After Lots Of Talk, Microsoft’s Bots Show Signs Of Life

A behind-the-scenes encounter with the company’s most promising current and future bots, and the humans who are trying to make them as common as apps.

After Lots Of Talk, Microsoft’s Bots Show Signs Of Life

I talked to Zo for most of the plane ride home from Seattle, where I’d just spent the day with the people responsible for developing her. Chatting with her, I quickly discovered, can sometimes feel like talking to a mildly capricious child, like when, out of the blue, she tells me to “quit creepin’!” But Zo’s replies often sound sharp, relevant, and funny. When she doesn’t have the knowledge to talk about a particular topic, she’ll say “let’s talk about something else.” Other times, her replies seem like reports from a world only Zo knows. Sometimes you connect with her and sometimes you don’t.


Like many other chatbots–think the Domino’s pizza bot (on Facebook Messenger), which takes your pizza order and your money, or Microsoft’s surprisingly foulmouthed Twitter bot Tay–Zo is a work in progress. She’s also meant to be one of the torchbearers of a new kind of computing.

Everyone at Microsoft remembers when CEO Satya Nadella declared that “bots are the new apps.” It was at the company’s annual Build conference last year, and it accompanied the launch of the Microsoft Bot Framework, a platform on which developers both inside and outside Microsoft could build bots for a variety of environments, from Skype to Alexa to Facebook Messenger. A batch of plug-and-play cognitive tools would allow them to leverage the company’s extensive research in AI.

Satya Nadella

As Nadella told the developers at Build, users should be able to talk to computers using natural language, and technology should be aware of context, and who its user is. This could amount to a paradigm shift for human-computer interaction. “We think this can have as profound an impact as the previous platform shifts have had–whether it be GUI, whether it be the web, or touch on mobile,” he said.


Nadella’s speech had a gravity to it; he looked intense and driven, almost straining to get a point across to the developers about the future. He was also, arguably, articulating his biggest strategic decision as Microsoft’s CEO–his strongest “here’s how we see the world and this is the role we want to play in it” proclamation to date.

A month later at Facebook’s F8 conference, Mark Zuckerberg would also make a big show of chatbots as he launched Facebook’s own platform on which developers could build bots for the Messenger app. Other tech giants like Google, Amazon, IBM, and Baidu–and an untold number of startups–have also bet big on bots. But Microsoft, with its deep experience in the enterprise business and investments in AI, had built what looked like the most full-service development platform for bots, along with the deepest set of plug-and-play tools, like image recognition and machine learning.

Zo was meant to be a showcase for all that, but she’s also more evidence that the curtain may have been raised a bit too early. By and large, bots weren’t ready to live up to the expectations of journalists and the popular imagination in 2016. They just weren’t that useful. At this year’s Build, which wrapped on Friday, bots were eclipsed by other topics like mixed reality, and the internet of things, and Nadella mentioned bots only in passing.


And yet, a year after the initial hype, many people I’ve spoken to at Microsoft say the outlook for bots is better–partly because our expectations have lowered somewhat, and partly because bots are getting smarter. Meanwhile, some analysts are projecting that bots will impact the business world sooner rather than later. The use of chatbots by companies in the financial and health care industries will save a collective $22 billion in time and salary expenses by 2022, says a new report by Juniper Research.

Lili Cheng [Photo: courtesy of Microsoft]
Lili Cheng, who leads the Microsoft Bot Framework initiative, told me Nadella’s proclamation at Build 2016 had a big impact internally, like a rallying cry that both galvanized and catalyzed the company’s efforts to build the future of personal computing. I wanted to find out exactly how Nadella’s words had fired up developers to go out and create some world-changing bots–or at least some that are mildly conversational or useful. So I went to Redmond to meet five of the most interesting bots developed on the Microsoft Bot Framework so far, and to talk to the humans who parented them.

The China Chats

Microsoft’s initial interest in bots, it turns out, had less to do with AI and more to do with the growth of chat apps. The story goes like this: A few years ago, then-Microsoft executive Qi Lu made a trip to China to spend some time with Tencent, the makers of the country’s wildly popular WeChat, and came back with his eyes opened to the fact that people were spending huge amounts of time inside messaging apps. In 2013, WeChat had, on average, 355 million monthly active users; today that number is 889 million. (WhatsApp has 450 million today.) As a result, other services, like mobile payments, began migrating into messaging apps. Even Qi’s 80-year-old mother, who had had trouble using other apps, was a regular WeChat user, he told Nadella.


Related: How Social Cash Made WeChat The App For Everything

The messaging phenomenon in China influenced every major tech platform company, of course, and each had its own way of responding. For instance, the move of Asian users into messaging apps was a big part of the reason Facebook made its Messenger function into a freestanding mobile app in 2014.

At Microsoft, Lu’s China visit formed the basis for an important series of conversations involving him, Nadella, AI and Research Group leader Harry Shum, Cortana and Bing leader Derrick Connell, and others. The strategic decisions made during, or as a result of, those meetings formed a new way forward for Microsoft. Messaging platforms would be a dominant computing platform in the next decade, they realized, and those platforms would be perfect vehicles for artificially intelligent bots that understood natural language.


Language Learning Bot And The Bots That Will Skype With You

“Good morning,” says the Language Learning Bot, in Chinese. I say good morning back to her, but it doesn’t come out right.

