Case Study
Faculty-Led AI at George Washington University
How the School of Engineering & Applied Science put course-grounded AI agents in the hands of instructors — student-centered by design, faculty-led in practice, and priced by usage instead of per student.
Professor Lorena A. Barba introduced a faculty-configured AI agent at GWU’s School of Engineering & Applied Science — one where the instructor sets the persona, supplies the course materials, and moderates how students interact with it. The point was never AI access. Students already had that. The point was AI grounded in the actual course.
Faculty-led
Instructors configure the agent
Course-grounded
Answers cite course materials
Any LLM
OpenAI, Gemini, Llama, Anthropic
Usage-based
No per-student license
Student-centered, faculty-led
Most campus AI programs start by buying seats and hoping for adoption. GWU started from the course: what should the AI know, how should it behave, and who decides? The answer to the last question is the instructor.
“What sets our approach apart is its focus on pedagogical design and cost-effectiveness. Unlike enterprise-wide solutions, the pay-as-you-go model could save institutions significant resources while offering greater customization.”
“Most are using ChatGPT, and not very well. Our AI agent addresses this reality by providing a customizable, course-specific tool that grounds AI responses in course materials.”
The instructor is the administrator
In the faculty demo, Professor Barba added course resources, set the agent’s behavior through prompts, and moderated student interactions — then showed the agent answering student questions with course-specific material and pushing back toward critical thinking rather than handing over answers.
Instructor-controlled persona
The faculty member defines how the agent behaves, what it will and won't do, and the tone it takes with students.
Course resources as the source
Readings, notes, and problem sets are loaded in, so the agent answers from the course rather than from the open internet.
Context-aware responses
Retrieval-augmented generation keeps answers tied to the supplied materials, reducing invented or off-syllabus responses.
Model selection, not lock-in
OpenAI, Google Gemini, Llama, Anthropic — the model is a setting, chosen per course and swappable as the field moves.
Visible usage and cost
Consumption is tracked through the AI vendor dashboard, so a department can see exactly what a course actually costs to run.
LTI embedding in the LMS
The agent drops into the existing course shell over LTI — students meet it where they already work, with no separate login.
Per-student licensing is the wrong shape for a campus
A per-seat license bills for every enrolled student whether or not they open the tool. Usage-based pricing bills for what a course actually consumes. At a university-sized population the difference is not a discount — it is a different cost curve.
| Model | How it bills | List price |
|---|---|---|
| ChatGPT | Per student, per month | $20 / student / mo |
| Microsoft Copilot | Per student, per year | $80 / student / yr |
| ibl.ai at GWU | Pay-as-you-go, plus bulk rates | Metered usage |
Published list prices at the time of the GWU program. On those figures, ibl.ai’s usage-based model came in roughly 85% below ChatGPT and 75% below Microsoft Copilot for the same population.
Enterprise-wide seat license
- Billed for every enrolled student, used or not
- One vendor's model for every course
- Generic assistant, no course grounding
- Cost rises with enrollment, not with value
ibl.ai at GWU
- Pay for the tokens a course actually consumes
- Model chosen per course, swappable anytime
- Answers grounded in that course's materials
- Department-level visibility into real cost
A program with a research design attached
At GWU’s TRAILS conference in February 2025, GWU and ibl.ai presented “Student-Centered Approach to AI in the Classroom” — a faculty-tailored program studied the way an academic program should be, with instruments and published findings rather than vendor anecdotes.
AI as tutor, coach, and teammate
Students used the agent for understanding material, for creative work, and as support inside coursework — three distinct modes, not one chatbot use case.
Real-world data collection
Surveys, focus groups, and faculty interviews tracked what the agents actually changed — for students and for the instructors configuring them.
Motivation and engagement as the outcome
The research looked past usage counts to what matters academically: whether the agents affected student drive and performance.
Best practices published for faculty
Guides and findings were written up so other instructors could adopt the approach — and shared openly to build AI literacy beyond one campus.
Professor Lorena A. Barba
School of Engineering & Applied Science, George Washington University
“ibl.ai’s generative AI in education is set apart by their unparalleled expertise and their commitment to empowering educators by putting full control in their hands.
Their platform enables instructors to customize AI agents, ensuring that the AI responses are always grounded in course materials—minimizing the risk of inaccuracies or hallucinations.
For any educational institution looking to lead the future of AI-driven learning, ibl.ai is the gold standard.”
One course, then a department, then a campus
The GWU program is deliberately reproducible. Nothing in it depends on a campus-wide procurement or a single approved model — an instructor can stand up a course-grounded agent, measure it, and hand the recipe to the next department.
Starts at the course
A department can begin on its own terms, without waiting on an institution-wide procurement cycle.
Configured, not coded
Faculty add materials and write prompts. Nothing about the setup requires an engineer on standby.
Scales on usage
Adding a department adds its consumption, not a new per-student line item across the whole population.
Run the same program at your institution
Course-grounded agents your faculty configure, on any model, priced by what you actually use — and deployed wherever your data has to live.