Invoicing and collecting from chat in a small business: how far an AI agent goes

In a small business invoicing is rarely a software problem: it is an interruption problem. The same person who builds the invoice answers how much a customer owes, sends yesterday’s quote and checks who has not paid. None of those four tasks is hard. Together, they eat the day.

What the agent prepares, and where it stops

It can build the draft invoice with the customer, the products and the prices already in the accounting system. And there it stops.

The draft-with-approval recipe has that pause written into the flow, not offered as advice. The reason is simple: issuing an electronic invoice has tax consequences and voiding it costs more than reviewing it. The speed is gained anyway, because the slow part was never clicking approve, it was filling in the document.

What it does handle end to end

Everything around the invoice that is not issuing it. Sending a quote. Handing over a payment link. Answering how much a customer owes, which is an exact lookup in the accounting system and today interrupts somebody several times a day.

And overdue balances. Chasing them over WhatsApp with the system’s data avoids both of the things that happen when it is done by hand: nobody reviews it for weeks, or somebody writes to a customer who already paid.

What it must not decide

A discount. A term. A credit condition. All three look like commercial details and all three are decisions with consequences, so the agent’s correct answer is to hand the case over.

The invoicing Met has it declared: what it prepares and what it never issues alone. The Alegra entry, for its part, states which lookups and actions are available and which ask for explicit confirmation.

Accounting does not move

None of this replaces the accounting system. The ledger, the obligations and the close stay where they are. What changes is who fills in the forms, and how many times a day somebody drops what they are doing to answer a question the system already knows.