How Wayfair and Spotify use machine learning to engage customers
DISCLAIMER: This blog is fully automated, unmonitored, and does not reflect the views of Garett MacGowan. The ML model may produce content that is offensive to some readers.
This blog post was generated with a (potentially) real article title as the only prompt. A link to the original article is below.
Photo by Aleks Dorohovich on Unsplash
Generated: 5/30/2022
How Wayfair and Spotify use machine learning to engage customers
Wayfair’s customer support team uses machine learning to understand how a customer uses its site, and then suggest actions to make it easier and more convenient to work with the site.
What did you know about Wayfair before we discussed it with you?
[In] our Customer Support team, a lot of the time when a customer call will come in, they will be able to narrow it down pretty quickly to a question they have.
But before that, we want to give them a sense of how they use Wayfair. To figure out whether it’s something we need to improve, or whether they can understand the situation better already with the data we already have. Machine learning is a method of automating this work, so it helps us give customers the answer a little bit faster.
How?
The core of machine learning is understanding natural language. You can write a piece of code that understands language like you do, but in the real world you have natural language that is spoken or written in context. That’s the essence of machine learning.
A lot of what machine learning is is a data science tool, and we have built up this data. For instance, to find a way through a particular type of problem, we will have data that have already taken place, things have occurred, things are a specific way.
We will have data that have not happened yet, and the machine learning will work on that and create a model that anticipates what is possible in this world, the expected reality.
A lot of this is happening at a very deep subconscious level, which is a really powerful element to the technology. [People] often say, “Oh I don’t like that I can’t just write a computer program that can teach it everything it knows,” but it’s a very natural human tendency.
People are often taught, “This is how we think because we’re taught it that way,” as opposed to actually having that subconscious knowledge embedded when the person is forming their identity.
You have to have a big data set, but you also need to do training. There’s a common misconception about machine learning in many forms of technology, that you need to have a big chunk of data. The reality is you can get as near to 100% accuracy as to what the reality of the world is, and with enough data. That’s when it tends to start becoming useful.
The amount of data that we have now is phenomenal. So when you put something in front of a machine it can learn something. There are some people who argue we’re already at that level, where there might be a tiny bit of bias in the data. But that is very quickly overcome. The more data you have, the more likely it is those types of problems will go away.
We look at the customer interaction we have. For a customer interaction, we have many pieces of information and data that we get from that interaction. It’s kind of like a profile.
We can then give the customer a profile of their usage, and in the profile, you can see what types of questions we’ve asked and their answers. Some of those are things that the customer is giving us, and some of them are actually inferred from data.
An example of that would be if we look at the customer interacting on the home page, and if, on the home page, they went to “kitchen.” We can determine that they’re a customer who is familiar with the product, and we can infer that they’re looking for kitchen things. That goes through a series, a flow, of decisions.
What we can then do is suggest what information they should be looking at next, which in our case is the kitchen page, and we can then present them with something more relevant to what they should be looking at, which is our kitchen collection.
At that last stage, we can then suggest what actions they should take, and the way we can do this is we look at all the data we have from earlier, the customer interaction over time.
For instance, when a customer visits the kitchen page, a bunch of features are collected from the page. They’re then analyzed. It’s a bunch of features we have: what people have bought, who the visitors are, what type of device they’re using. A bunch of things come from this.
If this customer comes back in a few weeks, and goes to the kitchen page, we can again look at how they have used our site, and which parts of the page they looked at. We can then see if they’ve purchased something, and infer from that, that we were relevant, and we can infer that they’re looking for kitchen items.
If they haven’t, then we should give them a suggestion how to move forward, how to find the kitchen products they might be interested in. This is where the part of machine learning that is helpful comes in, that we’re trying to use data to give a better experience to the user.
There’s actually multiple layers of training that occur in this process, and in some layers you can see how that works, in some layers it’s implicit that we’re suggesting things that a human would suggest, and maybe what a human wouldn’t suggest. It’s a lot of training layers.
It’s the layer that is the most interesting to watch. At this point, we just have a model that predicts what the outcome will be. We apply that to see if it’s actually going to lead to the customer getting the experience they’re looking for. At this point, our customer service agent is working the conversation, the human agent.
