Meet Amy, Your
Personal AI Assistant
X.AI wants its chatty bot to schedule your day
for you
--
Dennis Mortensen was 10 minutes late to our phone interview. His two daughters slowed him down that day, messing up his schedule. Given that he’s busy founding a software company, those delays are costly.
Mortensen has a scheduling assistant, Amy, so I ask him if Amy could help him manage the kids.
“Not just yet,” he says, though he laughs as he suggests that one day Amy might reserve Bathroom 1 or Bathroom 2 for him in the mornings. “I’m not sure my wife is going to buy that, when she gets that email.”
His assistant isn’t actually a person. Amy is an email bot. Amy (or Andrew — your choice!) just tries to do one thing: schedule meetings. After agreeing to chat, Mortensen copied Amy into our email conversation, asking Amy to set up a time.
A few minutes later, it followed up.
Hi Caleb,
Happy to get something on Dennis’s calendar.
Does Monday, Mar 16 at 10:00 AM PDT work? Alternatively, Dennis is available Monday, Mar 16 at 2:00 PM PDT or Tuesday, Mar 17 at 11:00 AM.
What’s the best number for Dennis to call?
Amy
So now I was emailing with an AI.
At this point I had a choice. Being someone who likes to test systems, I could try to get Amy to fail a Turing Test and write back with something I believed only a person would understand — even if it sounded weird. “Perf,” perhaps, or “Whoo-hoo, 10 in the miz-orning dances with my docket!” or “The temporal coordinate you describe with a heavy typeface integrates in the most optimized manner with my sequence of obligations.”
Or I could just schedule the interview.
The degree to which scheduling meetings totally sucks needs no introduction. Few parts of professional life are as clunky. For well over half a century artificial intelligence has been heralded as the panacea for our woes; after all, Isaac Asimov penned the 3 laws of robotics in 1942. With intelligent systems now being deployed to, among other things, count the number of cars on a bridge or write news articles, computers seem to have attained sufficient power and learning ability to finally force real change.
But asking a computer to maintain a complicated human interaction demands far more sophistication. Ten years ago as soon as you heard the automated voice of customer service you’d start howling “Operator!” Now those systems are almost — almost — bearable. On many company websites you can instant-message with a virtual agent about your issues. But if you get stuck, you can usually find a path back to a human. That path is important, because no AI can handle all the small quirks of human conversation and the nuanced needs of every customer. This is why many developers and upstarts zero in on a specific problem, or what Mortensen calls “verticalized AI.”
Apple’s Siri, Microsoft’s Cortana and “Ok Google” will answer verbal questions and try to anticipate a user’s needs. But these systems connect one person with the internet — a tricky enough problem already. They don’t mediate an interaction between two people. Even casual scheduling interactions between humans can get ambiguous. If you’re discussing “next Monday” on a Friday, the date in question is unclear. As far as I can tell there hasn’t been a successful piece of software that mediates a human interaction as common but complicated as setting up meetings. Plenty of scheduling apps exist, but none of them strive to beat a task-specific Turing Test.
Trusting a machine to get the nuances of human interactions right takes a large leap of faith. Amy can’t make mistakes, especially if X.AI is going to charge customers for the service. In business a misunderstanding can seem catastrophic— a botched meeting with a big potential client, say. X.AI hopes to prove that at its core scheduling a meeting is just an exchange of data — a problem AI should be able to handle with enough teaching.“This is one of those services where ‘almost good enough’ doesn’t cut it,” says Mortensen, a C-level veteran of analytics firms. “We need to be at 100 percent.”
Including its four founders, X.AI now consists of 35 employees, with all but two of them building the product. One of Mortensen’s co-founders, Alex Poon, had been engineering vice president at Visual Revenue, where Mortensen was CEO. The company was acquired by Outbrain in March 2013. Mortensen and Poon started their virtual assistant project in October of that year and did research by acting as each other’s personal assistants and pretending to be Amy. They’d cc each other into an email thread at the “Yeah, let’s meet/hangout/get coffee” phase and, behind the scenes, they catalogued the details they needed to schedule appropriately.
For in-person meetings, for example, Mortensen and Poon realized they’d have to handle such details as where the person might be traveling from and what kinds of venues were appropriate for a meeting, such as whether both individuals drink alcohol. Phone calls required knowledge of the person’s time zone and location (to handle international calls) and the quirks of both parties’ schedules, in case one person, say, only takes calls on Thursdays. All these conditions went into the design of Amy.
