Where should an AI demo stop and a human take over? (building this, launching Aug 25)

Hey Malte here, I am the founder of Demoloop. We're launching on Product Hunt on August 25.

Demoloop puts a "See it live" button on SaaS websites: visitors talk (or type) with an AI guide that demos the real product in an isolated sandbox and hands the sales team the transcript, qualification, and objections. This was something that I was genuinely looking for when I was building a very complex compliance tool and when I was working in software development five, six years ago.

The design question I keep wrestling with, and where I'd love this community's take: where should the AI stop? Right now our answers are strict, the agent only cites knowledge you approved, demos follow bounded paths, and a human is always one click away. But there's a real spectrum between "safe but limited" and "capable but improvising," and every AI-agent launch thread lately ends up debating exactly this.

So: if a voice AI was demoing your product to your prospects, what would you let it handle, and what must stay human? Pricing questions? Objections? Custom use cases? Actually using your app?

I'll take every answer into the build before launch day. And if you want to see where we land, follow Demoloop here and you'll get pinged on the 25th.