One common misstep: assuming bigger models and better benchmarks translate to a better user experience. Accuracy may not be enough. If a system is slow or resistant to correction, it alienates users.