As higher education institutions grapple with the ethics of using Generative AI in classrooms and examination halls – and some even deploy unreliable detection tools to catch students cheating – the Indian Institute of Management Bangalore has done something unique. It has allowed students ‘controlled’ use of AI in exams, with a twist.
During the June-July term this year, 78 MBA students across core Business Analytics and general MBA elective cohorts took their GenAI midterm not by putting pen to paper, but by training and deploying their own chatbots to answer exam questions on their behalf.
The initiative was spearheaded by Prof Subhabrata Majumdar, who built a custom, password-protected testing platform on Vercel – a cloud platform for web applications – in just two weeks using AI-assisted coding tools. To ensure academic integrity and data privacy, the platform operates as a secure cloud environment. The underlying Large Language Model (LLM) is strictly restricted to assigned course materials, preventing internet access or external data leakage.
According to Prof Majumdar, students were given a month to experiment, build draft versions, and evaluate different iterations of their chatbots before submitting one final bot for grading. The exam questions were identical to what a traditional midterm would test on Generative AI concepts, but supplied directly through the chatbot they built.
Because the platform was architected as a no-code Retrieval-Augmented Generation (RAG) workflow, non-technical students were not disadvantaged compared to those with coding backgrounds. Instead, grading evaluated prompt engineering, parameter tuning, and how effectively students operationalised course material.
“Chatbots that scored highly had prompts with specific and detailed instructions tailored towards answering questions realistically, keeping outputs brief, or augmenting low-quality answers with the model’s internal knowledge,” said Prof Majumdar.
The format also tested practical decision-making under technical constraints. When students discovered the platform struggled with complex PDF formatting, high-performing cohorts adapted by converting their course material into plain text to optimise accuracy. “It trains them for industry-relevant soft skills, such as decision-making under uncertainty,” added Prof Majumdar.
The approach reflects a deliberate departure from static, memory-based assessments towards practical application. Drawing from eight years of industry leadership prior to academia, Prof Majumdar noted that real-world execution hinges on problem-solving rather than rote recall. “Future business leaders with hands-on knowledge will be able to quickly relate to technical teams, grasp deployment limitations, and accelerate project timelines. A hands-on, no-code evaluation is far more effective for B-school students than traditional closed-book exams.”
Initial student feedback for the format was overwhelmingly positive, yielding course evaluation scores between 4.5 and 5 out of 5. Following the success of the midterm, Prof Majumdar plans to open-source the Vercel-based platform and collaborate with faculty colleagues to scale the testing framework across other courses at IIM Bangalore.