Where this project came from
Marta runs customer service at a food ingredients distribution company. Around 65 people total. She called in mid-January and said her team was not keeping up. Three people, volume growing, a hiring process that takes months and no guarantee the new person stays.
I spent the first week watching how they actually worked. Not interviewing them, but literally sitting next to them and looking at the inbox together. And I quickly saw what was happening: each of those three people was opening somewhere between 150 and 200 emails a day, and most of them were the same questions, phrased slightly differently by different clients.
"Did order number X go through?" Yes it did, the answer is sitting in the ERP, but to check you need to open the system, search, copy the status, write the email. Four minutes. Twenty times a day. Each of the three people.
The owner knew this was a problem. Marta knew. Nobody had time to fix it because everyone was buried in answering emails. Classic trap.
Automation 1: email classifier
Before building anything I went through three months of inbox history. Somewhere north of 40 percent of all emails were questions with a ready answer, where the answer was already sitting in data the company had in their ERP. Stock levels, order statuses, delivery dates, invoice history.
We built a classifier on n8n connected to Gmail via API. Every incoming email gets read, categorized, and sent one of three ways: automatic reply with data pulled live from the ERP, routed to the right person with a short note, or dropped into a manual review queue when the classifier is not confident.
That third bucket is not optional. An automation that blindly answers everything does more damage than just doing it manually. Complaints, anything where the client sounds annoyed, anything the model is not confident about, all of it goes to a human.
The first version of the automated replies sounded oddly formal. We rewrote the prompts twice before Marta stopped flagging them as "not quite right." And for the first week or so the team was manually spot-checking every outgoing automated reply anyway. That took about ten days to stop.
After things settled, the vast majority of those FAQ-class emails go out without anyone touching them. The team went from reading everything to managing a priority queue.
Automation 2: order statuses from the ERP
The second problem was even simpler to identify: clients were asking about order statuses because nobody was sending them those statuses. An order would come in, move through several stages, go out as a delivery, and the client had no information at any point unless they asked.
We connected the ERP via webhook to n8n. A status change triggers an automatic email to the client: what is happening with the order, an approximate timeline, who to contact if something is off. A language model writes the message from the order data so it does not read like a system notification.
In the second week we caught a bug where the template was pulling raw ERP status codes into the email body. A handful of clients received something that said their order was in status PROD_QUEUE_PENDING. We fixed it and resent. Nobody complained directly but it was embarrassing.
After that: a few clients replied to automated emails with thank-yous. They did not know it was automated. One wrote that "there is finally some communication from this company." Slightly funny, slightly sad.
Complaints about not knowing where an order is basically stopped coming. I am not putting a number on it because I do not want to invent precision I do not have.
Automation 3: Monday summaries (and the bug that will ruin your morning)
Every Monday started the same way: Marta and her team sat down and manually compiled purchase summaries for their forty biggest clients. What they ordered last week, what is on the way, what their running total looks like for the month. One summary took about 10 to 12 minutes to pull and format. Forty summaries is over six hours. By the time they finished it was noon.
We automated the whole thing: a workflow on n8n queries the ERP on Sunday evening, compiles the data for each client, generates the email, and sends it at 7:30 Monday morning. The account manager is CC'd if they want to add something personal.
And here I need to admit to a mistake, because this is exactly the kind of thing nobody writes about in case studies and they should. The first summaries went out on time the first Sunday after deployment. The problem was that n8n operates on UTC by default and I had not set the timezone in the workflow. The summaries were showing data from a date range shifted one hour back, so Friday orders placed after 11 PM local time were falling out of the report. Five clients wrote back Monday morning asking where their Friday orders were. We had to manually fix and resend to all forty. An extra week and a half of Monday morning I had not planned for.
The fix took five minutes. But the story is worth remembering if you ever build anything with a schedule in n8n: always set the timezone explicitly.
Unexpected side effect: in the first week three clients replied to the summaries and placed new orders. They saw they were running low on something they order regularly and restocked. The automated email was generating sales.
What it looks like a month later
Altogether the three automations recovered somewhere around 27 hours a week, though the first couple of weeks after deployment it was closer to 20. The team kept checking things the system was already handling. That is normal. It takes time to trust something new.
Marta told me a month in that her team is finally calling clients instead of just emailing them back. Before this they genuinely did not have time for it. Eight hours in a day, six of them on the inbox.
Project cost: one-time fee plus roughly 180 EUR a month in infrastructure. The return is hard to put a number on, but Marta is no longer planning a new hire for the department.
Where to find similar hours in your own business
I do not know how it works in your company specifically, but with every client I have worked with this time is there somewhere. Usually buried under processes nobody has touched since the company was half its current size.
The simplest thing you can do: ask someone on the team to spend a week tagging every email they touch. Not analyzing it, just attaching a label. After a week look at how many times each label repeats. If anything has more than 20 or 30 repetitions, you have an automation candidate.
Second: ask who in the company spends Monday morning compiling something manually. There is almost always someone.
Third: look at how many of your incoming emails are answers to questions where the answer is sitting in a system you already have. No tools needed, just skim the inbox from the last two weeks.
If you want to go through this together, the conversation is free.