It’s no great leap to say the future will be filled with AI applications designed for conversation.
Whether we’re telling our car to play music or asking our phone where to eat, voice interfaces already play a starring role in day-to-day life.
Yet practical advice for designing these systems remains in short supply.
Earlier this year, we had the opportunity to build a virtual assistant with a voice interface.
In the end, we learned a great deal about the importance of the user’s environment, about conversational design, and maybe most importantly, about shaping the personality of an AI.
In this post, I’ll explain all of those points in more detail to provide more context about what’s required for conversational design.
People are busy. Our audience for this application — enterprise leaders — are even busier.
After interviewing stakeholders, we quickly realized that just building an enterprise virtual assistant — even one with outstanding conversation capabilities — wouldn’t be enough. If the end-users didn’t feel a connection to the AI, they would just accomplish their tasks in a manual way.
This brings up two key points about virtual assistants specifically and AI more generally.
First, people are still reticent to fully trust machines with complex, highly important tasks.
If a CEO needs to book a meeting with their department directors, they need to know that all the appropriate people are included in the meeting.
The cost of forgetting someone are high, because it costs a lot to assemble a team of well-paid business directors. This anxiety is particularly acute for our users, but it exists in one form or another for most of the general public.
Second, humans assign personality traits to AIs as soon as they interact with them.
This behavior has been documented as far back as 1966 during research at MIT.
More recently, Stanford professor Clifford Nass noted the tendency of humans to bestow human characteristics on computers in his paper, Computers Are Social Actors.
For example, Nass observed research participants were courteous to computers when it wasn’t necessary. Further, he noticed people correlated a computer’s actions to the machine’s own agency, rather than its programming.
In short, people have a natural tendency to treat computers — and AIs in particular — as something more than soulless machines.
So it doesn’t matter if you design an AI without a personality. Your users will just interpret it’s personality as boring.
For our enterprise virtual assistant to succeed, it needed a personality. And not just any personality. The AI had to behave in specific ways to evoke a particular set of emotions in our audience.
It might sound intimidating to design personality traits for an artificial intelligence, but here’s a secret: you probably already know how to do the research.
We didn’t do anything outside of our usual human-centered design process to find the answers. We spent time on contextual observation, created empathy maps and mental models, and took the time to understand the end-users.
In the course of our research, we discovered users needed to feel three primary emotions to connect with the virtual assistant:
To evoke these emotions in our audience, we had to decide what type of personality traits would work best. Like all design decisions, we tested the results with our stakeholders and users, but this part took some intuition.
The final result was more nuanced than just three personality traits.
But for the sake of clarity, here are the three most important personality traits we designed for the application:
A lot of advice on conversational design focuses on copywriting, but actions can be even more powerful.
To create a sense of consideration in the AI, we designed the application to not ask for the same information over and over. On a technical level, we designed conversation trees that could adjust to new information mid-request.
So if the request followed a totally separate direction than the AI initially thought, it could adapt based on new parameters.
By not asking for redundant information and by having the ability to shift parameters several levels into a conversation, the virtual assistant gave the impression that it was considering the needs of the user rather than just performing a function.
For this project, our primary personas were business consultants and enterprise CEOs. These are people who expect their colleagues to be proactive.
To create an AI that seemed forward-thinking, we programmed the AI to make inferences based on additional information. So if the user wanted to schedule a meeting at a time when they already had a commitment, the assistant would recommend a different time that worked for all parties.
All the user would need to do is confirm the time and date, and the assistant would set it. Again, this is less an example of copywriting and more an example of actions driving a personality.
When the AI finds mistakes and makes connections based on other information, the user will feel assured that the assistant can be trusted with important tasks.
Of the three personality traits, this is the one that relied most heavily on copywriting.
To make design a respectful assistant, we paid careful attention to error states. If a request was impossible or information was incorrect, the assistant would offer a correction, but never imply a judgement.
After strenuous testing, we were able to confirm that these three traits were able to inspire the emotional response we needed from end-users.
Ultimately, designing the personality of a voice interface comes down to understanding what’s appropriate for the users’ environment. Instead of visual design elements, it’s the tone, words, and actions of the AI that will persuade people to make the application a part of their daily lives.
Voice interfaces and conversational design are more popular than ever.
But if designers want to create experiences that become a part of people’s lives, they have to focus build that’s both intuitive and personable.