Duke's Fuqua School of Business to roll out OpenAI-powered classroom analysis tool this fall

The system analyzes recorded class discussion and team meetings to give students and faculty feedback on participation, teamwork, and where understanding breaks down.

Duke University's Fuqua School of Business is rolling out an AI classroom analysis system built with OpenAI's API this fall.

Duke University's Fuqua School of Business is expanding an AI-enabled classroom system to every classroom in the school this fall, after raising funding to upfit the facilities. The tool, developed with OpenAI's API, records classroom discussion and student team meetings, then analyzes who contributed, how conversations developed, and where understanding broke down.

The system was developed by Scott Dyreng, an accounting professor and Senior Associate Dean of Innovation at Fuqua, working with colleagues including Robert Olinger, an assistant dean who had been building an AI system to analyze recorded discussion. It began in one classroom, expanded to two, and will now be available to any faculty member who chooses to use it. The work is detailed in a case study published by OpenAI on July 6, 2026.

The first major use case is C-Lead, Fuqua's first-year team orientation program, in which students are assigned to teams, create a team charter, and work together through required courses. The school is working through the logistics of applying the tool from the start of the program so teams can track whether they follow through on their commitments.

"If we could figure out who was speaking, capture a transcript, and analyze that transcript, we could learn a lot about what was happening in class," Dyreng says in the case study.

The tool addresses a long-standing problem in business education: participation can form a meaningful part of a student's grade, but is typically assessed from a professor's memory across multiple sections, sometimes with 70 students in a room.

How the system works

Ceiling-based microphone arrays capture classroom audio and determine where sound is coming from. Students check into a seat, rather than the room, through a modified attendance app, which connects speakers to the transcript without assigned seating. The annotated transcript is then analyzed through OpenAI's API using context from the faculty lecture, assessing whether a comment responded to a question, built on another T

The system runs on OpenAI's GPT-5 model family, using GPT-5.4 as the default analysis model and GPT-5.5 selectively. Participation assessment runs as an ensemble of model calls that vote, and a single lecture can consume hundreds of individual calls. The team began on GPT-5.2.

Dyreng says the workflow includes checks against treating a single model judgment as fact. "You tag the data, analyze it in one context, analyze it again in another context, and sometimes send it through different versions of a prompt to see whether the model agrees before sending feedback to students," he says. Transcripts can be anonymized before analysis and mapped back afterward.

Team meetings recorded after AI use weakened collaboration

The team meeting component grew out of a problem Dyreng observed in fall 2025, when he gave students unrestricted access to AI for assignments. Teams of around five students divided work up individually rather than meeting, he says in the case study, because AI could polish final submissions without peer review.

"We were trying to help students learn how to work with others, and introducing access to AI enabled them to reduce their interactions with others," he says.

His response was to require teams to upload audio recordings of their meetings, initially just to confirm they met. The recordings are now transcribed, speaker-labeled, and analyzed. In one team he cites, one male student and four female students worked together, and the male student spoke for around 50 percent of the time.

Each team receives around three pages of feedback per meeting, covering speaking time, the questions discussed, conclusions reached, and how members interacted, along with suggestions such as directing questions to quieter members.

Faculty reports before class

Faculty receive pre-class reports drawn from team meeting analysis. Dyreng says one report based on 15 hours of meetings across 15 teams described which concepts students understood, which they were struggling with, and which students or teams to call on in class.

The system also analyzes teaching across a full course. Dyreng says it showed students in his class used the vocabulary of a framework he teaches without understanding it deeply. "I've been teaching for nearly 20 years, and I haven't had access to this kind of insight before," he says.

According to the case study, Fuqua is also considering the workflow for admissions interviews and recorded mock interviews in career services. The school-wide rollout begins this fall for faculty who opt in, starting with first-year team formation and leadership work.