You are in a customer-facing meeting with the CTO and platform lead of a global bank evaluating OpenAI for document intake, summarization, and agent assist across mortgage operations. They expect 40 QPS average, 100 QPS peak during business hours, P95 latency under 1.5 s for interactive workflows, and batch throughput of 15 million pages/month, while keeping total platform cost under $0.045 per processed page. In the role-play, explain your proposed design choices, how you would isolate sensitive data, what failure modes you would call out, and how you would justify tradeoffs to both the hands-on engineering team and a skeptical executive sponsor who only cares about ROI, risk, and time-to-value.