A seller from Bulgaria succeeded in a German Amazon Marketplace without knowing the language, yet found a great way to aid his business. For that, he used Natural Language Processing. His name is Radoslav, and he shared his story with me at the Orange Hat Summit conference.
Getting Started With Amazon
When I was young, I was fascinated with computers and wanted to be a software developer. I was amazed at how "something" can be on the web, and everyone can access it and see it. I wanted someday to be able to do such things.
At first, I started trying to make simple HTML websites with a few buttons. A few years later became my full-time job and finally led me to study Computer Science at the university. After I graduated, I worked as a software developer for a few years.
One day a friend of mine asked me if I wanted to join him in his Amazon business and that's how it all started.
German Marketplace Specifics
The German marketplace is the biggest one in Europe. Also, it is very specific with all the German words. Not many sellers know for which keyword to rank their products.
Most people don't like Google Translate, because of its old fame, but it has dramatically improved in recent years. Especially for German, it's quite good. I can't say it's perfect, because nobody can be perfect in German, but it's okay. You can easily understand what customers want and make a simple and understandable reply.
Running Into Problems
As a seller, I can honestly say we have experienced everything. You name it, and we've experienced it. We've had our account closed, encountered fraudulent infringement claims.
The shipping agent lost a few of our bags with orders (we were doing FBM at the time) which resulted in dozens of tickets per day and everything else that you can think of.
I'm saying bags because we were doing a lot of FBM and we were sending all the orders from the day in bulk in a bag. Now imagine a few of those bags getting lost without you knowing. That's a few hundred orders! The customer requests quickly became almost 50 per day, without counting the replies.
We were and still are using an online helpdesk platform which has templates for answers, and you can quickly choose the reply you need to send just like Seller Support does. ;) It's reasonably fast, no more than a minute per answer.
However, the problem is that this 24/7 replying to the same customer messages is boring, because you do the same thing with the same messages over and over again, day after day, week after week.
One time there was nobody around to answer the messages one weekend, and I said I was going to do it. The first ten messages were okay, but on the thirtieth, I was "No way I'm gonna do this two more days. This task is so time consuming and boring".
So I started thinking of ways to automate this tedious everyday task. It reminded me of something I read a while ago, a quote from Bill Gates:
"I choose a lazy person to do a hard job. Because a lazy person will find an easy way to do it."
-- Bill Gates
And that's precisely what I did. Found an easy way to do it.
Applying Natural Language Processing
I have some basic background in Neural Networks from my years in university and a few online courses, but I can't say I have some profound knowledge in Natural Language Processing. What I did is I searched for already built platforms which had this service in them and just chose the one with the best UX.
I wanted to automate the replies to the customer tickets somehow, so our team can focus more on the daily tasks, and not on repetitive customer support. So what I did is I found a platform that provided text classification service and started experimenting with the customer messages and replies to see if I can somehow group them in categories. Most of the times the customer requests were the same: "Where's my order?", "I want an invoice," "I want a refund," "I have a problem with my products" and so on, so it was pretty easy to classify them.
The German tickets are very easy to be classified in the text classifier because the words are very complex and distinguishable. By just looking at a ticket you can easily understand if a customer wants a refund, resend and so on only by seeing some specific keywords.
After the classifier started getting good results, I connected it to the helpdesk software so that you can send the reply to the customer with just one click from the ticket list. Meaning you can answer all of the messages by just clicking a few buttons. Of course, if the classifier fails on some ticket, you still need to do it manually, but most of the time it's pretty precise.
When we had our fail with our bulk orders, there was a person that was answering customer tickets all day long. By integrating the message classifier, we were able to decrease the time to just a few minutes per hour, which is a pretty good optimization. We saved around 5 hours per day by doing that. Nowadays we use mainly FBA, so we don't get so many messages, but the few that we get, they still go through the classifier!
It decreased the tedious task of choosing which template reply to send to the minimum. Moreover, I say boring, because when you do it a few hundred times, you start wanting to do anything else besides that.
>>> How many hours are you spending on mundane customer service replies?
- Artificial Intelligence Saves This Seller Five Hours per Day of Customer Service
- https://www.scaleleap.com/zine/artificial-intelligence-saves-time/ Copy
- Roman Filippov
- Date Published