A controlled field experiment with 700+ Boston Consulting Group consultants compared task performance with and without GPT-4 access on two task types: those within the model's capability frontier (e.g., ideation, synthesis) showed 38-42% performance gains; those outside it (e.g. tasks requiring validated judgment on financial analysis) showed 13-24 percentage-point losses, driven by overreliance on incorrect AI output. The study finds that cognitive-heavy judgment tasks show the largest performance gains for lower-skilled workers (43% vs. 17% for higher-skilled), and that explicit training on AI capability boundaries moderates but does not eliminate the performance loss on out-of-frontier tasks. For financial institutions, the findings suggest both that AI productivity gains are real but narrow (contingent on task selection and worker expertise), and that blind deployment without capability-boundary training will erode performance on complex judgment tasks central to banking, trading, and credit analysis.
293 other tracked developments touching the same claims.