A tech company has developed a new large language model designed to act as an automated customer support agent. The model has achieved state-of-the-art scores on several academic benchmarks that measure its ability to recall specific facts from long documents. The product manager argues that these scores are sufficient to prove the model is ready for deployment. As the lead evaluation specialist, you disagree. Propose a more effective evaluation plan that uses real-world tasks. Justify why your proposed plan would provide a more accurate assessment of the model's usefulness to actual customers compared to relying solely on the existing benchmark scores.