Inbox census
For seven days, count everything that arrives: by type, by sender, by what happened to it. You cannot automate a mailbox you have never measured, and owners consistently guess wrong about their own top category.
Inbox Ops Autopilot replaces the inbox-operator role, the sorting, routing, and repetitive replying nobody put on the org chart, with a system. It reads every inbound email, classifies it, routes it to one owner, and drafts replies for routine requests. It never sends anything on its own. A human is on the send button, always.
The first build took longer while the trust gates were earned.
For seven days, count everything that arrives: by type, by sender, by what happened to it. You cannot automate a mailbox you have never measured, and owners consistently guess wrong about their own top category.
A named list of the decisions the business actually makes about email: answer, route, schedule, pay, escalate, ignore. Written in the client's own vocabulary. This taxonomy is the real intellectual property of the whole system.
Before any AI touches anything, auto-archive and unsubscribe everything that never needed a human: notification noise, newsletters, machine-generated receipts. The cheapest win, and it makes every later step more accurate.
Every email type in the taxonomy gets exactly one owner and one label. One, not two. "Whoever sees it first" is not a routing rule, it is the mechanism by which threads die quietly.
Now, and only now, the AI arrives. A language model reads every inbound thread, applies the taxonomy, and scores urgency. The output is just labels. A good classifier disappears into the furniture.
For routine request types, the system writes a reply and saves it as a draft. A human reads it, edits it, and presses send. The system has no send capability at all, by architecture, not by policy. This is the single most important design decision in the map.
Named VIP senders, money above a threshold, legal language, genuinely angry tone: those interrupt a human immediately. Everything else waits calmly in the queue until the next review window.
One structured brief replaces two hundred pings. Each morning the owner gets a single summary: what arrived, what was drafted, what is waiting on a decision, what got escalated.
Every week, review the misclassified threads and every draft the human rewrote heavily. Feed the corrections back into the classifier and the taxonomy. This is why it is a monthly system, not a script that rots in six months.
An inbox-heavy small business is typically paying for four roles, usually as fractions of everyone's day, including the founder's. The nine steps absorb the sorter, the router, most of the answerer, and the chaser's memory, while leaving every actual decision, and every send, with a human.
Decides what each message actually is.
Decides who owns it.
Types the same handful of replies on loop.
Bumps the threads nobody answered.
This space is full of invented numbers. Surveys like the SBE Council's find owners reporting around five hours a week saved with AI tools, and that figure is self-reported, which means treat it as sentiment, not measurement. I do not promise hour counts. I show mechanisms and let clients measure their own before and after.
No. It has no send capability. It classifies, labels, routes, and drafts. A human sends everything.
Misclassifications get flagged in the weekly drift review and fed back as corrections. Guard rules also flag threads that appear already settled so they are never re-answered.
$2,750 for setup, $395 per month for run and tune. One mailbox, targeting a three-week go-live.
Workflow automation tooling and a language model classifier. The stack is deliberately boring. The taxonomy and the trust gates are the product.
The execution, the trust gates, and the weekly tuning are the business.