In the last chapter, we introduced you to Dialpad’s Call Purpose detection feature and how your business can use it. In this chapter, we’ll give you some insight into our data exploration and process for creating this feature before revealing how these insights led to the development of the feature in the 3rd and final chapter.
While almost everyone has a fair idea of what phone calls are about, only the data can tell us what truly occurs between the caller and the called.
We went through hours and hours of contact center transcripts classifying calls, looking for common patterns in the call opening interaction with regards to when and how the purpose of a call is stated, and prioritizing the importance of these patterns.
We noticed a few patterns in the data, which confirm a few assumptions about call purposes in general and also helped us “teach” our model what to look for in order to find a call’s purpose. Specifically, there were three things.
1. The model had to learn where in the call to look for the Call Purpose statement:
Many people cut right to the chase when they call, especially in customer support settings.
They start with a quick and short greeting or introduction, and then immediately get to their reason for calling.
Good morning, I’m calling because I have a problem logging into the system.
Most other conversations (outside of customer support) start off with some small talk, often about the weather, before proceeding to the purpose of the call.
This is usually done to build rapport or maintain a relationship with the other party.
In the vast majority of calls, the purpose of a conversation is known within the first 40 seconds—but it can occur as far as five minutes into the call!
The model had to distinguish between different types of settings—support, sales, and a whole slew of other types of conversations—and learn where in the call to look for the Call Purpose statement in those settings.
2. The model had to learn to recognize who in the call is stating the call's purpose.
It’s reasonable to assume that the person who initiated the call expresses its purpose, though in certain cases it comes from the call recipient (like when a call is returned, repeated, or scheduled).
While the caller usually states the call purpose, the call recipient guesses, assumes, or inquires.
Caller: Hi, this is Maria. I called earlier today.
Call recipient: Hello Maria. Did you want to talk about the air duct cleaning services?
Some language patterns such as expressions of need may indicate the call purpose if it is stated by the caller, while being just prompts for clarifications if they come from the call recipient:
Caller: I need the floor in my house replaced.
Call recipient: I need your name and the customer number.
The model had to distinguish between similar sounding sentences coming from both the caller and the call recipient and learn to identify who is stating the purpose of the call.
3. The model had to learn which statement in the call captures the defining call purpose.
Some Call Purpose statements are long-winded elaborations of the problem at hand, while others are succinct statements that might require further questioning of their speaker to fully understand.
The nature of the reason for calling will also dictate the length of the call purpose statement—for instance, wanting to purchase a product versus describing an issue with a product.
Hi, I'm calling because we have an AC in our office that isn't working and it's kind of older and very very big and loud and I just wanted to know what your center can really do in terms of giving quotes and coming by to sort something out, especially if we can't come to you.
Hello, I want to order two air filters please.
I have a problem with order 13749.
There might be just one or several reasons for calling. In calls that have several reasons, the reasons could be interconnected or independent, planned ahead of time or arise spontaneously in the course of the call.
Sometimes, a customer has a litany of issues or complaints they want addressed and they may choose to state them one after the other as each is addressed or altogether in one large call purpose statement. Here’s an example:
So basically, I’m calling about that problem with a vehicle that we’re renting. It needs to be repaired. Take a look at this. I’m not sure if you’ve heard what happened but we were on the road, just made a quick stop, and then we couldn’t unlock it. I don’t know if it is a network issue or something is wrong with the locks. I spoke with someone named Michael about it, so maybe you can transfer me to Michael again? Or can he call me back? Maybe you can give me a refund for this because it wasn’t really my fault.
In cases where the reason is given out in pieces or in cases where there are two or more reasons for calling, the model had to figure out which one to denote as the defining Call Purpose statement for the call.
If you’re interested in the stats:
About 30% of calls contain explicit declarations of call purpose, using language such as “I’m calling because” or “the reason I’m calling is”
Another 30% contain expressions of desire: the caller wants or needs something
15% are all sorts of responses to “how can I help you” prompts
About 9% of calls start with a long-winded explanation of a situation directly following a greeting, usually on the caller’s side
Introductions of problems, concerns, updates, and announcements of various kinds together make up another 9%
The rest fall into the patchy “other” category or don’t have a call purpose that can be inferred from a conversation itself—for instance, in a call between two colleagues that’s based on external context known only by those two parties.
So, it turns out, phone calls are not as simple as they seem! In the next chapter, we’ll reveal how this deep dive into the call data helped us build the Call Purpose feature.
Don't miss the next installment of our Call Purpose series about how we built the AI models that underpin this feature.