By clicking “Accept All Cookies”, you agree to the storing of cookies on your device to enhance site navigation, analyze site usage, and assist in our marketing efforts. View our Privacy Policy for more information.

Get a demo

Webinar: 5 Strategies for Conversational Design

May 30, 2019

Hello everyone! My name is Edan Soroker and I’m a solution, I’m, excuse me, I’m a solution designer with Pypestream, and we’re here today to talk about conversational design – specifically about five key strategies that we employ to build beautiful, elegant, and effective conversations within the realm of conversational AI. We have a pretty quick agenda for today: first I’ll give a brief introduction to who we are at Pypestream and how we approach conversational AI from a bird’s eye view; then we’ll get into it talk about those best practices, our five key strategies to building conversations in an effective way; and at the very end we’ll have some time for open questions and answers.

Okay, let’s get to it! Conversational AI, what exactly do we mean when we say this term? First of all, there has to be an aspect of automation, of course this is AI, our solution provides responses that are dynamic to the user and contextually relevant without necessarily having any human intervention on the back end. Also, we are talking primarily about messaging based communication, so, as you can see here on the left hand side in a mobile or even a desktop view, we can include text as well as some richer assets such as buttons, menus, selections, maybe images, videos, and so forth. We also have a component of natural language understanding or NLU, this is the technology that a conversational AI solution utilizes in order to understand free text user inputs and utterances in order to make sense of them and be able to respond accurately. Last, we also have a component of back-end logic execution, usually via technologies such as API web integrations, which is how we can make our solutions be dynamic and transactional in their nature.

Who are we at Pypestream? We are a full-stack solution for conversational AI, we’re built for scale with enterprise grade security working with clients within the fortune 500 across a variety of verticals including financial services, telecom, insurance, and travel.

Okay, let’s get to it, building those conversations. So, our first strategy is don’t just know the user but you have to really try and be the user and our goal is to be a truly user centric-company in our approach to building solutions and we do this by employing specific tools and methodologies. Our overall attempt here, and really what we strive for, is to see the solution from the eyes and situation of our customers so that we can build the optimal experience for them – and how do we do this? First of all, at the very beginning of a new project or implementation before we start talking about steps or flows, getting from A to B, achieving success, any of that, we really begin with generating some user personas to ground us in that user-centric approach. We start with just maybe a name and a face, some key demographic information to give a little bit of substance to our persona and then we embark upon an empathy mapping exercise – that is focusing on the current state today before we have a conversational AI solution in place, how does this user persona achieve their goal and get what they want with completing their task. We do this by looking at what a persona is thinking, feeling, doing, and saying today in this interaction and our goals here are, first of all, of course, to uncover the different pain points that the user is experiencing so that we can build a valuable solution for them and improve this experience drastically with the use of conversational AI; and it’s also a methodology that we use to help us get a 360 degree view of really how the user is interacting with the company, with the tools available to them today and really help us, kind of, internalize this as we build and test out our solution.

Next, we shift our focus to the future state, so taking our user persona and this empathy map of what they’re going through when they’re trying to accomplish their task today, we look to the future with the help of a conversational AI solution and we generate what we call a project hill to guide our project. This project hill takes on a structure of who, what, and wow. Who, who is our user persona, will be able to accomplish what, which is of course the task or the given use case, and what is the given wow factor that we can really make that experience delightful and effective for our users. In addition to this, we really emphasize user advocacy at every step on the development process throughout the project. As a component of user acceptance testing we engage sponsored user groups and these are groups, that are really derivative from the personas that we had constructed at the beginning of the project, and we rely on these sponsored user groups to help give us honest feedback as they go through the solution and test it out truly within the users footsteps.

Now, with every key strategy here that I’m going to be discussing today, I list some key questions that I, as a solution designer, am consistently asking myself within the given theme. On the theme of empathizing and identifying with end users what I’m constantly asking myself is A, what is my users situation and emotional state throughout the interaction and truly this is with a beginning, middle, and end. When a user is first opening up our solution, coming into it before even beginning with a kind of a back and forth interaction, what are they dealing with? Are they coming from a place of frustration or distress or confusion, say if our use case is about, say, submitting a first notice of loss insurance claim and we have a user who is reporting property damage, right, this is a serious subject matter and users come in with a lot of stress. On the hand, maybe we’re building a solution for a use case where users are coming in from a place of curiosity or interest, and these kinds of considerations should really drive the tone of voice that we use in approaching the conversation and how we directly relate to our end-users.

