Why my AI teaching assistant is far superior to Claude, Gemini, or ChatGPT

By Arnaud De Bruyn, Professor of Marketing and Associate Dean for the Ph.D. Program at ESSEC Business School

Three days before the final term-project presentations, I had not received a single question from my students. I checked. Over the next two weeks, still nothing. Not one email, not one cry for help. I asked in class if everything was OK. Silence. I started to worry that the term project had quietly collapsed.

It had not. When the groups presented, the work was the most polished, professional, and conceptually sound I had ever seen in this course.

The students had been talking to an AI assistant I had built for them, one that helps them refine their term-project pitch. Yet it would be a mistake to conclude that any chatbot would have produced the same outcome. Drop Claude, Gemini, or ChatGPT in front of the same students, and the result would have been considerably weaker, possibly worse than no help at all. The reason is that the AI assistant did not bring the expertise; I did. The model only reflected, with infinite patience, the context I had given it. And that context was a decade of accumulated feedback, frustrations, dead ends, and recurring patterns I had seen students fall into.

This is, I think, where most AI-in-education projects get the framing backward. The implicit hope is that the generative model arrives with the substance, and the professor merely deploys it. That hope is misplaced. A general-purpose assistant has read everything and learned nothing about how my students stumble on this assignment. It will offer cheerful, generic, often confidently wrong guidance. It will let a group pursue a project that I could have told them in thirty seconds would yield nothing worth pursuing.

Term projects, in my data-driven marketing course, are an extraordinary learning experience: hands-on, practical, useful, and impressive enough on a CV that students still email me about them years later. They are also one of the most time-consuming exercises a teacher can run. Each pitch deserves a real reaction; each ambiguity deserves a push-back; each shaky methodological choice deserves the right counter-question. Multiply that by twenty groups, and the math becomes brutal. I almost abandoned them out of exhaustion.

What the AI assistant changed is not the depth of the feedback, it is the throughput. The instructions I gave it are, in essence, the same things I would say to a student in office hours, written down once and refined over several iterations: what counts as a real tension, why a single methodology beats three half-baked ones, when a topic is too narrow to find respondents, or when it is too broad to be interesting. The students get those reactions immediately, at any hour, on any draft, until they no longer need them. They can come back twenty times if they want to —and some did— whereas in previous years, more than a few groups were reluctant to send me a single email.

What makes this work is less the model itself than the scaffolding around it. The assistant is explicitly instructed never to write the project for the students, never to draft their survey questions, and never to declare a methodology choice on their behalf. Its job is to make them think harder, not to think for them. When a pitch is vague, it asks where the tension is. When a pitch promises three methodologies at once, it pushes back. When a topic sounds unfeasible (e.g., too niche to find respondents, too broad to deliver in a few weeks), it says so, plainly.

Behind that behavior sits a structured brief: the four methodologies we cover in the course, a catalog of typical use cases for each (segmentation, conjoint, positioning, and pricing), and a taxonomy of the recurring traps I have seen groups fall into over the years: methodology creep, missing tension, no access to respondents, excess or lack of ambition, and the eternal temptation to study the cafeteria. The model does not improvise this material. It draws on it.

I do not think this scales blindly to every course or every assignment. The assistant works here precisely because the context behind it is dense, specific, and earned.

Building an AI assistant is uncannily close to training a new employee. The managers who train the best juniors are almost always the ones who have done the job themselves, at the bottom of the ladder, long enough to know where the work hurts. They can anticipate the mistakes. A manager who has never been a foot soldier writes generic training material and hopes for the best. The same is true of building AI assistants: you cannot delegate to a model an expertise you have not earned, slowly, the hard way. The lesson I draw from this experiment is not that AI replaces teaching, nor that it amplifies it for free. It is that an AI assistant is only as good as the years of work you have put into being able to instruct it.