[Faerrg]

From my side, latency would probably be the first thing I’d keep an eye on, mostly because users notice delays before they notice small mistakes. We had a prototype chatbot once that answered questions correctly maybe 85–90% of the time, but when responses took 5–6 seconds people instantly assumed the system was broken. Funny enough, I later read a breakdown here: https://pitchwall.co/blog/artificial-intelligence-services-development-how-to-build-ai-that-works-in-production and it described a similar issue with real deployments — the tech can be solid, but the surrounding infrastructure and monitoring end up being just as important as the model itself. That matched pretty closely with what we experienced in practice.