Podcast | Building and training chatbots for conversational commerce

Jean
Shin

Director, Strategy & Content | Podcast Host of Mobile Interactions Now

24 min podcast
tyntec Podcast Episode 13

In this episode, Malte Kosub from Parloa shares his insights and tips gained from working with top brands across industries to make automated dialogs more helpful (and fun!). Get ready to walk on the happy path – and off-the-happy paths, too :).

Podcast Script:

Jean:
Malte, welcome to the show. We get that little intro in the beginning, but I'm sure they will like to hear more from you. So can you just introduce yourself really quickly?

Malte:
Yeah, sure Jean. Hi, I'm Malte. I'm one of the founders of Parloa. I have been dealing with conversation AI for more than three and a half years now. We started out with an agency, consulting big companies around Europe and then shortly realized that there is a lack of infrastructure for teams building conversation at AI, the building automated dialogues for customers. So we build Parloa for companies to enable their teams building those automated dialogues. I'm a business guy, founded an eCommerce store before, then I wanted to do something a little bit more deep tech. So I decided to do something with conversation AI. We are based in Berlin. We are more than 20 people here and I'm very excited to be here.

Jean:
Awesome. It's great to have you here today, especially because we've been getting a lot of questions regarding chatbots these days, especially with the some of the integration problems and chatbots being dumber than they hoped and all those fun questions. So I will love to just unpack some of the complexity or fun that's involved in making chatbot actually that's useful. So let me start right there. What are we really accomplishing by using chatbots that we weren't able to do before?

Malte:
Yeah, that really is a good question and I think it all is based on a development which started about 10 years ago when more or less deep learning and machine learning algorithm took over and made natural language understanding much, much better. So by now we are able to grasp much more complexity when a customer asks something to a company. We cannot just try to interpret what the user wants to say. So the intent, we can also extract some entity. I can ask my bank how much I spent last week on food. It's not just the intent, I want to know much did I spend. I also can ask to which times last week and the category, so the food. So the way the questions can be structured are much more complex. And through that development over the last five to 10 years and more is currently being implemented in companies.

Malte:
And I think when we talk about natural language understanding, we always need to talk about different parts of the natural language understanding. One thing is the automated speech recognition. The translation from an audio file into text, Microsoft and Google reached 95% accuracy rates. And 95% of the words can be translated right by these algorithms. And they released that in 2017, so about two to three years ago we reached 95%. And what is so important or interesting about 95%, a call center agent understands 95% of what the other person is saying. So we reached human level when it comes to translating an audio file into text about two or three years ago. That's more or less the first step when for example we want to automate the call center. But then very importantly for chatbots, we have the natural language understanding.

Malte:
So this is the software which tries to understand what the user is saying and tries to export the entities I just talked about. For example, a time or category of spendings and also in that space we are seeing a lot of development. Two years ago, roughly one and a half years ago, Google announced Bird, totally new or very big step in that part. So I think also in natural language understanding we're seeing a lot of development, we're seeing a lot of startups evolving. We're seeing the big tech companies investing a lot in that field. So yeah, it's evolving and we are seeing a lot of progress.

Jean:
That's sort of begs a question here. From businesses’ point of view, I understand these enabling technologies are really, really evolving quite fast. But is it more of that what's driving current deployments or is it more of customers that are actually hitting certain tipping point where the demand is just no longer can be ignored? Is it consumer demand or is it more of a tech readiness? What do you see?

Malte:
That is a really good question and I think there's no one answer to that. I think it depends on the channel for example, it depends on the use case. When we talk about digital assistance for example, Alexa and Google Assistant, there is a totally new demand coming up, so it's a demand which was not there before. Companies need to think about how to be present on those platforms, so that's totally new. When we talk about call centers for example, there is a high demand and you probably also know these old IVR system, please press the one, please press the two. And this can be done much, much better. So I would say it's not like direct customer demand but there is the possibility to make an experience better. So that's regarding call center. And for chat I think this is slowly starting that customers really demand that there is a chat more or less on which they can use to contact the company.

Malte:
So rather it is or it could be on the website, it could be WhatsApp, it could be in Facebook Messenger and I have the feeling that this is evolving, especially chat is evolving pretty fast and the demand is getting more and more important. On the other side, we see a lot of companies and we have been working for the biggest companies in Europe when it comes to conversation AI and they all started a few years ago because they saw there is something happening. And they all started with small POCs for small use cases and now are integrating those teams and the technology into their main infrastructure. So I think it's a little bit of both and it depends a little bit on the channel.

