Artificial intelligence has moved from pilot to production in underwriting, offering a concrete path to faster workflows and more consistent decisions in complex commercial lines. This controlled field evaluation quantifies how AI-enabled underwriting workbenches—combining automated data prefill, predictive risk models, and remote virtual surveys—affect operational efficiency, decision quality, and portfolio performance in real-world settings. We compare treatment teams using an AI workbench to those using traditional methods over 12 months, tracking time per submission, straight-through processing, underwriting accuracy, override rates, loss ratios, quote-to-bind ratios, and retention rates. System logs capture human-AI collaboration patterns to identify configurations that deliver optimal outcomes across case complexities. Grounded in Technology Acceptance Theory, Cognitive Load Theory, and Organizational Learning frameworks, we develop and test hypotheses about how collaboration patterns influence performance outcomes. Our findings demonstrate that robust prefill and advanced risk models reduce time per submission by 47% (Cohen's d = 1.23, 95% CI: 0.89–1.57) and raise accuracy by 28% (Cohen's d = 0.87, 95% CI: 0.54–1.20), while maintaining regulatory rigor through explainability and auditable trails. Human-in-the-loop oversight outperforms both fully manual and fully automated approaches on decision quality and consistency. A cost-benefit analysis reveals a 3.2:1 return on investment within 18 months, with participating carriers achieving 20–30% book growth through improved prioritization and focus.