Amy will also have to understand social graces. Amy can’t schedule a CEO to meet a salesperson at his cubicle. But that CEO might be happy to cab across town for an investor. And what about when two CEOs meet — does Amy evaluate whose company generates more recurring revenue? Mortensen hopes to open Amy and Andrew to the public sometime in 2015, but handling the nuances of human pecking orders will be one of the company’s biggest challenges.
The AIs of the world usually become adept at a circumscribed task by training on huge real-world data sets. Want a computer to recognize smiling? Have it analyze all the pictures of faces on the Internet. No easily accessible resource exists for the world of meeting scheduling. Gmail, Hotmail, Apple and Outlook servers contain such information, but the owners have little incentive to share. So after Mortensen and Poon set up the initial framework for Amy, X.AI is now teaching it by slowly opening up access to more and more humans. Mortensen didn’t share how many people are using Amy today, but he said it schedules tens of thousands of meetings each month. With every new person it learns a little more about how to process the patterns of habits and syntax, and funnels these data into Amy/Andrew’s overall understanding of scheduling.
When I signed up, I synced my calendars with Amy and gave her the three alternate emails I might use. Then it gathered my meeting preferences via an open-ended request — “You can also email me if there is anything else I should know to get started; e.g. you take most meetings at your office or hate meetings before 10 AM.” I had a feeling I hadn’t covered all important details when I answered, though. Then Amy just waits in the cyber-ether for the next time I’m scheduling a meeting. When one arises I simply copy Amy into the thread and the AI takes it from there — though Amy does forward me correspondences to make sure I’m comfortable. Such a thoughtful new hire.
As I communicated with Amy to set up my meeting with Mortensen, I found that I was actually nervous about what I wrote. Should I make a human sentence? “Hi Amy, Monday at 10 AM is best for me.” Or since Amy is actually a machine should I get straight to the data? “Yes: 10 AM Monday.” But that feels rude! As chats with AIs grow more common, I wondered if we are going to start homogenizing our speech out of fear that the computer will get it wrong.
“It was one of those things that I was afraid of,” Mortensen says. “We’re trying very hard to not introduce syntax.” (Sometimes Amy responds with smiley emojis, which seemed to trigger the Uncanny Valley for me.)
Behind the scenes the machine is slicing the email into simple data — timestamps, names, etc — and sifting through the text body of the email to figure out, among other things, who is involved in the scheduling and what social nuances might come to bear. X.AI even disaggregates the email signature into seven possible parts. “There’s a lot of juicy stuff,” Mortensen says of the signature.
Amy assigns a confidence rating to each email it’s going to send — meaning the percentage chance that Amy correctly understood the questions and content and is replying with a logical next step. If the AI falls below a certain threshold, an alert goes out to the X.AI team for help. But not all issues are ones of syntax. In a mock attempt at scheduling a meeting with my brother, he tripped up Amy by asking about safe places to lock up his bike. Amy forwarded me his question.
“It doesn’t look like it’s a message I can provide an answer to so I suggest you follow up with Brett directly,” Amy said. So, now I’m back in the meeting-scheduling loop. X.AI wants to automate these little, customer-specific details, too. But what if bike parking is always full by 9:30 am or the meeting is in a rough part of town, meaning it’s not a good idea to lock up at night? To fulfill its potential, X.AI intends to teach Amy complexities beyond yes and no.
But for arguments sake, let’s say Amy does learn to schedule meetings for the correct “next Monday.” And it can figure out when a CEO will take a cab across town, or know to scan Google Street View for places to lock a bike. Assistants take on plenty of other tasks for their bosses. Maybe Amy starts booking flights, hotels and rental cars, distributes the powerpoint slides of this week’s board presentation, and orders everyone’s favorite sandwiches from the caterer, all without being asked.
It’s easy to get carried away. Maybe Amy gets promoted. Amy is now a master of the back and forth, weighing all the options and preferences from either party. Maybe, I suggest to Mortensen, it gets good enough to negotiate business deals. A contract is just a set of optimal data points, right?
Dennis puts the brakes on. “I try not to think about it,” he says of the promises of AI. “We’ve been disappointed so many times.” Right now X.AI is focusing on scheduling meetings. That’s it. “Once we’re good at that, then I’ll move on to the next idea.”
