Company AI agent: one inbox, orders, enquiries and mailings you can actually measure

A made-to-order office furniture manufacturer, 25 people, a factory floor near Wałbrzych. Everything entered the company through the biuro@ inbox (the office address) and through the owner: orders from regular customers, quote requests, complaints, questions about dates, replies to the mailings the accountant sent from Outlook once a quarter. The owner read it all himself and handed it out. We measured it: around three hours a day, out of a working day that was already eleven hours long.

Mailings are a separate story. The company knew how many emails it had sent. It didn't know how many people replied, how many of those became orders, or who had asked to be unsubscribed. Those requests also went to biuro@ and sometimes got lost, which under GDPR is a real problem rather than just a mess.

How an agent differs from an automation

The automation from the post about orders does one thing: an email comes in, a document goes out. An agent gets a goal ("hand out today's work, keep an eye on deadlines, run the mailing") and a set of tools: the inbox, the ERP, the job sheet, the calendar, the owner's WhatsApp and the mailing system. It decides for itself which to use, and writes every decision down in a log. That log turned out to matter more than I assumed, more on which below.

One Tuesday from the log

Instead of listing features, I'll show one day, taken from the agent's log in week three.

At 6:00 the agent sends the owner a report on WhatsApp: what's in production today, which invoices are overdue, what's waiting for a decision. At 8:10 an order comes in from a regular customer for six desks; the agent looks the customer up in Subiekt, takes the discount from the contract, creates a ZK (a customer order document) and replies with a date from the production calendar. At 9:30 an enquiry arrives for 20 desks for a new office. Here the agent only prepares a draft quote from the price list and sets it aside for approval, because it's a new customer and a large amount.

At 11:00 it sends a mailing to 1,240 companies in the "offices with 20+ people" segment, with UTM links. At 11:40 the replies start coming in; the agent sorts them into "interested", "question", "not now" and "unsubscribe me". Each gets a task, and "unsubscribe" leaves the list the same day, with no human involved. At 14:05 a complaint: the agent doesn't handle it, it just summarises it, passes it to Marek in production and sets a 24-hour reminder.

At 17:30 the wrap-up: 14 tasks handed out, three waiting for the owner's decision.

The day the agent invented a date

At 15:20 that Tuesday a customer emailed asking when he'd get the cabinets he'd ordered. The agent replied: "by Friday". In Subiekt the completion date was two weeks away. The agent didn't check the ERP; it most likely took "Friday" from an earlier thread with a different customer that it had in its conversation memory. We never found out for sure.

We learned about it the next day from the log rather than from the customer. Without the log it would have come out on Friday, from an annoyed customer. With the log it came out on Wednesday morning, and the owner called the customer himself before he had a chance to be surprised.

The fix: every date, amount and stock level in a reply has to come from the ERP or from the job sheet, and the agent records in the log where it got it from. If it can't find one, it writes "I'll check and get back to you" and creates a task. Six weeks have passed since that change without a second case. The owner still reads the log once a day over coffee. He says it takes five minutes and he sleeps better.

The first mailing that could actually be measured

That Tuesday's mailing after 30 days: 1,240 sent, 612 opens, 97 replies, 23 quote requests, 6 orders worth around PLN 41,000 in total. Before, the company only knew the first number.

Two things came out of it. 29 people replied "not now" and will get a reminder in 60 days; before, those emails got lost in the inbox and nobody went back to them. 14 people asked to be unsubscribed and all of them left the list the same day, with a log entry you can show if anyone ever asks.

What still doesn't work

The agent struggles with a customer who changes his mind in three consecutive emails within an hour. It makes three tasks out of it instead of one. The workaround: tasks from one thread within the last two hours get merged into one, but sometimes it merges too much. I'm still working on that.

The second issue is cost. An agent that reads whole threads and checks three systems for every email costs more in API fees than a simple automation; at this client it's around PLN 350 a month. Against three hours of the owner's day that's small change, but you need to know it before the start rather than after the first bill.

Efekt: The owner got back around three hours a day. Orders from regular customers go through without him. The mailing has numbers for the first time, and unsubscribe requests are handled the day they arrive. One slip with a date and one more rule.