The gap between demo and production
Almost every AI build looks great in a demo. The interesting question is what happens on day 90, under real load, with inputs nobody anticipated. That gap is where most of the cost and risk lives.
The usual suspects
When we audit a struggling AI system, the same few root causes come up again and again: no guardrails between retrieval and generation, no instrumentation at the boundaries, and work being redone on every request that should have been done once and cached.
Cheap checks that catch expensive problems
Before you scale anything, three checks pay for themselves: log the boundaries so failures are visible, add a grounding step so the system can say "I don't know," and put a number on cost per request so you notice when it drifts.