Back office automation for a small business: what to hand over first
Sergei Pak · designs AI operating systems · русская версия
The back office is everything a customer never sees. Mail, bank statements, invoices, reports, reminders to a contractor. In a small business the owner usually does this personally, in the evening, after the real work. Back office automation means these operations run without them: on a schedule, when a letter or a statement arrives, and only what needs a decision reaches the owner.
Five places the time goes
Most owners I've talked to name almost the same list:
- Mail and messengers. Hundreds of messages, a few dozen of which matter.
- Bank and books. The statement came in, and nobody has time to match it until month end.
- Documents. A contract, an invoice and an act sit in a chat, and nobody checked whether the amounts and dates agree.
- Reports. The same tables opened every morning just to see where the money is.
- Deadlines. Who owes what to whom next week lives in one person's head.
Hiring and purchasing are back office too, but in a small company they rarely take more than a couple of hours a week. They're a poor place to start.
A rule first, then a model
The most common mistake is handing everything to a large language model right away. That kind of sorting is hard to check. I go the other way: first I measure how much a plain rule can close, and only the remainder goes to a model.
Here's my own mail (measured over the 30 days to September 19, 2026). 3,514 messages arrived, and judging by sender address and unsubscribe links about 98% came from a robot. 91% of all mail turned out to be marketplace seller notifications, mostly "new order", about a hundred a day. A table of rules by sender and subject template sorts them, no model at all. Real people sent an estimated 60 messages that month.
The decision model works after that, on the ipak.dev and Acceptum work mailboxes. It checks for phishing and for instructions aimed at the assistant, and for mail from people it also decides whether a reply is needed. It picks each answer from a fixed list and reports its confidence. A call costs a fraction of a cent, and every decision can be checked and explained.
Before and after
My mornings used to start with mail, the bank and a couple of spreadsheets. Now, by six, the mail is sorted, the cash position is recalculated after the overnight statement, and at 08:45 Telegram brings cards for the questions that can't close without me. I decide with a button. Right now the business runs about two hours a day without me at a screen, and the target for December is four. How I measure that is in the post on hours of autonomy.
For an owner the order of work is simple:
- List what you do by hand every day.
- Count the monthly volume for each item.
- Close with rules everything that follows a template.
- Give a model only what's left, and decide in advance what the system does when the model isn't sure.
Where this lives
All of these processes sit in one system around the owner: memory of the business, data collection, a decision layer and automation on top. I call it an AIOS; more in what an AIOS is. How the bank gets matched to the cent is covered separately, and how the system works overnight while everyone sleeps is in While I sleep.
FAQ
- What is back office automation?
- The back office is the internal work a customer never sees: sorting mail, reconciling the bank with the books, documents, reports, chasing contractors. Automating it means these operations run on a schedule and on events without a person, and the owner only gets what needs a decision.
- What should a small business automate first?
- Whatever repeats every day and eats the most time. Usually that's incoming mail and messengers, matching bank statements to the books, and contract deadline reminders. Start by counting: how many of these operations happen in a month and how many a plain rule can close without a model.
- Do you need AI for this?
- Not everywhere. In my mail, 91% of messages over a month are marketplace seller notifications, and a table of rules by sender and subject sorts them, no model involved. A model comes in where a rule falls short: mail from people, free-form documents, judgment calls. That way every decision is easy to check.