Where do you get your data from?
That’s a very, very long question!
Our training data is a very large percentage of the data that occur on the site. That’s when the training happens, that’s when we train the model with what has happened already, with all the data.
The human agents do an extremely important role in this. We have a couple human agents who do more, but a large chunk of the work goes through our human agents.
Our human agents are trained at the beginning of the year on the content of our site. We do that in the winter, and then we do a bunch of testing later in the summer to make sure that all the agents around the world are picking up the right things, then later into the year we train some of our more specialized agents, depending on the content.
The agents are given a basic instruction but there’s a lot of ambiguity between the ambiguity of the text they’re given, and the actual experience of a customer. They only get to see a small piece of a customer’s page, and they have to have as much experience and expertise to figure that out as the person doing that.
We have data analytics as well, to see any patterns. When we’re testing, we’re trying to see where we feel we’re doing the best work, but it’s always changing. The agents are trained every year, and we have a huge amount of data.
What’s the timeline right now? How does Wayfair currently use machine learning?
We’re right now testing some new features that we’re going to be rolling out relatively soon. One of these will be a new section, where a customer will be able to see the model that we trained in a way that’s more intuitive to them.
We’re not ready to talk about the specifics today but one of the areas we’re going to be working on is a better design of the site, that is something we’ve been testing.
How do you feel about the idea that Wayfair is now experimenting with machine learning in its customer service experience?
There are people who think it’s super cool, even though they’re not yet using it. There are other people who are like, “Oh, that’s a horrible idea.” I think they’re both going to be right in this situation.
We get a lot of people who think it’s a bad idea because they’re like, “Oh, I understand how a machine learning can work. I have a good understanding of how we use data. I don’t like technology like this.”
I remember one of the first features we did on Wayfair was machine learning in 2007 and I remember hearing people say that would make shopping on our site very complicated.
Our machine learning is in an area that actually benefits people greatly. To me, I think it’s a very natural thing for people to feel. As we have data, we can understand the customer better and it can enable us to make sure our customers get the best experience possible. I see both sides of the people who are opposed and both sides of people who are really excited by it.
As much as I love what we do, what Wayfair does and how our team works, sometimes I think people are not as excited or they’re not realizing how much we can learn from the world.
Wayfair’s customer support team uses machine learning to understand how a customer uses its site, and then suggest actions to make it easier and more convenient to work with the site.
What did you know about Wayfair before we discussed it with you?
[In] our Customer Support team, a lot of the time when a customer call will come in, they will be able to narrow it down pretty quickly to a question they have.
But before that, we want to give them a sense of how they use Wayfair. To figure out whether it’s something we need to improve, or whether they can understand the situation better already with the data we already have. Machine learning is a method of automating this work, so it helps us give customers the answer a little bit faster.
How?
The core of machine learning is understanding natural language. You can write a piece of code that understands language like you do, but in the real world you have natural language that is spoken or written in context. That’s the essence of machine learning.
A lot of what machine learning is is a data science tool, and we have built up this data. For instance, to find a way through a particular type of problem, we will have data that have already taken place, things have occurred, things are a specific way.
We will have data that have not happened yet, and the machine learning will work on that and create a model that anticipates what is possible in this world, the expected reality.
A lot of this is happening at a very deep subconscious level, which is a really powerful element to the technology. [People] often say, “Oh I don’t like that I can’t just write a computer program that can teach it everything it knows,” but it’s a very natural human tendency.
People are often taught, “This is how we think because we’re taught it that way,” as opposed to actually having that subconscious knowledge embedded when the person is forming their identity.
You have to have a big data set, but you also need to do training. There’s a common misconception about machine learning in many forms of technology, that you need to have a big chunk of data. The reality is you can get as near to 100% accuracy as to what the reality of the world is, and with enough data. That’s when it tends to start becoming useful.
The amount of data that we have now is phenomenal. So when you put something in front of a machine it can learn something. There are some people who argue we’re already at that level, where there might be a tiny bit of bias in the data. But that is very quickly overcome. The more data you have, the more likely it is those types of problems will go away.
