Andrew Funck

HEC Paris · Economics and Decision Sciences · 2026–27 job market

Fields

Innovation, Finance, Labor, Behavioral & Experimental

Job market paper

Human vs. AI in Startup Evaluation: Evidence from an Early-Stage Accelerator

Organizations that screen early-stage ventures increasingly draw on artificial intelligence to inform or substitute for expert human judgment, yet little is known about why algorithmic and human assessments diverge. I develop a Bayesian-learning framework in which an evaluator's assessment combines a prior belief about venture quality with a signal extracted from available evidence, weighted by the evaluator's relative confidence in each. The framework implies that if artificial intelligence extracts more precise signals from venture-specific evidence than humans do, relative to the precision of their respective prior knowledge, its evaluations should respond more to venture-specific evidence and rely less on priors. I test this hypothesis using 579 applications to a French startup accelerator, comparing more than 24,000 scores assigned by expert human judges on 15 business criteria to counterfactual scores generated for the same applications by a large language model prompted to follow the human grading instructions. Consistent with the framework, AI grades are systematically more responsive than human grades to application content, especially for negative information. Linking scores to venture survival, employment, and capital raised, I find that AI scores are at least as strongly associated with performance as human scores. These results are robust to prompting strategy and look-ahead bias but sensitive to model choice: newer, more capable models converge toward human evaluative patterns with no gain in their association with outcomes. These findings suggest that general purpose AI models can already match the quality of human evaluators in screening ventures.