I see the same mistakes repeatedly:
Fine-tuning when prompting works. Teams spend months fine-tuning only to discover prompting solved the problem. Try prompting first.
Using low-quality data. Garbage in, garbage out. excellent examples beat mediocre ones.
Having data doesn't mean you should fine-tune. Ask: Have I tried prompting with this data? Is it high-quality?
Skipping evaluation. Without metrics, you can't know if fine-tuning helped.