Voice · Anonymized

A voice system that knows what every call cost, and can stop them all.

Multi-provider voice with per-call cost attribution, deduplication, intent scoring, and an operator kill-switch.

This is a production voice system we built and ran, described here without call volumes, providers, customers, or any commercial outcome, because that delivery is confidential. The point of publishing it anonymized is the engineering, which is the part a buyer of voice automation actually needs to trust.

Voice automation fails in a specific, well-documented way. It is not that the agent cannot hold a conversation — it usually can. It is that spend runs away, the same number gets dialled twice, and when the bill arrives nobody can say which calls it came from. Gartner attributes cancelled agent projects to escalating cost and inadequate risk controls, not to model capability. A voice system that cannot account for its own cost, or cannot be stopped instantly, is that failure waiting to happen.

The operational risk, then, is a cost-and-control risk before it is anything else. The system had to be built so that cost was attributable at the level of a single call, and so that a human could halt the whole thing at once.

What the system does

The system is multi-provider: it does not depend on a single voice vendor, and the routing sits under our control rather than the provider's. On top of that sits per-call cost attribution — every call carries its own cost, so spend is a number you can trace to the conversations that produced it rather than a lump sum at the end of the month.

Before any call goes out, numbers are deduplicated, so the same contact is not dialled twice by a system that lost track of itself. During the call, transcripts are scored for intent, turning a conversation into a signal that can be acted on rather than a recording nobody replays.

Behind all of it is an operator kill-switch: a hard stop a human can pull to halt every line at once. This is what “inadequate risk controls” looks like once it has been solved rather than described — the ability to end an autonomous process immediately, without waiting for it to finish what it started.

What we can prove

The published facts, and the claims we refuse to invent.

Hard, published facts

A multi-provider voice system.
Per-call cost attribution.
Deduplication before dialling.
Transcript intent scoring.
An operator kill-switch.

Not claimed, on purpose

No call volume, minutes, or spend figure.
No cost reduction or conversion outcome.
No languages, providers, or CRM named.
No customer, industry, or campaign detail.

The list above is the whole claim. There is no “we cut cost-per-call by X%”, no conversion lift, no client quote — because none of that is published evidence, and inventing it to dress up a page about cost governance would be self-defeating.

What this means for an automation build

The lesson generalises past voice: an autonomous process that can spend money needs a cost cap and a kill-switch before it needs anything clever. The order matters. Capability without control is the configuration that produces a five-figure afternoon and a cancelled project. Control first is what makes the capability safe to switch on.

Concretely, that means attributing cost at the unit of work — here, a single call — so that spend is legible while it is happening, not reconstructed after the bill lands. It means a hard stop that a human can reach without a meeting. And it means deduplication and other guardrails that stop the system from wasting effort or breaking a rule simply because it lost its place. These are the parts of a voice build that never demo well and matter most in production.

It is why every build we ship carries cost governance and a kill-switch as standard rather than as an upsell. The interesting part of an agent is what it can do; the part that decides whether it survives is what happens when it misbehaves.

What we would not hand to the machine

We would not run an autonomous dialer without a human able to stop it. The kill-switch is not a nice-to-have that can be added in a later phase — a voice system without one is a system we would decline to put into production, because the failure mode it prevents is precisely the one that gets these projects cancelled.

And we would not remove the operator from the loop on cost. The system makes spend attributable and stoppable so that a person can govern it; it is not designed to be trusted to spend without oversight. Autonomy on the conversation, human authority over the money and the stop button — that is the line, and it does not move.

Will it reach production on your workflow?

Score your workflow before you talk to anyone: six questions against what actually stops automation reaching production. The result is immediate, no email required, and it will sometimes tell you not to hire us.

Take the readiness test See the Calling Agent

The same cost governance and kill-switch are built into our Calling Agent — the voice workflow we deploy for lead qualification, booking, and reminders.