A startup is building an AI-powered writing assistant. Their initial architecture: every user keystroke triggers a call to GPT-4o. After launch, their monthly API bill is $47,000 for 10,000 active users. Their CTO says: "We need to cut API costs by 80% without degrading quality." A systems architect reviews the architecture and identifies three wasteful patterns. What are they, and what is the cost-reduction approach for each?