The University of Michigan Gave 9,000 Students an AI Teaching Assistant. Here’s What Happened.
In April 2025, the University of Michigan’s Ross School of Business launched one of the largest deployments of an AI agent in higher education. Built on Google’s Gemini, their Virtual Teaching Assistant works across 72 courses and 26 schools. It guided them through problems, adapts to specific curricula, and reports back to faculty with real-time analytics on where students struggle.
Early results showed higher engagement. Instructors got deeper insight into learning patterns. And the study, co-led with Northwestern University’s Kellogg School of Management, is still expanding.
This wasn’t a flashy demo, but a quiet infrastructure decision, which signals a larger shift: now, AI agents in education are operational systems handling enrollment, student support, research workflows, and administrative load at scale.
The global AI in education market reflects that shift. Estimated at $5.88 billion in 2024, it is projected to reach $32.27 billion by 2030, growing at a CAGR of 31.2%. That growth is driven by institutions that treat AI as an operational investment, not a teaching experiment.
What Exactly Is an AI Agent in Education?
An AI agent in education is a software system that can perceive a task, make decisions, and execute actions with minimal human supervision. It goes beyond a chatbot that answers FAQs.
A chatbot waits for a question, then responds from a fixed script. An AI agent can authenticate a student, check their enrollment status against the university ERP system, verify prerequisites, flag a scheduling conflict, and suggest an alternative section. All in one interaction.
This distinction matters for decision-makers in educational institutions because the operational value of AI agents sits in workflow execution, not in conversation.
According to the EDUCAUSE 2025 AI Landscape Study, 57% of higher education leaders now consider AI a strategic priority. But only 39% have AI-related acceptable use policies in place. That gap — between intent and infrastructure — is where AI agents create measurable impact.
Where the Operational Gaps Are
Before looking at solutions, it helps to see the concrete, observable problems educational institutions face today.
Response time pressure. Students expect immediate answers. Support teams at mid-size universities handle thousands of inquiries per enrollment cycle about financial aid deadlines, course registration, transcript requests, and housing. Most of these are repetitive and predictable. Staff members answer the same 50 questions in slightly different forms hundreds of times per semester.
Administrative workload on faculty. A 2025 study by SaM Solutions reports that 87% of educational institutions globally have integrated AI tools into at least one area of operations, from tutoring to grading. Yet much of that integration is shallow. Faculty still spend hours per week on attendance tracking, grading for completion, progress reports, and responding to routine syllabus questions.
Enrollment and retention friction. Getting a prospective student from first inquiry to enrollment is a multi-step process involving admissions, financial aid, academic advising, housing, and registration. Drop-off happens at every stage. Many institutions lose applicants not because of program quality, but because the process is slow or confusing.
Research inefficiency. Literature reviews, data extraction, citation management, and report writing consume weeks of researcher time. These are structured, repeatable tasks that follow clear patterns.
Disconnected systems. Most universities run separate platforms for their Student Information System (SIS), Learning Management System (LMS), financial aid, HR, and communications. Data lives in silos. Staff manually transfer information between systems. Errors compound.
Six Ways Institutions Are Using AI Agents Right Now
1. Agentic Virtual Teaching Assistants
The University of Michigan’s Virtual TA is the most documented example. Powered by Google Gemini, it operates across courses in financial technology, operations strategy, analytics, and statistics. The agent provides 24/7 access to course-specific support, explains complex concepts on demand, and acts as a practice partner. Critically, it is designed to never give away answers — it guides students through reasoning instead.
For faculty, it delivers analytics dashboards showing commonly asked questions, engagement patterns, and areas where students consistently get stuck. This allows instructors to adjust their teaching between sessions rather than discovering problems only at exam time.
The EAB (Education Advisory Board) describes how this model differs from previous chatbot tools: the agent walks students through material, coaching them rather than just answering.
Operational result: Faculty spend less time on repetitive Q&A. Students get support outside office hours. Institutions collect structured data on learning patterns at scale.
2. Student Enrollment and Support Agents
Georgia Southern University deployed an AI agent called GUS, built with DRUID AI. The university was struggling with an outdated SMS-based communication system that required significant manual management by senior staff.
GUS handles natural language inquiries across the entire university lifecycle: enrollment questions, HR processes, campus services, financial aid, and academic advising. It authenticates students, connects to the university’s ERP and Student Information System, and provides personalized responses.
Results reported by the university: over 35,000 conversations handled, measurable improvements in student engagement, and a 2% enrollment increase generating $2.4 million in additional revenue. The university now plans to expand GUS’s capabilities to proactively address student needs rather than just responding to inquiries.
Columbus State University followed a similar path, deploying a Student Information Agent connected to its Banner system and SSO infrastructure for real-time, personalized student support.
Operational result: Reduced manual workload on administrative staff. Faster response times for students. Direct revenue impact through improved enrollment conversion.
3. AI-Powered Research Acceleration
Researchers at AMD and Johns Hopkins University developed Agent Laboratory, an AI framework that automates literature reviews, experimentation, and report writing. The system uses specialized AI agents working in sequence: one gathers and analyzes research papers, another plans experiments and prepares datasets, and a third runs experiments and generates documentation.
The measured result: 84% reduction in research costs compared to existing autonomous methods, at approximately $2.33 per paper versus $15 with previous approaches. Research quality standards were maintained.
Johns Hopkins has since established a dedicated AI Agent Lab at the Carey Business School, focused on designing AI agent workflows for business and healthcare challenges.
