You are reviewing a generative AI system that answers user questions and may refuse, answer directly, or route to a safer fallback. The team wants a clear evaluation approach that balances safety, factual accuracy, and usefulness, and they need a framework for deciding when the model should answer versus abstain.
How would you ensure AI responses are safe, accurate, and helpful?