Relying on external APIs network failure points and unavoidable latency from the round trips. There is also the AI API rate limits that come into play. We might find that for critical workflows, local compute is the only reliable architecture.