“Human vs. AI in Startup Evaluation: Evidence from HEC Incubateur” Draft out soon!
Using the same applications, evaluation criteria, and discrete scoring scale employed by human experts at the incubator of a leading French business school, I generate counterfactual evaluations of early-stage ventures with a state-of-the-art AI model, and develop a Bayesian evaluator framework that decomposes human–AI grade divergence into two structural channels: differences in prior beliefs and differences in signal-extraction precision. The empirical analysis establishes two main facts. First, human and AI grades differ systematically across evaluation criteria and venture characteristics, with AI responding more strongly to text-based signals, especially on negative and verifiable information. Second, AI scores are significantly more predictive of ex-post venture outcomes (survival, employment, and capital raised) than human scores, among both admitted and rejected applicants.
Working Papers
“Decomposing Venture Evaluation: Learning about Belief Formation” with Thomas Åstebro and Francesco Giordano 2nd Round Revision Strategic Entrepreneurship Journal
Evaluating early-stage ventures is difficult because observable signals of future quality are weak and noisy. How do evaluators form beliefs about venture quality, and do their subjective assessments add to the information already available about a venture? Using more than 1,600 screening decisions from a venture incubator, we examine how judges use application information to evaluate ventures. Judges' beliefs about a venture's attributes well explain their recommendations, but the beliefs themselves are only weakly associated with venture quality. We find some systematic miscalibration in those beliefs. However, the miscalibrations seem to even each other out when examining the correlation between recommendations and quality. Beliefs and recommendations predict quality no better than ventures' observable attributes. Experienced judges become more selective and draw on information beyond the recorded criteria, but identify high-quality ventures only slightly better than novice ones.
“Cognitive complexity, expertise and prediction accuracy in venture selection” with Thomas Åstebro and Francesco Giordano
Venture capitalists, business angels, funding agencies, and incubators evaluate ventures, a difficult task where decision uncertainty is high. We examine how the degree of cognitive complexity and human expertise affects judges' admission recommendations at an incubator. Judges read an application, use preset criteria to score it, and form an intuitive overall judgment to accept or reject the application. We model and test how cognitive complexity and judge expertise affect this judgment through a Bayesian classification model, with implications on classification uncertainty and accuracy. Judges demonstrate poor accuracy in evaluating venture quality, but we show that the decision environment is so noisy that they can't do much better. A Bayesian model of judgment captures much of the decision variation. Complexity raises uncertainty and lowers classification accuracy, while expertise reduces uncertainty and improves accuracy only for moderately complex cases.
Work in Progress
“Employment Creation through Psychological Entrepreneurship Training” with Thomas Åstebro, Marcos Balmaceda, Bruno Crépon, Mona Mensman, Naja Pape and Mathis Schulte RCT on-going in collaboration wtih France Travail.
Through a large-scale randomized controlled trial conducted in collaboration with France Travail, this project evaluates whether online psychological entrepreneurship training targeting Personal Initiative and Negotiation skills affects business creation, employment, and related labor-market outcomes among jobseekers. Beyond the average effect of the training, it investigates two further questions. First, on narratives: whether the framing of the training in invitation emails affects course take-up and completion. Second, on chatbot provision: whether delivering the training with different versions of AI chatbot assistants improves learning outcomes. The experimental infrastructure is in place, including the courses, online platform, surveys, IRB approval, AI chatbot variants, the randomization protocol, and the legal and operational set-up with France Travail for recruitment, data access, and data security. Pilots begin on a Qualtrics-recruited sample in July 2026, followed by pilots of the full experimental path with France Travail in Fall 2026, the experiment at scale expected in Spring 2027, and data analysis from Summer 2027.
“The impact of role models on gender stereotypical beliefs about educational choices.” with Adam Altmejd, Thomas Åstebro, Mona Mensman, Ali Mohammadi and Karl Wennberg pre-registered RCT on-going.
This project studies how gender stereotypes about STEM shape the gender gap in STEM education, and whether role models addressing those stereotypes can shift teenagers' beliefs and actual educational choices. In a pre-registered RCT covering roughly 12,000 Swedish students aged 15 to 18 across about 300 schools, classes are randomly assigned to one of nine treatments or to a control group, following a 3×3 design that varies the type of presenter (female role model, male role model, or study counselor/teacher) and the category of gender-stereotypical beliefs about STEM addressed in the one-hour presentation. The study then records students' subsequent educational choices and surveys them on their perceptions of STEM occupations and their educational and occupational preferences. Primary outcomes are enrollment in STEM education, stated STEM preferences, and beliefs about STEM careers. Presentation materials, surveys, and the class-level stratified randomization protocol are in place, and recruitment of schools and professional role models is ongoing. We are currently running the fourth wave of data collection (school interventions), rolling out visits across southern and central Sweden over autumn and winter 2026 to 2027. Survey responses will be linked to administrative records on realized educational choices for analysis per the pre-registration. The project is supported by the Royal Swedish Academy of Sciences and the Kamprad Family Foundation.
“The Impact of Academic Accelerators on Startup Performance”
This paper estimates the causal impact of admission to the Incubator of a leading French business school on venture performance (funding, employment, and survival) exploiting the incubator's rank-based admission process to implement a fuzzy multi-cutoff regression discontinuity design across multiple admission batches. Fuzzy RDD estimates uncover a 25–40 percentage-point jump in admission probability at batch-specific score thresholds and deliver large, statistically significant local average treatment effects on fundraising for marginal admits, alongside positive but noisier employment effects and null effects on survival; density and covariate-balance tests support the design's validity. Ongoing work estimates the marginal threshold treatment effect (MTTE) to characterize how the effect of admission varies along the score distribution, and explores the heterogeneity of these effects across venture characteristics.
Pre-PhD Publications
“Quantitative Analysis of the Costs and Benefits of Delegating Certain Tasks regarding the Implementation of Union Programmes 2021–2027 to the Executive Agencies” with Gabor Katay, Matteo Grigoletto, and Ian Vollbracht JRC Technical Report 121255, 2020
