appears when AI support makes the problem structure, evidence, alternatives, and justification more visible to the learner. Conventional prompt–response chatbots leave the learner responsible for sequencing the task, checking evidence, and judging when work is complete — a limitation that matters most in applied courses such as e-commerce, where students must compare business models, weigh market and operational constraints, and defend strategic recommendations.

Agentic here is deliberately qualified: it does not mean unrestricted autonomy. The framework applies bounded educational agency, in which goals, roles, data sources, tools, checkpoints, stopping conditions, and final decisions are defined or approved by human educators, and each agentic function must trace to a learning requirement, assessment purpose, or AI Governance control. The authors distinguish an agentic workflow from advanced prompt engineering: a well-designed prompt may request step-by-step output, but an agentic workflow preserves task state, allocates functions, checks completion conditions, returns to earlier stages when evidence is insufficient, and records material decisions.

The three conditions

Typical AI role Main learning support Key limitation or safeguard
Traditional e-learning none Instructor-designed content, activities, feedback, assessment Limited real-time personalization and feedback
GAI-chatbot-supported prompt–response assistant Explanation, summarization, ideation, drafting, feedback in isolated interactions Learner remains responsible for sequencing, verification, and completion judgment
GAI-agent-supported bounded workflow partner Task decomposition, evidence organization, comparison, verification, reflection, checkpoint-based progression Requires explicit boundaries, transparency, instructor oversight, and preservation of student agency

Architecture and governance

AGAI-HE is organized as three interacting layers:

  1. Pedagogical workflow layer — the instructional process itself: task interpretation and planning, instructional activity, formative assessment and feedback, verification and remediation, summative assessment, and course improvement.
  2. Agentic support layer — specialized educational agents for course design, content curation, tutoring, task decomposition, decision support, assessment support, learning analytics, verification, and reflection.
  3. Human supervision and governance layer — acceptable AI use, pedagogical boundaries, privacy rules, disclosure requirements, source verification, instructor checkpoints, integrity mechanisms, and final human accountability.

The framework was constructed through design-science phases (problem, objectives, artifact, demonstration, evaluation, communication) and is presented as an extension of the authors' earlier chatbot-assisted course framework, moving from episodic conversational support to supervised multi-step orchestration. Design requirements and a course-lifecycle allocation of instructor and student responsibilities are specified in accompanying tables.

Perception-based validation

The exploratory study (12 May – 9 June 2026) used a Bulgarian-language questionnaire with a within-respondent design: the same respondents rated all three conditions. After screening, 130 usable responses were retained (90 women, 40 men; 75.4% reporting medium digital skills). Prior chatbot exposure was near-universal (94.6% at least some use), while agent exposure was lower but substantial (85.4% at least some use, 61.5% using agents at least sometimes).

  • Both AI conditions beat traditional e-learning. Friedman repeated-measures tests found significant condition effects for learning enhancement (χ²(2) = 33.658, p < 0.001, Kendall's W = 0.129), personalization (56.410, p < 0.001, W = 0.217), decision-making support (18.007, p < 0.001, W = 0.069), and workflow support (6.766, p = 0.034, W = 0.026) — effect sizes very small to modest. Holm-adjusted Wilcoxon comparisons confirmed higher chatbot ratings than traditional e-learning on all four domains (e.g., personalization Δ = 0.677, r = 0.738), with the same pattern for agents.
  • The agent–chatbot gap was not significant. Chatbot and agent means were very close (e.g., learning enhancement 3.835 vs. 3.858; personalization 3.723 vs. 3.735), so the study does not demonstrate an agentic advantage.
  • Risk, trust, and adoption. Perceived risk was moderate (M = 3.333, SD = 0.893) while trust and adoption intention were more favorable (M = 3.615, SD = 0.805). 69.2% endorsed continued use of agents, 63.1% supported integrating them into learning activities, and 61.5% wanted training in effective agent use. The strongest single endorsement was for combining all three approaches (69.2%), whereas only 33.1% agreed agents were more effective than chatbots and 45.4% trusted agents under instructor guidance.
  • An unexpected association. Perceived risk correlated positively with continued-use intention (Spearman's ρ = 0.317, p < 0.001). The authors explicitly reject a causal reading and interpret it as informed adoption: more engaged or experienced users recognize both the value and the limits of GAI. The cross-sectional design cannot separate awareness, exposure, self-selection, and reciprocal influence.

What this means for practice

  • Instructors. Introduce agents where the workflow itself needs sequencing and checkpoints, not as a general upgrade: agent and chatbot conditions did not differ significantly on any of the four domains (learning enhancement 3.835 vs. 3.858).
  • Instructors. Keep goals, checkpoints and final decisions with learners and educators, and hand control back at the verification and reflection stages, so orchestration supports reasoning instead of completing the task.
  • Learners. Ask for explicit agent-use training rather than assuming access is enough — 61.5% of the 130 respondents wanted training in effective agent use, and only 45.4% trusted agents under instructor guidance.
  • Curriculum designers. Plan the course around hybrid support: combining traditional teaching with both chatbot and agent support drew the strongest endorsement (69.2%), while only 33.1% agreed agents were more effective than chatbots.
  • Administrators. Build disclosure requirements, source verification, privacy rules and instructor validation into the workflow layer itself rather than issuing them as separate policy, and treat students' risk awareness as part of informed adoption rather than resistance.

Limitations

  • The evidence is 130 usable questionnaire responses (90 women, 40 men) from one e-commerce course, using a within-respondent design in which the same students rated all three conditions.
  • No objective performance indicators were collected: the study measured perceptions and did not evaluate grades, submitted project quality, decision accuracy, task-completion time, knowledge retention or transfer. It used no random assignment, no baseline-to-post change and no longitudinal transfer measure, and the authors label the propositions the questionnaire could not test as requiring experimental, longitudinal or process-based evidence rather than as supported.
  • Condition effects were very small (Kendall's W = 0.026 to 0.217), and the key agent-versus-chatbot contrast was not significant, so the study cannot rank the two AI conditions; the authors read the null as possibly reflecting students' limited practical experience with agentic workflows and the novelty of those workflows, and suggest agentic value may become visible only under sustained, authentic implementation.
  • Only student perceptions were captured — no instructor perspective — the authors describe the implementation as an initial conceptual model rather than an optimized instructional system, and the paper is a non-peer-reviewed preprint.

Connected Concepts

Connected Articles

Citation

Ilieva, G., Yankova, T., Ruseva, M., Klisarova-Belcheva, S., Georgiev, P., & Totkov, G. (2026). Agentic Generative AI in Higher Education: Perceived Benefits, Risks, and Implications for Learning. Preprints.org (preprint, not peer-reviewed).

Metadata

Study designSurvey
DisciplineBusiness Education
LevelHigher Ed
Intended audienceLearners, Instructors, Administrators, Curriculum Designers, Researchers
Technologies and techniquesGenerative AI, Human-in-the-Loop, Pedagogical Agent, Personalized Learning
Research methods and evaluationDesign-Based Research
Institutions and policyAI Governance
Equity, ethics, and responsible useAI Use and Disclosure Statements, Trust
Foundations of AI in educationAcademic Integrity, Agentic AI
Page kindFramework
SourceHistoryEditChat

Embed this page

Copy the code below to embed a chromeless version of this page in a learning management system or other website. The embedded view hides the site header, navigation, and footer.