AICoP: CAiSEY an AI-powered Course Tool

The AI Community of Practice closed its spring season with a session that ran against the prevailing anxiety about AI in higher education. Dan Wang, the Lambert Family Professor of Social Enterprise at Columbia Business School, walked attendees through CAiSEY — a voice-based AI tool he founded that has students talk through their reasoning before class rather than outsource it. He was joined by co-founder Jill Cohen and curriculum and operations specialist Joanna Chouraqui.

The through-line wasn't the technology. It was a re-framing: the same tool many faculty fear as a shortcut can, built differently, force students to slow down.

A Problem the Case Method Couldn't Dodge

Dan teaches strategy through the case method, where the learning comes from students debating one another. It only works when students arrive prepared — and as he put it, students don't always speak up, either because they're hesitant or because they haven't done the reading. When ChatGPT arrived in late 2022, that second problem sharpened into a crisis. It became trivial to skim a summary instead of the case, or to walk in repeating an argument the model produced rather than one the student reasoned through.

CAiSEY is Dan's answer. Rather than another text chatbot, it drops a student into a voice-first, real-time debate with what the team calls a "knowledgeable peer" — adversarial enough to push, friendly enough to keep going. In a live demo, Parixit played a CBS student prepping a case on Netflix's content strategy, arguing for original content while CAiSEY pressed him on cost control and global scale. What struck him afterward was that the format made him slow down and think the question through — the opposite, he noted, of how AI usually gets used.

Evidence, Not Just Enthusiasm

Dan is a social scientist by training, and he was candid about not trusting satisfaction surveys. A spring 2025 pilot with 275 MBA students produced roughly 1,300 conversations; he reported that about 90% of the free-response feedback was positive, with most students citing sharper critical thinking and a stronger sense of preparation. But he wanted causal evidence.

So the next semester, the team ran a randomized controlled trial with 760 students in the required Strategy Formulation course, assigning each to use CAiSEY in two of twelve sessions at random. The result Dan highlighted: students who used the tool twice before the midpoint were 18% more likely to report feeling more comfortable participating in class. He stressed the size of the input — about 20 minutes of total engagement producing a measurable confidence shift. He described it as a nudge that punches above its weight.

Students also pointed to inclusion. Many at Columbia speak English as a second language or have learning differences like dyslexia and dysgraphia, and several described CAiSEY as a lower-stakes way to practice. Because it's plainly an AI and clearly just practice, Dan said, students tried out arguments they might have self-censored in front of a professor — and often redid conversations to experiment.

Under the Hood

CAiSEY runs on a voice-native model — OpenAI's Realtime API — rather than the usual speech-to-text-and-back pipeline, with an architecture layer that lets the team swap in other voice-native models. Dan pointed to two payoffs: near-zero response lag, and the model's ability to register tone and emphasis that text strips away, which is what makes the exchange feel like a conversation.

Around that sits a system prompt that defines CAiSEY's peer persona and keeps it on task, shaped by decades of teaching notes and research on productive disagreement — Dan singled out Harvard Kennedy School professor Julia Minson's work on having difficult conversations. Cohen added that the knowledgeable-peer framing, pitched at the grad-student level, is part of why it resonates: a real peer in a real meeting sometimes misremembers a fact or repeats a point, and that realism is part of the practice. The tool has also been red-teamed as part of security review.

Crucially, instructors stay in the driver's seat. They provide the discussion topic, supply teaching materials as CAiSEY's knowledge base, and can add custom instructions — a common one being to hold the discussion as if it were taking place at the time of the case.

From One Class to Forty

What began in Dan's classroom has spread to more than 4,000 students across 40 courses at 23 institutions, reaching well beyond business into leadership, ethics, accounting, creative writing, and pilots in engineering and medical education. Johanna described the common thread as any setting where a student is reasoning and thinking critically out loud, citing work with a writing composition course, a veterinary course, and epidemiology at the medical campus.

For instructors, Dan summarized one professor's experience at HEC Paris: saved prep time, better discussion from better-prepared students, a clearer window into student learning, and improved teaching ratings. He was careful to say he couldn't promise the last one for everyone.

On privacy — a recurring AICoP theme — Dan was unequivocal that adopting institutions own their data and that CAiSEY does not train on student submissions or instructor materials. For Columbia community members, the tool is offered essentially at cost through a chart string, with the team absorbing API costs.

Takeaway

CAiSEY is a working counterexample to the idea that classroom AI mainly erodes effort. By making students talk their way through a problem before they walk in, it adds friction where most tools remove it. With early experimental evidence, fast cross-disciplinary uptake, and a firm stance on data ownership, it points toward a version of AI in education designed to deepen learning rather than route around it. As Dan put it, the goal is a tool that helps every student find their voice.

Anyone can try the Netflix demo at caisey.me, and instructors can request access to build their own assignment.