Ajay Patel and a group of researchers from Wharton, Carnegie Mellon University and Harvard Business School have introduced a benchmark showing that frontier AI models can produce strong answers across open-ended business disciplines, extending measured progress beyond mathematics, coding and other tasks with easily checked answers.
The researchers' BusinessCaseBench working paper, first submitted on July 17th and revised on July 21st, uses 238 business school cases and their instructor solutions to create 615 questions across 18 disciplines. The questions cover strategy, finance, leadership, business ethics, marketing and business-government relations, among other fields.
Patel, the paper's corresponding author, conducted its methodology, experiments and analysis. His co-authors are Wharton professor Kartik Hosanagar, Carnegie Mellon professor Ramayya Krishnan, University of Pennsylvania professor Chris Callison-Burch and Harvard Business School professor Karim Lakhani.
The findings drew attention on July 28th when Ethan Mollick (@emollick), a Wharton associate professor and co-director of the school's Generative AI Labs, argued in a thread on X that the data contradicts claims that model gains are concentrated in verifiable fields. Mollick wrote that large language models remain "unreasonably effective across so many different disciplines" despite uneven performance within and between fields.
Mollick's interest is grounded in both management research and company-building. According to his Wharton profile, he co-founded a startup before entering academia and earned an MBA and PhD from MIT Sloan. His current research focuses on how AI changes work, entrepreneurship and education.
A benchmark for ambiguity
BusinessCaseBench asks models to handle the kind of incomplete, contested information found in management decisions. Representative tasks include advising a hospital on a difficult promotion, evaluating the fairness of a municipal voting system, analyzing an ethical problem in AI hiring software and writing a judicial opinion on a pharmaceutical patent settlement.
Each model received a complete case and one exam-style question in a single prompt. The systems had no web retrieval, tools or opportunity to ask follow-up questions. Their answers were checked against equally weighted rubric items derived from instructor solutions.
The paper's primary comparison tested OpenAI's GPT-5.4, Anthropic's Claude Sonnet 4.6 and Google's Gemini 3 Flash Preview. Under the researchers' standard partial-credit scoring, Claude Sonnet 4.6 scored 88.4%, GPT-5.4 scored 87.2% and Gemini 3 Flash Preview scored 81.6%.
Those percentages measure the share of expected rubric elements covered. They should not be read as conventional course grades or proof that the systems could independently make the underlying decisions inside an operating business.
The stricter results show the gap. When a response received credit only if it satisfied every rubric item, Claude Sonnet 4.6 completed 49.6% of the questions, GPT-5.4 completed 47.6% and Gemini 3 Flash Preview completed 32%. Even the leading model omitted at least one required element on slightly more than half of the cases.
The researchers describe the outputs as strong drafts that still require review. Models performed better on structured analytical and explanatory work than on open-ended advisory tasks such as identifying business opportunities. Performance also varied more by discipline than by whether a question was numerical, non-numerical, subjective or objective.
Progress outside math and coding
The paper's clearest result comes from its comparison of four OpenAI models evaluated on the same questions. Standard scores rose from 63.9% for GPT-4 Turbo to 87.2% for GPT-5.4, a 23.3 percentage-point increase over roughly two years. Complete-answer performance climbed from 13.2% to 47.6%.
The gains extended to non-numerical and subjective questions. Under complete-answer scoring, non-numerical questions improved by 36.2 percentage points, compared with 30.4 points for numerical questions. That supports Mollick's contention that stronger mathematical and coding performance has arrived alongside broader gains in analytical judgment.
Fewer than 7% of the benchmark's questions were difficult enough that every tested frontier model scored 70% or lower. A hypothetical system selecting the best answer from the three primary models for each question would have scored 92.8% under partial-credit grading, compared with 88.4% for the strongest individual model. The remaining errors were often model-specific omissions rather than questions that none of the systems could approach.
The benchmark's limits
The researchers used Gemini 2.5 Flash as an automated judge, then checked the evaluation method with three annotators who had business-school grading experience. Automated partial-credit scores had a moderate correlation with human scores, and the annotators rated 96% of the automated grades acceptable or mainly acceptable. The paper says disagreement remained at the individual-question level, reflecting the subjectivity inherent in grading open-ended cases.
BusinessCaseBench also tests a controlled slice of professional work. It is English-only, single-turn and built around the business-school case method. It does not measure whether a model can gather fresh evidence, negotiate with stakeholders, accept responsibility for a decision or adapt as organizational conditions change.
The benchmark joins a growing set of evaluations aimed at economically useful work. OpenAI's GDPval covers tasks from 44 occupations across nine sectors, while BusinessCaseBench concentrates on analytical judgment and the instructor rubrics used to train business students.
For AI companies, that distinction matters. High partial-credit scores mean customers can obtain plausible analyses cheaply. The complete-answer results show where those products remain exposed: missing assumptions, unaddressed stakeholders and incomplete recommendations that may look polished until an experienced operator checks the work.