A voice agent that takes the claim over the phone

The bit everyone skips

Most conversations about AI in claims start at assessment — fraud scoring, valuation, settlement recommendations. Fair enough, that's where the money is.

But the claim has to get into the system first, and that part is still a person typing.

Someone rings in after a loss. A handler picks up and starts typing while they talk, trying to keep pace with the date, the description of what's gone, the crime reference number, whether there's a valuation, whether there are photographs. Half of it lands in a free-text box. Anything missed becomes a callback tomorrow. And because the notes aren't structured, someone re-keys the important parts into the claims system afterwards.

Nobody designed it that way. It accumulated.

What we built

A voice agent that answers the claims line and has the conversation instead.

Not a phone tree. Not a form read aloud with a synthetic voice. It listens, understands what's being described, and asks the follow-ups a handler would ask: when did it happen, what was taken, do you have a crime reference number, do you have a valuation or receipt, do you have photographs.

As the conversation goes on, it writes the claim record. Each field is structured, and each one carries the agent's own confidence in what it heard. A crime reference read out digit by digit scores high. An estimated value mentioned in passing, mid-sentence, doesn't — and that's the point.

By the time the caller hangs up, the claim is open in the queue, the transcript and recording are attached, and they've had a secure link to upload their documents.

Where the person stays

This is the part that decides whether a build like this is worth having.

Anything uncertain is flagged, not guessed. If the agent isn't confident it heard a value or a reference correctly, that field is marked for a handler with the audio to check against. It doesn't quietly commit a plausible number and hope.

Distress escalates immediately. A caller who is upset, confused, or simply asks for a person gets one. There is no clever retention loop.

Every decision that costs money stays human. The agent takes details. It does not decide whether a claim is covered, quote a settlement, or turn anyone down. Those are handler calls with handler accountability, and moving them to a model would be the wrong kind of clever.

Why it's built this way round

The instinct with voice AI is to see how much of the job it can take. That's backwards for anything regulated.

The right question is which part of the job is genuinely mechanical. Capturing what someone tells you, accurately, in a consistent structure, every time regardless of who is on shift and how busy the queue is — that is mechanical. Deciding what a claim is worth is not.

Split it that way and you get something you can actually defend: intake gets faster and more consistent, handlers spend their time on judgement instead of typing, and every decision that matters still has a name against it.

Concept build. All data illustrative. No client work is described here.