AI education focuses on model accuracy, but real production work demands navigating six crucial trade-offs that rarely appear in coursework. Sara Nóbrega's analysis examines the decisions engineers face once models go live, from infrastructure choices to human oversight strategies.

The build-versus-buy dilemma has evolved dramatically. A 2025 survey found 95% agree building offers more customization, while 91% say prebuilt platforms ship faster. For teams processing over 1M daily requests, API per-token costs can erode margins, though staffing represents 70-80% of self-hosting expenses. Many teams exceed LLM budgets by 340% due to poor cost attribution.

Model complexity creates hidden maintenance burdens. Google's CACE principle shows how small pipeline changes trigger unexpected system-wide effects. Applied ML projects often gain just 2% accuracy from complex models while paying 18 months of debugging overhead. Clean, small datasets frequently outperform larger noisy ones, especially in medical applications.

Real-time inference isn't always necessary. Most business problems—nightly churn scores, weekly recommendations—work perfectly with batch processing at fraction of the cost. Prompt engineering suffices for most tasks, while fine-tuning runs roughly $10k in compute and six weeks of preparation. The key question remains: who actually owns this system in six months?