Jean:
I understand your company works with the different interfaces, voice and chat apps and all those. But we're starting to see a lot of chatbots actually appearing inside of messaging apps and things like this. Why you think this is happening now?

Malte:
I think that why more and more chat channels coming up and I think, is that the question, why are we getting more and more channels in the ecosystem. And I think that is because first of all, sure the technology evolved over the last years and users are getting used to write a lot via their smartphones for example. And that's where the customers are. They are at WhatsApp, they are on Facebook Messenger, they're getting used to chat to get their message out. And I believe that companies want to be there where the customers are. And if it's easier to write to a company via WhatsApp because the people are there, than searching for a hotline on the web and it takes you two minutes to even find the hotline. And then call them and wait four minutes in the queue. It might be a lot easier and faster to just write them.

Malte:
And maybe the first 30 seconds is automated and maybe then you can directly write or chat with an agent. So I think companies already want to be there where their customers are. First and second, when we talk about bots, the technology got better over the last five to 10 years and also for platforms it takes some time to integrate and make it possible. And now we're seeing that first [inaudible 00:10:09] with WhatsApp that they make it possible for companies to also integrate good bots and systems to write with customers. So maybe these are reasons for that change.

Jean:
I was chatting with some of my colleagues in terms of some of the value added things that happens when you have these interactions somehow integrate to the backend. Is there inherent advantage where texts based interactions, how it's used with the analytics versus some of the voice interface or something like that? Is this part of what's driving in the backend as well?

Malte:
Yeah, I think chat has the advantage versus, when we compare it to voice that we don't have the automated speech recognition, so we doubtedly have the text there. We don't need to translate the audio file into text. So there's one component which can cause some errors, which you don't have in chat. Then we have a visual component there, so we can show something to the user and always when we can show something to the user, it actually can be faster. Because you can read much faster than you can listen to someone who answers to you. So I think the chat inference is a lot of potential and advantages versus voice.

Malte:
However, if you talk to someone via voice, you as a customer can articulate your questions much faster. You can speak roughly about 140 words a minute and you can write, I don't know, it depends a little bit like roughly 50 or 60. So you're much faster and when you speak. But for when it comes to a bot, I think having a chatbot is a little bit easier because you have this visual interface. You can have suggestion chips or a suggestion button. And so you can give the user some feedback what the user can say or what the user can click. And in voice it's a little bit more complicated. So I think there are some advantages when we talk about chatbots versus voice bots.

Jean:
That's really interesting. So we talked about the having service in a way that answering questions in a fast way. And I consider myself as a rather impatient consumer when it comes to that. So can you take us into the back room of companies doing business, is what really is involved using systematically this kind of technology to speed up that interaction to the level where today's consumers can be satisfied with.

Malte:
Okay. So I think that's a very important question. What should companies do to use that technology and to build cutting edge experiences? I think there are different parts which companies need to do. The first thing is they need to understand that conversation that AI is a whole new kind of knowledge they need to integrate into their company. So there are voice user interface designers or conversational designer who build a speech model who know how natural language understanding works.

Malte:
So you first need to build up a team, a dedicated team maybe together with an external agency or you can just do it internally which knows how to work with conversation AI. This team then needs to train the bots if you want to do parts of it in an automated way, which trains this bot. Because if you start to build a bot and then you deploy it and then you just started with it, the journey just begun and then the actual work starts. Because you need to make the bot much better.

Malte:
You start maybe with 20%, 25% automation. Then next week you're 26 and 27 and so every day you need to look at the past conversations, look what worked, what didn't work. And then you need to use that. We call it training data to make the conversation better. And just to name one example, someone ask a question and the bot respond, "Sorry, I can't help you with that." Or, "I don't know what you meant." So now I need to take that sentence and to integrate into my speech model so the next time someone asks this question, it gets answered right. And there, you or companies need people sitting there just doing that. It's actual work. And I think a lot of companies think, "Well it's AI and everything is magical." Well yeah, there is some machine learning in there, but in fact at the end you need people to see what is working, what is not working to train the model.

Malte:
So you need a team first, you need the team to work on it on a daily basis if you really want to build a good bot. And last but not least, you need to have good APIs in place. And if you just want to build an FAQ bot without any connection to a backend, this might be a good start. But if you really want to build good use cases, you need to connect the dialogue, the experience with your backends. And we experienced a lot of trouble with that because a lot of companies starting to build a lot of good APIs and scaling APIs and fast APIs. But you really need to have those to build end to end use cases. And if you have all those three things in place, I think you really can build good experiences. But it's not something which is there automatically if you start and then it's there. It is a process and you need to keep optimizing it. But those three things I think are important.

