AI projects fail in production for boring reasons: bad data, no monitoring, no owner, and prompts pretending to be architecture.
No eval loop
Teams ship demos without a test set of real user questions. Accuracy drifts. Support learns to distrust the bot.
Integration as afterthought
CRMs, ticketing, auth, and billing live outside the prototype. Production needs webhooks, retries, and idempotency — not a curl example.
Cost and latency surprises
Without caching, routing, and model selection, bills spike at scale. We design cost controls up front.
What works instead
Start narrow, measure weekly, keep humans in the loop, and treat AI as acceleration for engineers who already ship. That is how we build agents and RAG that survive contact with customers.