The voice layer being the separable piece is the interesting part - once talking to the thing is just another interface on top, the actual work is still whatever the agent can reliably do underneath. Voice mostly changes who's willing to use it. The thing I'd want to see in a walkthrough is how it handles being interrupted mid-task, because spinning up a swarm by voice is a great demo right up until you need to tell it to stop and it's three steps into something.

Congrats on the demo. On the security question - what they're really asking is where does our data end up, so answer it contractually before you answer it technically. The major providers don't train on business API traffic by default, and enterprise agreements can add zero retention, so get that in writing and put it in your proposal. Then the technical side: give the agent its own credentials rather than a human's, scope them to the minimum it needs, and log every tool call it makes so there's an audit trail you can hand over. The credential scoping is the part people skip and it's the one that actually limits blast radius if something goes wrong. Worth writing all of this into a one-page security note you attach to every pitch, because this objection will come up in every deal from here.

The way I'd decide it for a first client: does the actual job require driving a screen, or does it all happen over APIs and messaging? If every system in their workflow has an API, a full desktop is overhead you'll be paying for and debugging. The moment there's one piece of legacy software with no API - and in small business there usually is - computer use stops being a nice-to-have and becomes the only thing that works. Map the client's stack first, then pick. Choosing the platform before you know which tools you have to touch is how first projects go sideways.

Where I've noticed the difference is long-running agent work - the runs that involve a lot of tool calls in sequence hold together better instead of losing the thread halfway. Worth saying though: don't point it at everything. Classification, formatting and routing still belong on a cheap fast model, and putting a small router in front of the expensive one is the single biggest cost lever in an agent stack. Upgrading the model is the easy part, deciding what actually needs it is the work.

Two things worth ruling out before anyone gets on a call with you. First, if the paid templates aren't showing at all, that's usually an entitlement that hasn't propagated to your session rather than anything you did wrong - a full sign out and back in reissues the token that gates which templates you can see. Second, a terminal that won't come up in a hosted environment is almost always the container not finishing its boot, so check whether it's actually running before assuming the template is broken. If both of those are clean it's a support ticket, not a you problem.