What Your Intelligent Revenue Department Is Actually Built On

Ask ten vendors to demo an AI phone agent and they will sound remarkably similar. The voice is natural. The qualification questions are sensible. The booking flow works. And then you sign, and three months later you are trying to work out why the thing that demoed so well is quietly missing things.

The demo tells you almost nothing, because the demo is a controlled environment. A single call, a cooperative caller, a clean calendar. What separates a system that holds up from one that does not is what sits underneath it, and that is worth understanding before you buy, not after.

**The reasoning layer.** Relay is built on Claude, Anthropic's AI model. The brand name matters less than the capability floor it sets: understanding what a caller actually means when they say 'my roof is leaking but only when it rains hard,' separating a genuine emergency from a question, and responding in a way that does not sound like a decision tree. Weaker models fail in exactly the situations customers remember, the ambiguous, emotionally loaded, or unusual calls. Conversation data is processed by Anthropic to generate the agent's responses, and Anthropic does not use that data to train its models.

**The infrastructure layer.** This runs on Render's managed cloud infrastructure, for the boring reasons that turn out to be the important ones. Calls arrive around the clock and in unpredictable bursts. A system that degrades under load, or goes down during a storm when every roofer in the county is being called at once, is worse than no system, because you have stopped watching the phones yourself.

**The integration layer.** This is where most AI phone products quietly fail. Answering a call is the easy part. Booking it into the calendar you actually use, writing the record into the CRM your team actually opens, and triggering the follow-up sequence that already exists, that is the work. An AI that books appointments into a system nobody checks has not booked anything.

Relay connects to phone systems, forms, CRM platforms, calendars, email, and SMS. The point is not the length of the integration list. The point is that a conversation which starts on a phone at 9pm ends up as a record in your pipeline by 9:01, with the follow-up already scheduled.

None of this is visible in a demo, and that is the honest reason to ask about it. The question worth asking a vendor is not 'what can it do' but 'what happens when the call is unusual, the volume spikes, and the booking has to land in my actual calendar.' The answers live in the stack.