Most small businesses do not have an email problem. They have a repetition problem. The same twenty questions arrive every week: delivery times, refund rules, prices, "is this in stock", "can I change my order". Someone answers them by hand, often late at night.
An AI email agent takes that repetition off your plate. Here is how the ones I build work, and where the limits are.
What an AI email agent actually does
An email agent is a small program that sits between your inbox and a large language model such as Claude or GPT. Every few minutes it runs the same loop:
- Read new messages from Gmail, Outlook or any IMAP inbox.
- Classify each one: support question, sales lead, complaint, invoice, newsletter, spam.
- Look things up in your own material: help pages, price list, order system, past replies.
- Draft a reply in your tone of voice.
- Decide whether to send it, hold it for approval, or forward it to a person.
- Log everything so you can see what happened and why.
The model is only one step. Most of the value comes from the plumbing around it: what data it can see, which actions it may take, and who gets asked when it is unsure.
The approval rule that keeps you safe
The most important design choice is deciding what the agent may send without asking. I use three levels:
| Message type | Default action |
|---|---|
| Routine questions answered by your docs | Send automatically |
| Sales leads, price negotiations, anything with money | Draft, then ask for approval |
| Complaints, legal, refunds above a limit | Forward to a person, no draft sent |
Approval should take seconds. I usually send the draft to Telegram or email with three buttons: Approve, Edit, Skip. Owners approve most drafts while waiting for tea.
Grounding replies in your own data
A model with no context will guess, and guesses in customer email are expensive. The fix is grounding: before drafting, the agent retrieves the relevant parts of your knowledge base and order data and passes them to the model with a strict instruction to answer only from that material.
When the material does not contain an answer, the agent should say so internally and route the email to a person. An honest "I don't know" is a feature, not a failure.
What it costs to run
Running costs are mostly API usage. A typical support email plus its context is a few thousand tokens, so even a busy inbox costs far less per month than an hour of staff time per day. Smaller models handle classification cheaply; a stronger model writes the replies that customers read.
How I roll one out
- Week 1: connect the inbox, import help content, run in draft-only mode.
- Week 2: review drafts together, tune tone and rules, add the approval buttons.
- Week 3: switch routine categories to auto-send, keep money and complaints on approval.
- Ongoing: a daily summary email shows how many messages were handled and which needed a person.
Signs you are ready for one
You are ready if the same questions make up most of your inbox, the answers already exist somewhere in writing, and slow replies are costing you sales or reviews. If every email is unique and high-stakes, an agent helps less; a good shared inbox might be enough.
The goal is not to remove people from customer service. It is to remove the copy-paste, so people spend their time on the conversations that need judgment.
If you want an email agent connected to your own inbox and order system, tell me about your setup and I will suggest a first version you can test within a couple of weeks.