What's the first production problem you hit after integrating an LLM?

When we first started working with production AI systems, we assumed model quality would be the hardest challenge.

Instead, we kept seeing teams struggle with operational issues like latency, cost, retries, rate limits, and observability.

It made me wonder whether the biggest challenge in production AI has shifted from model capability to infrastructure.

For those of you running AI in production:

• What was the first real problem you encountered?

• Was it what you expected?

• What would you do differently if you started today?
  
  I'd love to hear real experiences from builders.