Providers now pick codes via EMR prompts and AI, but general-purpose models hit only 50–59% coding accuracy. Who is checking the code?