I feel like I have to keep reminding people at work that LLMs are more expensive and have higher latency and less predictable behavior than plain old deterministic software.

Those tradeoffs must be considered when deciding whether a feature requires regular software or AI.