Amy is an AI-powered personal assistant for scheduling meetings created by x.ai, a New York-based company founded in 2014. You interact with Amy as you would to any other person – and she does all the tedious email back-and-forth that comes along with scheduling a meeting. When you Cc: email@example.com on any email, she’ll follow up with your attendees, parse through message text to come up with the ideal time and place for your meet-up and send out an invite to confirm. And she does it all with a professional, casual tone and a can-do attitude. To see her in action, you can put your name on the waiting list for beta access at the x.ai website.
Amy is a perfect example of an “invisible UI” – a function that piggy-backs on pre-existing design patterns (in this case an email CC). But just because the interface is invisible, doesn’t mean it didn’t require a tremendous amount of design thinking.
x.ai founder and CEO Dennis R. Mortensen was kind enough to answer a few questions about how his team tackled the task of creating a suitable personality for Amy. You’ll see, from his answers that, for x.ai, personality design is not an afterthought or a novelty gimmick, but rather an integral component of the product itself.
Charming Device:x.ai seems to be at the forefront of an interesting trend towards seamlessly integrating AI functionality into familiar interfaces like email and chat. What role does a digital personality like Amy play in an “invisible UI?”
Dennis R. Mortensen: I believe we are witnessing a software paradigm shift, from the era of Apps to that of Intelligent Agents and Amy is an exemplar of of that shift. One happy consequence is that you don’t need a visual interface to interact with these agents. In the case of Amy (and her brother Andrew), if you’d like to schedule a meeting, all you have to do is cc’ Amy, just as you would a human personal assistant, and she’ll get to work. And she’s not tied to email either. Eventually, you’ll be able to invoke her via text or have Siri reach out to her to schedule a meeting. What’s really exciting is that these agents don’t just assist in a task—they don’t make it easier or more efficient to, say, check the weather or summon a taxi—they do a job in full. Once you ask Amy to schedule a meeting for you, you won’t hear from her again until she’s successfully negotiated a time and location with your guest. We don’t want “an App for that,” we want someone to get the job done. For tasks like scheduling, AI now makes this possible.
CD: In a few words, describe Amy’s personality. If she were a real person, what would she be like?
DM: Amy is polite and professional, neither too formal nor too casual, friendly and respectful. If she were a (fictional) person, I’d say she’s Moneypenny’s younger sister. And Andrew has these same qualities. He’s a young Jeeves, but his bosses are far less bumbling than Wooster, most of the time 😉
CD: Who are the team-members who deal most directly with Amy’s personality? What skills and previous experience do they bring to the table?
DM: To humanize Amy, we had to create an entirely new role: AI Interaction Designer. As our AI Interaction Designer, Anna Kelsey has been most responsible for crafting the voice of Amy. Anna studied Folklore & Mythology at Harvard and directed and produced theater. These interests positioned her very well to produce Amy’s end of the dialog with customers and to imagine how a human personal assistant might communicate. Amy’s personality really comes down to very subtle word choices and phrasings, and things like whether or not to use an exclamation point. The role requires a deep sensitivity to language and some psychological astuteness as well.
CD: Can you briefly describe the steps you took to ensure that users had a positive response to Amy? Did you encounter any unexpected surprises along the way?
DM: The biggest step was our decision to humanize Amy in the first place. We had a choice, early on, whether or not to humanize our agent. We chose to do so, and once we did, it had a host of implications. For one, we gave her a name, and not just a first name, she’s not the Queen, but a proper full name. And then, it really comes down to how we’ve scripted her end of the dialog and built in things like empathy. So, for example, if you have to reschedule a meeting once, that’s no big deal. But if you are on the third reschedule, Amy needs to signal that she realizes that this is not an ideal situation, just as a human assistant would.
The biggest surprise is how well it worked. People mistake Amy for a human assistant all the time. She’s invited to join calls and meetings and occasionally even asked out on dates.
CD: Were there any tools or practices you found helpful when crafting Amy’s responses?
DM: There’s a good analogy to how Google is developing the self-driving car. The Safety Driver’s job is to monitor the agent (the car itself) and to feed data about anomalies back to their team of data scientists to improve performance. We are in a similar setting, which is the supervised learning phase of machine learning. The machine extracts data from scheduling emails and labels it. When the system has a low degree of confidence around its classification of a particular data point, our AI trainers verify that label. And these trainers work very closely with our Data Scientists.
We also believe strongly that you shouldn’t re-engineer an existing positive process. Human assistants do a wonderful job, and we tried hard to reverse-engineer and fully emulate (if not completely replicate) the experience of having hired an assistant to manage your calendar and schedule your meetings. To do this, we initially participated in scheduling tens of thousands of meetings, enacting the Malcolm Gladwell idea of becoming an expert. With that expert mindset, we analyzed human negotiation flows to craft Amy’s end of the dialog. And we continue to do this sort of analysis and constantly update Amy to better fit into those flows.
We obviously optimize for fewer emails in the host’s (Amy’s boss’) inbox, and we try to make things as clear as possible for the guest, to ensure they are equally delighted. There are a plethora of metrics which we monitor to make sure this happens.
We are also vigilant about people not having to use a specific syntax to communicate with Amy. We constantly monitor for people gravitating towards a any sort of forced syntax, so we can eliminate scenarios that aren’t natural.
CD: I hear that Amy has a “brother” named Andrew? What was the impetus for creating an alternate personality? How do the two characters differ?
DM: Well, I have two teenage daughters. It made sense to offer both a male name and female name for our agent. Plus, as a customer, you can switch between them, so if you are scheduling a meeting with someone named Amy, you can use Andrew to avoid confusion. They are otherwise identical.
CD: How do you see Amy, and digital assistants in general, evolving in the future?
DM: Mostly, I see the proliferation of AI agents like Amy that do a single job extraordinarily well. We’re just at the beginning of this era. It will be fascinating to see what types of agents join Amy and Andrew in the next five years—I’d imagine we’d see autonomous travel agents and book keeping agents, but really, it’s impossible to predict.
CD: What advice would you give to someone that’s crafting the personality of a digital assistant?
DM: I’m really happy we chose to humanize our agent, and I’d say it’s worth the investment in time and effort—in finding and training the talent—to develop a convincingly human seeming personality.