Stanford University’s Virtual Lab takes a different approach, using a large language model to simulate an entire research team. An LLM principal investigator agent orchestrates specialized agents with different scientific backgrounds to examine interdisciplinary research questions.
Operational result: Research teams complete literature reviews and preliminary analysis in hours rather than weeks. Cost per research output drops significantly. Faculty time shifts from data gathering to interpretation and decision-making.
4. Campus-Wide AI Innovation Platforms
Arizona State University has moved further than most institutions in treating AI as operational infrastructure. In 2024, ASU became the first university to collaborate with OpenAI. Since then, it has launched the AI Innovation Challenge, which has resulted in over 700 faculty and staff projects. ASU also built CreateAI Builder, an internal platform that enables the campus community to build AI-enabled products in a secure environment.
The university’s approach is notable for treating students as AI builders, not just users. Through the AI Acceleration Student Innovation Challenge, student teams design and prototype AI-powered solutions for campus operations, mental health support, and learning accessibility. ASU’s student workers also built an AI-powered chatbot for prospective student engagement as part of a collaboration with Amazon Web Services.
The Arizona AI Challenge, launched in late 2025, convened over 90 students from ASU, Pima Community College, and South Mountain Community College to build AI tools for neurodivergent learners. One student team created Nerva, an AI assistant that breaks tasks into manageable chunks and schedules them in a calendar with supportive feedback.
Operational result: AI capability spreads across departments through internal tools. Student-built solutions address real operational needs. The institution builds AI infrastructure, not just AI pilots.
5. Automated Administrative Workflows
The most immediate operational value of AI agents in education is in tasks no one wants to do, but everyone depends on. These include:
- Grading for completion — AI agents evaluate whether assignments meet basic submission criteria, freeing faculty for qualitative feedback
- Scheduling and registration — Agents check prerequisites, flag conflicts, and suggest alternatives, reducing registration errors
- Financial aid processing — Agents guide students through applications, verify deadlines, and answer eligibility questions
- Survey and feedback collection — Automated outreach with adaptive follow-up based on student responses
According to Workday’s analysis, 86% of students say they are already using AI in their studies and expect their institutions to do the same. But 80% say institutions are not fully meeting those expectations. The gap is primarily in operational integration, not in teaching.
Ithaca College’s approach illustrates a measured path to adoption. They established a Presidential Working Group on AI, launched faculty AI Mini-Grants ($500 stipends to develop AI-integrated course units), opened an AI Exploration Lab staffed by student guides, and are developing an agentic AI system called Aurora for proactive student guidance.
Operational result: Hours of repetitive administrative work are reclaimed per week per staff member. Error rates in registration and scheduling decrease. Students receive faster, more consistent service.
6. Custom AI Agents Built for Institutional Workflows
Off-the-shelf AI tools handle generic tasks well. But educational institutions have unique combinations of legacy systems, regulatory requirements, enrollment processes, and reporting structures.
This is where custom AI agent development matters. A custom agent can be trained on an institution’s specific course catalog, financial aid rules, HR policies, and communication standards. It integrates with the existing technology stack — Banner, Canvas, Workday, Salesforce — rather than requiring everything to be rebuilt.
At Lab51, we build custom AI agents for businesses and institutions that need AI systems designed around their operational reality. For educational organizations, this means agents that connect to existing SIS and LMS platforms, respect data privacy regulations, and execute multi-step workflows like enrollment processing, student success interventions, or cross-departmental reporting.
The difference between a generic chatbot and a purpose-built AI agent is the difference between answering a question and completing a task.
The Numbers Behind the Shift
| Metric | Data Point | Source |
| Global AI in education market (2024) | $5.88 billion | Grand View Research |
| Projected market size (2030) | $32.27 billion | Grand View Research |
| CAGR 2025–2030 | 31.2% | Grand View Research |
| Institutions with AI as strategic priority | 57% | EDUCAUSE 2025 |
| Institutions with AI acceptable use policies | 39% | EDUCAUSE 2025 |
| Students already using AI in studies | 86% | Workday |
| Research cost reduction (Agent Laboratory) | 84% | InfoQ / Johns Hopkins |
| Enrollment revenue lift (Georgia Southern) | $2.4M from 2% increase | DRUID AI |
| AI faculty/staff projects at ASU | 700+ | ASU News |
Why Educational Institutions Need to Act Now
Three things are converging. First, student expectations are shaped by consumer technology. They expect instant, personalized, 24/7 service. Staff teams cannot deliver that at current headcount levels without AI support.
Second, the institutions that build AI infrastructure now will have a compounding advantage. The University of Michigan’s data on student learning patterns gets more valuable with each semester. ASU’s 700+ internal AI projects create institutional knowledge that competitors cannot replicate overnight.
Third, regulation is catching up. The EU AI Act classifies education as a high-risk domain. Institutions that wait will face both competitive pressure and compliance requirements simultaneously. Starting now means building on your terms, not someone else’s timeline.
The practical first step is not to buy a platform. It is to identify three to five operational workflows where staff spend the most time on repetitive, rules-based tasks. Financial aid inquiries. Registration support. Post-admission follow-up. Progress reporting. These are the processes where AI agents deliver the fastest, most measurable return.
If your institution needs AI agents built around your specific systems and workflows, Lab51 works with organizations to design, build, and deploy custom AI agents that integrate with your existing infrastructure.