Jean:
When you describe the bot to start say, "Oh sorry I didn't get that." All those scenario, you are describing my past couple of weeks. I was working with a chatbot developer for a simple nonprofit organization, FAQ. Q&A, and you start facing yourself. Some people don't say no, they say nope, sure thing. And completely different way, yeah. But as I was keep looking at revisiting the flow and keep working with this script, I was starting to be a little resentful and I was going, "Okay, are you training me or am I training you," kind of scenario. So now here's your chance. Is there some tips, tricks as a pro that what they have to really expect when they're start really training the bots. So that it doesn't get into the anger mode I was on last week.

Malte:
Yeah. That's a very good question. I think you will, in the next years you will not be able to answer 100% of the questions. So there will always be 30, 40 maybe 20 maybe 60% of questions which are answered wrong. So there must be something which satisfies the customer. And we really try to, and that's maybe a tip for everyone doing conversational design. Try to think about, not the happy path. Also try to think about the error path and try to do a very good error handling. So try to, first of all, one more time can the user rephrase what he said? And maybe at the second time you can be also a little bit ... And be transparent, say, "Well sorry, I'm not able to answer that. I'm still learning. I'm learning every day." So I think it's important to set the expectance right for the customer. And don't say, "Well I can do everything." And then you can't everything and the customer's unsatisfied.

Malte:
Be transparent. Say, "I'm learning." And then you can be, it depends a little bit on the company's profile, but you can be a little bit like be friendly to them and maybe have a little joke. And I think this really makes an experience differently and maybe, even if you can't help the user, the user say, "Well they try, they get better. And I know it's not perfect." So it's always important, not just think about the happy path and always and spend a lot of time thinking about error handling and about those situations when something is not working. I experienced that a lot of companies spend too less time on the things which are not working than just focusing on the happy paths.

Jean:
I absolutely love what you just said. Because it is a completely different way of looking at my last week's experience. And it also makes me think, I think at the end of the day, even machine interaction, machine to human interaction, I think ... I may not be remembering every response I received, but it's the feeling that I got. I interacted with the machine. There was a better than expected interaction, that little smile that I can walk away with. So I absolutely, if I imagine my user leaving the experience with that, that's what I'm hearing. So you put me into a completely different mode of thinking about this.

Jean:
With that goal though, what you are describing also opens up this, like there's a functionality of a chatbot and there is this, there's a style part of this almost, like a personality. And as you mentioned, it depends on what the company is and if there is an appropriateness to this, right? What kind of personality you take on. Because you may not, some people may not understand the interaction just saying "Hey I'm still learning." I'm like, "Then why are you already deployed, right?" So tell me a little bit of like how brands can think about this in terms of a matching up with their personality that they want to personify.

Malte:
Yeah. Three thoughts on that. I think the first thing a persona really needs to fit to the brand. I think to be friendly and maybe to be a little bit funny is always good in such a dialogue. Enter teaching and parenting, "Well, we're just learning." But each company needs to decide that by themselves. But it's important to think about it. A lot of companies just like having a lot of developers develop the experience, but then it's a experience developed by developers. If you know a company would develop a webpage just by developer. You have UX-designers and frontend designers and then they have developers implementing that. And I think it's the same when it comes to conversational development, you need developers to develop it and you need also designers to design it. And one thing is you have a corporate identity for your page and you also should have a corporate identity for your bots.

Malte:
Second thing is, you said, "Well you should first deploy it when it's perfect." And that's impossible right now. So just try to give the user an impression that there is a point when we need to start and we thought we might be able to start right now and you are helping us to make it better. We need your feedback. I think you need feedback from real users to really make it better. Because if we both together try to generate a lot of training data, we will never be able to think about all different ways a customer can express themselves.

Malte:
So there are 5,000 ways to ask for the weather, 5,000. We both would never be able to come up with all the different ideas. So you need training data, you need real customer feedback. And the third thing is, in best case you have real agents which are then being connected to the users. So say, "Well unfortunately I can't answer that. I'm still learning. I'm still learning every day. But I really want to help you so I'm connecting you to the real person." And if you do that you can learn and even if you just automate 20% or 25%, the rest you are connecting to a real person and then the customer at the end is satisfied. So I think that to just have a bot without real persons at the end can work. I would suggest, always suggest to have a real person. So to have a human handover when the bot is not working properly or just can't answer the question.


 

Part 2 of the interview with Malte Kosub will be released in two weeks, following this release of Part 1.