So, here we see just some examples of our design thinking approach and the tools we use to understand our end-users and empathize with them. On the left is an empathy mapping diagram where we take our persona and have distinct quadrants for what a persona might be saying, doing, thinking, and feeling in the current state, uncovering those pain points. In the middle we have a formulated user persona visual where here we take those pain points that we’ve been covered from empathy mapping and start translating them into proposed solutions where we can really generate those who, what, wow project hills and create a better future with the use of conversational AI for our personas. And to the right we see a sample user journey where we map out the
various steps it takes a user to accomplish their task today as well as what it would take for the user to accomplish their task tomorrow with the help of the solution looking for where those gaps are, where those pain points are and how we can help alleviate them and streamline the process for our users.

Onto strategy number two. We want to engage our users quickly and efficiently. Of course, we want to be very conversational and organic in the way that we approach our conversations but it is key to remember that at the very core of a conversational AI solution we are, in fact that, we are a solution and people use it as a tool to achieve what they want. Nine times out of 10 when an end user is coming to interact with conversational AI it’s not because they’re bored but it’s because they have something that they want to do, something that they want to accomplish, and in today’s digital age we definitely have a slim window to demonstrate the value that we can bring to our end users before they might move on and look for something else.

So how do we do this, how are we efficient and quick in the way we engage? First of all, it is extremely important to illustrate our capabilities of the solution upfront at the very beginning of the interaction, before it requires a single step or a single stroke of the keypad from a user it should be clear to the user what they can expect and what this tool can accomplish for them. This really provides two key benefits – so, first of all, we show that straightforward path to value add for the user as they can see hey I’m an
end-user, I’m entering this conversation with an AI solution, I need to accomplish X and when that solution shows me upfront hey I can do X, Y, and Z for you I as a user and that much more likely to engage because I see that it can directly accomplish what I’m looking for. In addition to that, it just sets the conversation up for success because we establish clear expectations from the user, we can see that the solution is emphasizing that it can accomplish X, Y, and Z so I’m less likely to ask about A or B but really try to stay on track with how we can create a successful transaction.

Next, we always want to look at the text length within our messages and just the way we approach conversations and keep things quick and direct wherever possible. It’s very common today, in the market, you can see a lot of solutions out there where companies might be looking to essentially rejigger some of the content that they have on their websites within a conversational AI solution and it just doesn’t work very well. If you think about how we message each other, our friends, our family it’s always quick and to the point, no one’s interested in reading essays with in a conversational AI solution and we definitely need to keep that in mind as we make sure that all content that we provide is truly conversational in its nature. A key question that I ask myself around this theme, as I’m building and testing my solutions, is, is every click really necessary and by this I mean, of course, we want to make sure that we minimize the number of steps it takes the user to achieve what they’re looking for, but it’s also about demonstrating that. Every step along the way should be demonstrably bringing a user closer to resolution for their task.

Next up, strategy number three: we want to balance flexibility and guidance. So, today in the market we see a wide variety of solutions that really span the whole spectrum here. Some solutions out there are very guided and they offer finite options at every single step and are quite restricting in the interaction. Whereas other solutions out there are extremely open-ended, relying very heavily on NLU and free text and take more of a question/answer, question-answer kind of structure. In truth there’s no real golden rule, there’s no universal right or wrong, and this balance is very dependent on the particular use case we’re trying to solve for and the particular solution that we’re building, but there are definitely some key components that we want to keep in mind as we’re looking to strike that balance to truly build an effective conversation. Something to keep in mind, with very flexible conversations that are reliant on NLU and free text, is that they make a very organic and conversational feel which is, of course, what we’re after, but they are also, on the other hand, a little challenging to complete a transactional use case where at the end of the day we do need some concrete sets of information. And, of course, on the flip side, if you take a look at a conversational flow that might be heavily guided, those could be very effective at getting transactional use cases completed but we always have to watch out because they can feel a little bit formulaic in their approach – so, of course, there is that balance.

Another key theme within this topic of flexibility and guidance is the idea of cognitive load, which we want to keep in check. Cognitive load is referring to the idea that having more options and selections does not necessarily lead to happiness or success. You can think about this maybe as the paradox of choice, analysis paralysis, or one of my personal favorites, K.I.S.S or keep it simple stupid, but at the end of the day, what this really means is we want to keep the number of choices to the minimum necessary to maximize user engagement because as those selections become too rich and too plentiful users get possibly a little bit paralyzed and might not engage quite as much as we would like. So, key questions that I ask myself within this realm is, at every single step with in a conversational flow what is the logical next step? If I, as our user, of course, going through the conversation in the shoes of my user persona, what is the next step that I’m always looking, for what might I want to do next; and as a designer am I making that next step obvious and apparent. It should not be hidden, it should be very easy and accessible, we want to make solutions that are intuitive and this is the way that we do so. We want to lead our users with prompts and calls to actions that
are outcomes based and help show the user what we can do for them.

