Why AI Should Never Send an Invoice on Its Own

An AI that can answer a phone, write a quote and book a crew can also, technically, send the invoice. We decided early that it won't. This note explains where we let AI act on its own, where a person signs off, and why the line sits where it does.

The rule

If an action costs someone money, commits the business to a price, or can't be quietly undone, a person approves it. Everything else the AI may do on its own, and log.

In practice that means the voice agent can answer, ask questions and book a visit into an open slot. Ledger can draft a quote, an invoice or an insurance claim package in seconds. But the quote isn't sent, the invoice isn't issued and the claim isn't filed until someone on the team has looked at it and pressed the button.

Why not let it run?

Because the failure modes are lopsided. A slightly awkward booking costs a phone call to fix. A wrong price on a quote can cost the job, or the margin on it. A mistaken claim can cost a customer relationship. The time AI saves on drafting is large; the time a person spends approving a good draft is small. That trade is worth making every time.

Make approval fast

A sign-off step only works if people actually do it carefully. So every draft arrives with what it was built from: the call transcript, the photos, the rates it used and anything it was unsure of, highlighted. Approving a good draft takes seconds; spotting a bad one is easy because the reasons are right there.

Measure the drafts

Every approval and every edit is a signal. If people keep changing the same line on AI-drafted quotes, that task's model isn't good enough yet, and our evaluation sets get a new case. Over time the drafts need fewer edits, but the person stays in the loop.

Say it's AI

When the voice agent answers a call, it says it's an AI assistant. Callers can ask for a person at any time. Being clear about it costs nothing, and it's the right way to build trust in software that talks to the public.