Teardown: the inbox that drafts itself

One of our ecommerce clients came to us because their mail had stopped being a queue and started being their day. A founder's personal inbox, the shared team addresses, and messages arriving from their store and their social accounts on top.

Around fifty customer emails a day in peak season, and nothing converged anywhere except on a person. Someone was clearing the backlog late at night.

What they asked us for was not an AI that answers customers. It was a way to stop having to read everything.

The obvious build points a model at the inbox and lets it sort. We did the opposite.

Plain rules run before anything else. Blocklists, allow-lists and pattern matches the team edits themselves from the dashboard. Newsletters, noreply notifications and system mail never reach the model at all.

In the month on these screens that removed 5,536 messages, and every one is still counted against the rule that dropped it. The model only reasons over mail a person would actually care about, which made everything after it both cheaper and sharper.

It read the newest message in a thread and politely asked a customer for something they had already sent us. The fix was not a smarter model, it was a sharper rule. Read the whole thread, not the latest message.

It guessed at prices, and it once offered a discount that did not exist. So pricing came out of every drafted reply. Drafts point at the website now instead of quoting a number.

A spam subject carrying the right keyword got flagged urgent with high confidence. Internal forwards between team mailboxes were dropped as noise until the rules learned what a forward looks like. Each one of those was a rule we had not written yet.

Anything touching money, legal or safety is caught before the model has a chance to be clever about it. Refunds over an agreed threshold, chargebacks, legal correspondence and anything diet critical go to a named person instead.

They land in the alert queue with the reason they landed there, the mailbox they came into and how long they have been waiting.

Opening one shows the classification, the recommended action and the whole thread, so whoever picks it up is not starting from a subject line.

Nothing sends itself. Every reply the system writes waits in that mailbox's own Drafts folder, marked as a draft, until a person reads it and presses send.

In the month on these screens it drafted 1,052 replies and flagged 1,204 more for a human. Their founder works two queues and a morning digest instead of a midnight inbox.

There is a path to auto sending and it is strict on purpose. A category has to run in shadow mode for two weeks with almost nothing edited before anyone will discuss it, and money, legal, dietary and complaints never qualify at all. Nothing has graduated yet.