How an AI customer service rep handles a support inbox

A support inbox receives three kinds of mail. Customers with a question. Customers with a problem. And everything else — the agency offering to redo your website, the newsletter you forgot subscribing to, the invoice from a supplier, the person who found your address on a list. In most small businesses all three land in the same place and get the same treatment, which is to say they wait for whoever checks the inbox next.

An AI customer service rep sits on that inbox and treats the three kinds differently. This article walks through what actually happens to a message, in the order it happens, because the order is most of the design.

First, it decides whether to answer at all

Before a single word of reply is written, the message goes through a sorting step that costs almost nothing. Mail that announces itself as automated — out-of-office replies, mailing lists, bounces, anything your mail provider already scored as spam — is set aside on the strength of its headers alone. No model is involved. That catches the obvious cases for free.

Then a small, fast model reads the opening of what is left and answers one question: is this a customer, or a prospective one, writing to this business? A quote request is. A complaint is. Somebody asking whether you cover their area is. A cold pitch for SEO services is not, and neither is a receipt, a recruitment email, or a message with no request in it. Those are kept — you can read every one in the dashboard, labeled as bulk — but nobody replies to them and no credits are spent on them.

Then it answers from what you gave it

The reply is written from your own material: the pages of your website you pointed it at, documents you uploaded, and a description of the business you wrote during setup. It is not drawing on a general idea of what a plumbing company probably does. If your site says you do not service anything north of the river, it says so. If your site says nothing about pricing, it does not invent a price.

That last point matters more than it sounds. The most common way an AI assistant embarrasses a business is by answering confidently about something it was never told. The guard against it is not cleverness, it is scope — the rep is built to answer from what it has and to say when it does not have it. We wrote about the failure modes to watch for separately; the short version is that a good answer and a confident answer are different things, and the product is tuned for the first.

Then it decides whether to hand over

Three things cause a conversation to reach you instead of being closed. The customer asks for a person. The customer is upset — not merely frustrated by a problem, but angry in a way the model recognizes. Or the request is outside what the rep was told to handle: a refund it cannot issue, a dispute, a question it has no material for.

When any of those happens, the full thread is forwarded to the address you chose during setup, with a note saying why. The customer is told a person will follow up. The rep does not keep answering in the background, and it does not pretend to be the person who will follow up. The handover is an email to you, not a mystery.

What it costs you per message

Replies are billed in credits, and one credit is roughly one typical reply — the exact rate is by length, so a one-line answer costs a fraction of a paragraph. A seat includes a monthly allowance, from 250 credits on the Starter plan at $19.99. Unused credits roll over and do not expire. When the balance reaches zero the rep stops answering rather than running up a bill, and the dashboard tells you so.

Messages that are sorted out as bulk cost nothing. Neither does a thread that is handed to you — the rep's part of it was already paid for, and your reply is yours. If you want to understand the credit arithmetic in more detail, the questions page covers it without the marketing.

What it will not do

• Act inside your other systems. It cannot issue a refund in your payment processor, change an order, or look up an account it was not connected to. It can tell the customer what the process is and hand the thread to you.
• Claim to be a person. It uses the name you gave it and says it is an AI assistant when asked. A customer who feels misled later is a customer you lose twice.
• Promise things. It will not quote a price your site does not state, commit to a date, or guarantee an outcome. Where it does not know, it says so and offers to get you.
• Answer junk. A cold pitch gets no reply, and the sender learns nothing about whether the address is monitored.

What setup looks like

You forward your support address to the rep's inbound address, or you give it its own. You point it at your website, and it reads the pages. You choose where handovers go. That is the whole thing — most people are done in an afternoon, and you can try one against a fictional business before you set anything up.