From the customer’s perspective, none of this matters. They don’t care whether the model reasoned correctly or whether an API timed out. They called because they wanted something done. That’s why I would measure a successful end to end task completion rather than conversational accuracy. That’s why the conclusion makes sense to me. Agentic voice isn’t really just an AI problem. It’s an orchestration problem. The model, tools, permissions, business rules, backend systems and humans all form one system. If those pieces aren’t designed together, improving the model probably won’t fix the underlying failure mode.