Seven days of the self-serve chat agent with no card, pointed at your own site and your own documents. It answers from your content, not a demo corpus, so the thing you evaluate is the thing you would ship. The voice line is open too if you would rather hear it than read it.
Voice, CRM write-back, and calendar booking are build-tier features, because each one needs credentials, an escalation policy, and an eval pass before it touches a real customer. Shipping them behind a self-serve signup is how vendors end up with agents that double-book. You can see them working on the live demo line instead.
Ask it the five questions your team actually gets, including the two you hate answering. Then ask something it should refuse, and something slightly out of scope. Deflection rate on easy questions is not the metric — the metric is whether the hard ones escalate cleanly instead of being answered confidently and wrongly.