Next, onto strategy number four: utilize rich media effectively. Using imagery and visuals can absolutely make or break a conversational experience depending on how we use them. Of course, we have to think about conversational AI as more than just messaging, it’s not just text, and the fact that we have resources like images and videos available to us, at our fingertips, absolutely means that we should leverage them to enrich the experience. Still, on the other hand, we should be a little bit careful because if we get too exuberant with the way that we use some of these assets it could potentially overshadow the utility and value that we’re creating by assuming maybe a little bit too playful or informal. So, of course, this relates very heavily to some of our previous strategies especially number one with relating to our end users, understanding what that tone of voice ought to be as we engage them and help them through their situation.

Still, there are some best practices that we like to employ at Pypestream about how we leverage imagery for practical reasons that could still also help engagement. For example, we like to open our conversations at the very onset, usually with the graphic, maybe even an animation just to help ground the user and help concentrate them within the experience, of course, to engage them and it also gives us an opportunity to, maybe, express a little bit of our brand, as well as, maybe, even showcase some of the capabilities that the solution can do for the user, in a way, like we said in strategy number two we want to engage effectively, it’s a great way to show that rather than tell. In addition, when we talk about navigational actions and a more guided approach, we like to utilize universal icons for some of those persistent actions. For example, as we see here, a back button or a home icon for a main menu, using these icons are just universally understood by users and help make the experience as intuitive as it could possibly be.

So, key questions on this topic that I’m always asking myself as a conversational designer are A, if I’m a user interacting with solution would I get bored with it, which, of course, we don’t want and can be a symptom of having a conversation that’s too text heavy and too dry, but, on the other hand, if I’m interacting with the solution would I lose patience with it, which is also a risk if we’re overly exuberant with those images and we overshadow that value we want to showcase. So, as long as we use rich media effectively, for practical reasons and, of course, paying homage to some of our previous strategies, we can prevent both of these outcomes and truly make a nice, rich experience for end users.

Okay, so now we are going to switch gears a little bit and see some of these strategies in action. So, here we have our conversational AI interface for Alister Bank, a financial institution, and what do we see upfront here at the very beginning of the interaction? First of all, we have this opening graphic which is prevalent, it’s colorful, it’s lightly animated which, of course, just helps engage the user, help them concentrate within this widget, establish a little bit of the brand and tone we’re Alister Bank but we’re also taking slightly informal, more casual approach here to the way we’re engaging the user. Our opening text is very minimal and to the point: Hi. I’m Allie your Alister Bank virtual assistant, how can I help you today? We do have NLU and free text enabled but we also have some key options here up front and our main menu to help guide the user and help them understand what the solution is capable of doing. So, therefore, if I type in something like which credit card is right for me, we can hope that the conversation will be able to understand, via NLU, what the user is asking about. And, as we can see here, that is indeed the case, we’ve triggered our natural language understanding, the solution understands we’re looking for a credit card and how does it respond – another image, keep the conversation rich and then again brief and to the point messaging: Looking for a new credit card, happy to help. We are acknowledging what the user has said providing that dynamic contextually relevant response and guiding them along the path – Let me ask you a few questions to determine which rewards card is right for you. Again, straight and to the point, we’re injecting a little bit of emojis and imagery here to stay consistent with our tone of voice because we’re adopting a little bit of a more casual approach here, but still, we can see here that this is clearly about the user, this is the solution trying to find the right card for me, not the right card for Alister Bank. And, of course, as we can see with our button options here, using some of those persistent icons for ease and navigation. Okay sounds good, let’s enter into this flow. So, what do we see here? This is a carousel selection, again referring to how we utilize rich media, this is an asset type that we like to employ here at Pypestream, just a nice, rich way to combine imagery with headlines, maybe some additional text, a call to action, all within one rich selection kind of interface. So, I’m going to go select travel rewards and continue on with our flow here. Next, we have a straightforward drop-down selection. Like I said, we want to utilize our rich assets in an effective way, we don’t necessarily want to have that image rich carousel at every single question, so as to not exhaust the user and create fatigue, this particular question is a little bit more simple, so we keep it regimented and ordered, minimize the cognitive load and show, kind of, the minimum required amount of options here necessary to continue with the use case. So, let’s select three to five trips per year. Again, we have another carousel just a nice way to combine maybe logos with headlines or company names, or airline names in this case. We’ll select the Fly Falcon and we’re continuing on here and, again, talking about that notion of contextually relevant responses and demonstrably bringing the user closer to resolution, we want to keep the user on track and let them know that we’re making progress and we do that here just by suddenly saying, hey we’re on the last question, just so that we can encourage
the user to keep going for that one final step. So, I’ll say, actually nope never, I do not want to pay an annual fee for my credit card, and here we are, we’ve gotten to a milestone in this use case. Again, showing an image with a slight animation to just enrich that experience a little bit, as well as some of the imagery and emojis that were utilizing here within our messaging, and here at this junction, within the use case, this is a major decision point for the user where we can see a number of actions that I, as a user, might be willing to take. Some users might be at a point where they want to explore more information about this card, others might want to go ahead and apply immediately if it’s already looking great, and some might want to take a look at some of the other options and do some more comparison work, but like I said in a previous strategy we want to always look ahead to see what is the next logical step that I, as a user, might want to take and make sure that those are apparent and obvious and as easily accessible as possible.