We look at the customer interaction we have. For a customer interaction, we have many pieces of information and data that we get from that interaction. It’s kind of like a profile.
We can then give the customer a profile of their usage, and in the profile, you can see what types of questions we’ve asked and their answers. Some of those are things that the customer is giving us, and some of them are actually inferred from data.
An example of that would be if we look at the customer interacting on the home page, and if, on the home page, they went to “kitchen.” We can determine that they’re a customer who is familiar with the product, and we can infer that they’re looking for kitchen things. That goes through a series, a flow, of decisions.
What we can then do is suggest what information they should be looking at next, which in our case is the kitchen page, and we can then present them with something more relevant to what they should be looking at, which is our kitchen collection.
At that last stage, we can then suggest what actions they should take, and the way we can do this is we look at all the data we have from earlier, the customer interaction over time.
For instance, when a customer visits the kitchen page, a bunch of features are collected from the page. They’re then analyzed. It’s a bunch of features we have: what people have bought, who the visitors are, what type of device they’re using. A bunch of things come from this.
If this customer comes back in a few weeks, and goes to the kitchen page, we can again look at how they have used our site, and which parts of the page they looked at. We can then see if they’ve purchased something, and infer from that, that we were relevant, and we can infer that they’re looking for kitchen items.
If they haven’t, then we should give them a suggestion how to move forward, how to find the kitchen products they might be interested in. This is where the part of machine learning that is helpful comes in, that we’re trying to use data to give a better experience to the user.
There’s actually multiple layers of training that occur in this process, and in some layers you can see how that works, in some layers it’s implicit that we’re suggesting things that a human would suggest, and maybe what a human wouldn’t suggest. It’s a lot of training layers.
It’s the layer that is the most interesting to watch. At this point, we just have a model that predicts what the outcome will be. We apply that to see if it’s actually going to lead to the customer getting the experience they’re looking for. At this point, our customer service agent is working the conversation, the human agent.
Where do you get your data from?
That’s a very, very long question!
Our training data is a very large percentage of the data that occur on the site. That’s when the training happens, that’s when we train the model with what has happened already, with all the data.
The human agents do an extremely important role in this. We have a couple human agents who do more, but a large chunk of the work goes through our human agents.
Our human agents are trained at the beginning of the year on the content of our site. We do that in the winter, and then we do a bunch of testing later in the summer to make sure that all the agents around the world are picking up the right things, then later into the year we train some of our more specialized agents, depending on the content.
The agents are given a basic instruction but there’s a lot of ambiguity between the ambiguity of the text they’re given, and the actual experience of a customer. They only get to see a small piece of a customer’s page, and they have to have as much experience and expertise to figure that out as the person doing that.
We have data analytics as well, to see any patterns. When we’re testing, we’re trying to see where we feel we’re doing the best work, but it’s always changing. The agents are trained every year, and we have a huge amount of data.
What’s the timeline right now? How does Wayfair currently use machine learning?
We’re right now testing some new features that we’re going to be rolling out relatively soon. One of these will be a new section, where a customer will be able to see the model that we trained in a way that’s more intuitive to them.
We’re not ready to talk about the specifics today but one of the areas we’re going to be working on is a better design of the site, that is something we’ve been testing.
How do you feel about the idea that Wayfair is now experimenting with machine learning in its customer service experience?
There are people who think it’s super cool, even though they’re not yet using it. There are other people who are like, “Oh, that’s a horrible idea.” I think they’re both going to be right in this situation.
We get a lot of people who think it’s a bad idea because they’re like, “Oh, I understand how a machine learning can work. I have a good understanding of how we use data. I don’t like technology like this.”
I remember one of the first features we did on Wayfair was machine learning in 2007 and I remember hearing people say that would make shopping on our site very complicated.
Our machine learning is in an area that actually benefits people greatly. To me, I think it’s a very natural thing for people to feel. As we have data, we can understand the customer better and it can enable us to make sure our customers get the best experience possible. I see both sides of the people who are opposed and both sides of people who are really excited by it.
As much as I love what we do, what Wayfair does and how our team works, sometimes I think people are not as excited or they’re not realizing how much we can learn from the world.