Every Turn Has to Earn Something

I've been heads down for the last 4 years building agents with a focus on voice, and the more I build, the more convinced I am: latency, voice quality, and getting agents to reliably take actions are increasingly solved problems. The hard part is behavioural science and conversational UX.

Voice deployments have a trust problem. Every caller has been burned by an IVR, and they show up to your agent already braced for disappointment. You're not starting from neutral, you're starting from negative.

For a long time I assumed the answer was better engineering. We've shipped all of it. It matters, but it's not what separates a great deployment from a mediocre one.

What separates them is how much value each conversational turn creates.

Here's the moment it clicked for me. An agent that opens with "How can I help you today?" is just an IVR with a nicer voice. It makes the human do all the work of declaring intent cold. An agent that opens with "I see your delivery didn't make it this morning. I can help. When do you need it by?" does something completely different. It uses what it already knows, invites a real answer, and might reveal that reshipping is the wrong answer entirely.

All the deep engineering work we do at Ada is in service of that one moment. The latency work exists so it feels natural. The model work exists so the agent can respond to what comes back. The infrastructure exists so it scales.

But the moment itself is design.

That's how you earn back trust.