Universities Are Testing AI Agents Here’s What’s Actually Working

A deep dive into a GovTech article exploring AI agents in education, where universities are seeing early success in administration but still struggling with teaching and learning applications.

Transcript

SPEAKER_00 0:00

If this episode makes you think, please let us know in the comments and support us by subscribing and leaving a review. Thank you. Today we are exploring a GovTech article by Abby Sauerwine titled AI Agents in Education, What's Working and What's Missing, published in May 2026. Now this article is interesting because it moves the conversation beyond chatbots and content generation into something more advanced. It focuses on agentic AI, which essentially refers to AI systems that can carry out sequences of tasks independently. Rather than simply responding to prompts, these systems can make decisions, interact with software tools, monitor outcomes, and adjust their actions in pursuit of a goal. And according to the article, universities are already beginning to experiment with these systems in quite practical ways. But what becomes clear very quickly is that the enthusiasm is not evenly distributed across education. Administrative uses are progressing rapidly, while teaching and learning applications remain much more contested. The article explains that Agentic AI only recently entered mainstream education conversations around 2023 and 2024. Nate Ober from Amazon Web Services describes these systems as being able to use tools like databases or learning management systems to plan and execute workflows in a loop until a task is completed. Earlier versions struggled with anything beyond a few simple steps, but the technology has advanced quickly enough that some systems can now operate independently for extended periods. Now that sounds impressive, but one of the recurring themes in the article is caution. Nicole Engelbert from Oracle warns that we are still in the very early stages and suggests decision makers should be skeptical of claims that this technology is already widespread or fully mature. That tension between excitement and caution runs through the entire piece. The clearest success stories so far are on the administrative side of education. According to the article, these are seen as easy wins because they involve repetitive, high-volume processes that institutions already struggle to manage efficiently. One example comes from the Illinois Institute of Technology, which automated transcript processing. Tasks that previously took around a month were reportedly reduced to a single day through AI-driven workflows handling intake, international grade conversion, and integration with customer management systems. Another example comes from a Highline College in Washington State, where an AI-powered financial aid status tracker significantly reduced student inquiries. The article reports a 75% drop in emails, phone calls, and in-person visits related to application status. Now these examples are important because they reveal where AI currently fits most comfortably in education. These are not deeply human, relational, or pedagogical tasks. They are operational bottlenecks. And when AI removes friction from those systems, the value is relatively easy to measure. The article also describes how these systems are becoming integrated into learn and management platforms like Canvas. In one example, an instructor can issue a natural language request such as grant the student an extension, and the AI agent can then carry out multiple connected actions automatically. It updates the due date, schedules reminders for the student, and prompts the instructor later about grading arrangements. Now this might sound small, but it points to something larger. AI agents are not just responding to requests, they are coordinating workflows across systems, and that starts to shift the nature of administrative work itself. The article identifies academic advising as another particularly promising area. Agentic AI systems can model different course pathways, optimize schedules, and prepare recommendations before routing them to human advisors for review. So again, the AI is not necessarily replacing human decision making, but reshaping the preparation and organization around it. However, once the article moves from administration into teaching and learning, the tone changes noticeably. This is where resistance begins to emerge, and the reason is fascinating. Administrative inefficiency is generally seen as unnecessary friction. Nobody believes students benefit from waiting a month for transcript processing, but educational struggle is different. In teaching and learning, some forms of difficulty are actually valuable. The article refers to this idea as productive struggle. The process of wrestling with concepts, making mistakes, revising thinking and gradually developing understanding is often central to learning itself. Jake Burley from the University of Massachusetts argues that education is fundamentally different from other institutional tasks because there is something deeply personal about the learning experience, and that creates a challenge for AI systems designed around efficiency and task completion. The article gives a particularly striking example through a tool called Einstein, developed by the Startup Companion. This AI agent integrates directly into Canvas and can automatically complete assignments for students. Now just think about what that means for a moment. If the system can complete the assignment itself, then the educational process risks collapsing into output generation. The assignment still exists. The grade may still exist. But the learning experience in the middle starts to disappear. And this is where the article becomes much more philosophical. Because once AI can reliably complete academic tasks, educators are forced to confront a bigger question. Why do students complete assignments in the first place? What is the purpose of the activity? Is it to demonstrate knowledge, practice thinking, develop a resilience, or simply produce a correct answer? The article suggests that AI exposes weaknesses in educational design. If students are easily tempted to outsource tasks entirely, perhaps they do not fully understand the value of those tasks themselves. Now some institutions are trying to navigate this tension creatively. The article highlights instructors building custom GPT tutors trained on their own course materials, allowing AI to support students within carefully designed boundaries. The University of Luxembourg is also experimenting with AI agents throughout the instructional cycle, including lecture preparation, real-time translation, transcription, and post-lecture analysis. So rather than replacing learning, these systems are being positioned as infrastructure around learning. Still, the article repeatedly returns to the issue of reliability. Ober warns that as workflows become more complex, the margin for error compounds at every stage. And this introduces an important principle that I think educators should pay attention to. The acceptable level of AI error depends entirely on the stakes of the task. A slightly imperfect course recommendation might be manageable. An incorrect financial aid calculation is far more serious, and that principle applies in classrooms as well. There is a big difference between AI supporting brainstorming and AI evaluating understanding or making decisions about student progression. The article also raises concerns around privacy and regulation. Some AI vendors entering education may not fully understand frameworks like FERPA or COPPA, which creates additional risks when systems are deeply integrated into student data and institutional workflows. So alongside the excitement around capability, there is a parallel conversation emerging around governance, accountability, and trust. Now if you step back from the article as a whole, I think the most interesting insight is this. Education appears much more willing to automate administration than learning itself. And that distinction matters, because it suggests that even in an AI-rich future, institutions still see teaching and learning as fundamentally human processes, or at least processes where human judgment remains central. AI agents may handle scheduling, reminders, workflow management, and repetitive support tasks extremely well. But once they move into spaces involving intellectual struggle, meaning making and personal development, the conversation becomes much more complicated. And perhaps that is the real takeaway from this piece. The future of AI in education may not be defined by whether machines can teach. It may be defined by how carefully institutions decide which parts of education should never become fully automated in the first place. That's all for today. Thanks for listening, and I'll see you in the next episode.