How we cut our back-office routine from 150 hours a week to 10

We run a trading company. There's a warehouse, an accounting system, a few online marketplaces and a group of wholesale buyers who order from us regularly. Since February 2026, most of the routine behind all that has been done by AI employees. Below is how we set it up and what we got wrong along the way.

Where the time used to go

A lot of our day went on moving numbers from one place to another: checking stock in our books against each marketplace, retyping wholesale orders into invoices, answering reviews, putting together price lists for buyers.

By our estimate this came to around 150 hours a week, about three people's worth of work.

Hiring was the obvious option, but we didn't want to. Each hire costs a full salary, and a new person takes a long time to learn the small stuff: which buyer gets which discount, what a product is called on each marketplace, and so on. So in February 2026 we started handing the work to AI instead.

By "AI employee" we mean an AI that has its own area of work and access to our systems. For example, it creates invoices directly in our accounting software.

The first job we handed over was stock, because it was the most tedious.

Stock

Something would sell out on one marketplace and still show as available on another, or a delivery would arrive and the marketplaces wouldn't get updated in time. We'd end up with canceled orders and penalties from the marketplaces.

Now an AI employee compares stock in our books with every marketplace and fixes any mismatch itself. This is the only thing it changes without asking us first, because stock has to match the books right away and can't sit waiting for someone to approve it.

When an item drops below its minimum, it messages us with a draft purchase order attached. The order only goes to the supplier after we say yes.

Invoices

Wholesale buyers send orders however they like: email, a chat message, sometimes a photo of a handwritten list. One of us used to retype each one into the accounting software, and things got lost or mixed up.

These days the AI employee reads the order, creates the invoice in our accounting system and shows it to us. We look it over, approve it, and it goes to the customer. We issue 10 to 15 invoices on a typical working day.

Listings and reviews

Every marketplace has its own rules for titles, required fields and photos. The AI employee updates our listings and descriptions to fit them, and we check the changes before they go live.

We get two or three reviews a day, and they used to sit for a long time before anyone answered. Now the AI employee drafts the reply, and it goes up once one of us has read it.

Emails to wholesale buyers

We email our regular buyers about new products and price drops. That used to eat a lot of time, because you had to pull together the price list and write to each buyer without mixing up who pays what.

Now the AI employee collects what's new and what's in stock and writes each buyer an email with their own prices. We get a batch of drafts, check them and send.

Profit by channel

This is the one we never really got around to before. To know what each marketplace and our wholesale business actually earned you have to pull the sales and subtract commissions, shipping, advertising and returns. We did it rarely, and not very accurately.

Now a short message lands in Telegram every morning with revenue, costs and profit for each channel and what changed since the day before. Yesterday's figure is provisional, because marketplaces send their commission and shipping reports late, so the exact number comes in a bit later.

Once we saw the numbers every day, we noticed things we'd missed before, including some obvious leaks.

You can also ask questions in the same chat, like which wholesale buyer ordered the most this month, and get the answer right away.

How we give it tasks

We message it in Telegram the way we'd message a colleague, often as a voice note. Something like "raise the price of this item 10% everywhere except wholesale". It transcribes the voice note, works out the new prices and sends them back. We reply "yes" and the prices change. If something's unclear, it asks.

We didn't have to learn a new app. Everyone on the team already uses Telegram.

What we got wrong

Review replies all sounded the same. At first the AI wrote polite, generic replies along the lines of "Thank you for your feedback, your opinion matters to us." Customers notice that kind of thing. What fixed it was showing it a pile of our old replies written by hand and explaining how we actually talk to people.

Product names got mixed up. The same product had different names in our books and on the marketplaces. The AI treated them as different items and the stock numbers drifted apart. We put together a shared list matching the names, and that fixed it.

We gave it too much freedom early on. At the start, the AI employee could email a customer on its own, and one day it sent a customer an email we'd have worded very differently. That's where the rule came from: nothing goes to a customer until we've approved it.

Each time something like this happens, we add a rule to its instructions.

Two rules we stick to

We see everything before it leaves the company. The AI does drafts, reconciliations and reports on its own. Emails to buyers, invoices, review replies, listing changes, supplier orders and price changes all need our approval. Stock is the one exception. Checking all of this takes us about an hour per working day.

Each AI employee sees only its own part of the business. The one that answers reviews can't see our margins. The one that calculates profit doesn't write to customers. We decide what each one can access, and we can revoke it at any time.

Where we are now

We never hired anyone for these jobs. The routine that used to take around 150 hours a week now takes about 10, and half of that is us checking the AI's work.

Each morning we get a rough profit figure for the day before.

If you want to try this yourself

1. Start with one job, the most tedious one you have. There's no need to hand everything over at once.
2. Decide up front what the AI can do on its own and what needs your approval. Anything that leaves the company is worth checking yourself.
3. Give it your real emails, review replies and price lists, so it writes the way you do.
4. After two weeks, compare how long the job used to take with how long it takes now. If you've barely saved any time, pick a different job.

It worked well enough that we now set this up for other companies too.