We’re turning here to our key strategies. We’ll move on to strategy number five, all about making data-driven decisions. So, within conversational AI we’re really lucky to have a wide variety of data resources that are rich for us to look at, to help analyze and see how users are interacting with our solutions and how we can better design them and better optimize them. For instance, we can look at transcripts and free text analysis to take a look at how users are interacting directly with NLU, what kinds of questions users are asking to make sure we are answering them, as well as even looking at feedback surveys to see what users are saying when they are entering free text, providing feedback. In addition, user traversal data is extremely valuable to look at. So this is when we take a look at an entire conversational structure, from start to end through the different steps, and how users are interacting with every step, where are they going, where are they navigating and how are they getting from start to end. In addition, we definitely like to leverage user feedback surveys as a great way to get that additional feedback point from end users as they interact with our solutions and we can actually employ some of our previous strategies to inject those feedback points, and to inject those feedback points, and do so in a tactful way that does not inhibit the use case. For instance, just offering a quick survey at the end of a session after the user has already completed the use case or even asking for quick thumbs-up, thumbs-down to gauge the relevance of the responses we might be providing to users for their questions. So, some key questions that I’m asking myself as I’m analyzing user interaction data is A, to see if users are getting stuck at any particular flows and what might be causing that, as well as it looking for trends, particularly with negative feedback, if we see that a lot of users who are providing negative feedback might have something in common, maybe they all went to the same portion of the conversational flow, for instance, what is it about that portion that we can optimize and improve.

So, let’s just shift gears once again to Alister Bank what happens if I end the conversation here after I’ve gotten my credit card
recommendation? I’m immediately prompted with a quick end of session feedback survey and what do we see? First of all, quick, brief, to the point, essentially a one-liner message, two options, happy or sad, minimal cognitive load, universal icons, green happy, red sad and, of course, we can even inject a little bit of extra personality and wit into it as we look at those dollar signs for the happy face. So, of course, we’re going to say we are happy and there we go the session is fully over and that’s a quick way to get users to engage with a feedback survey and get us that much needed feedback to optimize our solutions. Okay and there we go, those are our five key strategies and now we’ll open it up if there are any questions.

Okay it seems like where there are no questions so far, so just to let everyone know this link will remain active so even after the recording is finished people will be able to provide questions.

Okay we have one question: Is there a way for the user to follow up and apply for that credit card? Absolutely there is. so we do leverage API integrations, like I said at the very beginning, as a way for our conversational AI solutions to integrate with other external systems. So, within, say the Alister Bank’s different web services, we would be able to actually engage in that conversational flow to apply for the credit card, gather whatever additional information we might need for the user and be able to
interact with Alister Bank’s systems via a API calls to actually establish that credit card application.

Okay, it looks like, oh, we have another question: What other types of transactions do you support? We can support any kind of transaction, really, as long as there is an available API service with whoever the client is that we’re working with, we can ingest any kind of API’s, whether they’re SOAP or restful or so forth. So, as long as that service exists that the external application can read or write to our application we can make any of those transactions come to life.

Okay, thank you everyone again. As I mentioned before if there are any additional questions this link will remain active, you also have my contact information here at the bottom if you want to send me any questions or follow-ups and thank you again for your time. Have a great afternoon.