Back to blog

Case study: how our AI agent sorts inbound email for five websites

Between 25 July and 21 September 2026, our inquiry agent screened 541 emails with simple rules before any AI model saw them, and only 115 went on to the AI classifier. In the same period it filed 31 new buyer inquiries into our CRM and prepared 22 reply drafts. A person on our team reads and sends every reply; the agent never sends email itself.

We run five websites, each with its own contact form and mailbox. This is the agent that reads those mailboxes for us. We built it for our own business first, and it uses the same approach we take on AI agent builds for clients.

The problem

Five mailboxes fill up with the same mix: a real buyer now and then, and around them sales pitches for web design and SEO, automated reports and notifications, our own test messages, and form submissions from bots. Somebody had to open every message to find the few that mattered, then work out who the buyer was, whether they were serious, and what to ask the factory.

Two things made this slower than it looks. Contact forms send mail from our own address, so the real sender has to be read from the message itself. And a buyer's follow-up ("thanks, when do you ship?") often doesn't look like an inquiry at all, so a simple filter throws it away.

What the agent does

It runs twice a day, at 09:15 and 16:40 China time, and works through the same steps each time:

1. Reads all five site mailboxes. It takes the sender from the form fields or the reply address, never from the form's own sending address.
2. Sets aside the obvious mail with plain rules first. Sales pitches, automated mail, internal messages and bot spam are recognised by cheap checks that cost nothing to run. Only what is left goes to an AI model. The sender's email provider is never treated as a spam sign, because real buyers write from Gmail and Outlook too.
3. Joins follow-ups to the right lead. A reply in an existing conversation is added to the lead we already have, so one buyer doesn't turn into five leads.
4. Scores each real inquiry A, B or C. The score uses only what is in the email: a company domain, a named company and country, a phone number, a quantity, how detailed the specification is. C doesn't mean "reject". It means "don't spend research time yet".
5. Checks where the sender really is. It compares three things that should agree: the country the email was sent from (based on its IP address), the country the sender says they are in, and the language they write in. When they don't match, the lead is marked for a person to look at.
6. Writes a request for the factory in English and Chinese. The English version is kept as the record. The Chinese version is a short list of what the buyer needs, ready for the factory. If there is nothing real to ask a factory, it says so instead of making something up.
7. Emails us a digest. Each inquiry gets a short lead code. To answer one, we reply to the digest with the code and what we want to say, and a draft appears in that site's Drafts folder.
8. Files leads into the CRM. New inquiries go into our CRM in one direction only. A second message from the same person is added as a comment on the existing lead, and nothing a person changed by hand in the CRM is overwritten.
9. Prepares a first reply draft. For each newly filed inquiry it writes a draft in the site's Drafts folder, inside the buyer's own thread. It skips anything older than 7 days and any thread a person has already answered.

The numbers

All figures cover 25 July to 21 September 2026 and come from the agent's own logs.

The useful figure is the first one. Most of what arrives in a business mailbox can be sorted without an AI model at all, so the AI only sees the messages where judgement is actually needed.

What we deliberately left to a human

• Sending replies. The agent only ever saves drafts; its code never sends email. A person reads each draft, edits it if needed and presses send.
• Deciding what to say. When we answer a lead through the digest, the instruction comes from us; the agent turns it into a draft.
• Doubtful leads. When the sending country, the stated country and the language don't agree, the agent flags the lead and stops there. A person decides.
• Changes in the CRM. The agent only adds leads and comments. Moving a lead through the pipeline, or correcting it, is done by hand, and the agent never overwrites that work.
• Old mail. Nothing older than 7 days gets an automatic draft. Those leads wait for a person.

What it would take for your company

If your team spends time sorting a shared inbox or contact-form mail, the same pattern applies: cheap rules for the obvious mail, an AI model for the rest, a written score for each lead, and a draft waiting for a person to approve. The work is mostly in writing down your rules: what counts as a real buyer, what a good first reply says, and what the agent must hand back to a person.

For the build itself, see AI agent development. If you would rather start with a simpler workflow, such as routing forms into a CRM without an agent, see AI automation. Our list of AI agent use cases for small business covers other processes that fit the same approach.