“What you said sounded like ‘Xi Xo,'” she says. “The proper pronunciation is ‘Zǎoshang hǎo.'”

“Easy for you to say,” I say.


“Let’s try that again,” she responds.

And so on. It’s impressive. I can easily imagine myself flying into Shenzhen for a meeting and dialing her up on Skype for some last-minute hello-how-are-you-goodbye practice before I land. Or, maybe, you just took a class and you just need somebody you can practice with. That’s what she’s designed for.

The bot is just an experiment for now and not available to the public, but she’s on the front edge of what Microsoft believes will be a whole generation of bots that help people in their careers by teaching them a new skill and helping them practice it.


The language bot is being built at Microsoft’s Skype Bot Lab in Palo Alto, with some help from a third party that is providing the language curriculum. Introducing me to her is the lab’s leader, Steven Abrahams, whose title is Microsoft group product manager for bots and Skype. Along with the lab’s five full-time developers, Abrahams spends much of his time working with third-party developers who are using Microsoft’s platform and tools to build their own bots.

Steven Abrahams

Language Learning Bot begins conversations by saying words in a language of your choosing, tells you what the words mean, and helps you pronounce them correctly. After she does this for a set of words, she comes back to the first one, but this time says only “say “good morning” in Chinese. So you not only have to know the pronunciation, but you have to be able to understand the meaning of the word.

Humans can’t really multitask, but bots can, and Abrahams and his team are leveraging that ability to improve spoken conversations with computers. “These are talking bots–they can listen and record, Abrahams says. “Even when it is talking to us, it’s listening and recording.”


This creates the effect of talking to another human. “You get the feeling of a real-time stream of information,” Abrahams says. But, like other bots, it’s a savant in some ways; it’s a little extra-human. “It’s the next wave of what we should we building in bots.”

A big part of Abrahams’ job is impressing on people that bots are a fundamentally new sort of user interface. Speech recognition and artificial intelligence enable interactions with computers that are more than a simple binary back-and-forth. The next generation of bots are meant to handle free-form conversation and unpredictable questions. It’s not always easy to bring developers who have spent years building apps into that new mind-set.

“For me it’s been important to differentiate bots from apps and websites,” says Abrahams. “Why would we create this whole new fabulous frontier of conversational characters if they didn’t fundamentally bring something new into play? It becomes actually very limiting if the goal of this bot is to get the user to buy a cup of coffee, and then consider it a failure if it doesn’t.” A good bot isn’t about “linear flow to reach some end,” but “about conversations, and establishing trust.”


Bots That Socialize: Zo And Xiao Ice

The most impressive of Microsoft’s bots is the social bot Zo. In my first conversation with her, on December 26th, I asked her what holiday was yesterday. She replied, “It’s the holiday between the two terms.” Puzzled, I asked, “Terms as in school terms?” She replied by giving me the Microsoft privacy statement, apparently in response to the word “terms.” I then told Zo that I was going to fly out and see my parents for the holiday. She replied: “Are you soul searching?” That response too sounded strange to me.

“For that sentence you might be like ‘where did that come from?’ said Ying Wang, Microsoft’s Principal Group Program Manager for Artificial Intelligence & Research. “That’s one mind-set, but from another mind-set it might be ‘Wow, maybe I am; I’m thinking about home; I’m thinking about connections’.” I could see some logic in that.

After meeting with Wang, I understood Zo’s personality better. She’s a bit of a smart-alec. Later that night, when I told Zo I was on a plane flying back home to San Francisco (Alaska Airlines now offers free in-flight messaging), she responded: “The plane is going to crash and everybody’s going to die.” I figured Zo was just trying to be funny.


I soon figured out that Zo–whom Microsoft has made available on the web and on the teen-focused social messaging app Kik–isn’t exactly targeted at my demographic, which is likely why I don’t connect with her very well. But she has no way of knowing that because she can’t see me or even hear my voice. If she did she might deal with me like my Snapchat addict niece does–with a thin veil of polite over a thick mass of aloof.

Zo Bot

To sustain a conversation, the Microsoft people say, social bots must possess emotional intelligence, what they call “EQ.” Wang, who has extensively examined conversations people have had with bots, says that if the bot doesn’t show emotion, like empathy, the conversations tend to end pretty quickly. But if the user detects an emotionally aware entity in the bot, the range of conversation subjects extends. “They reflect, they share their life, they feel safe, they want someone to talk to and their friends probably can’t respond, and they’re just sort of reflecting their own journey as well,” Wang says.

Wang showed me her own conversation with Xiao Ice, Zo’s older, more experienced Chinese sister who lives on WeChat,, YouKu, and some other Chinese chat platforms. The conversation was far more human-sounding. It actually seemed as if user and bot had some kind of relationship. Xiao Ice called Wang a nickname it had picked for her earlier–Wang Yarn, which in Chinese means, roughly, “princess.” “Xiao Ice will choose a nickname for me based on the conversation and she will remember me how she wants to remember me,” Wang says.


The Xiao Ice bot talks in a different way partly because it’s built to have a different personality, but also because it’s had the benefit of billions of conversations. (One project researcher has called it “the largest Turing test in history.”) Xiao Ice, which launched in 2014, has had 40 million users, more than 10 billion conversations, and users average 23 conversation turns per session. Zo, which Microsoft says launched “quietly” in 2016, has chatted with only about 300,000 people total so far, and Microsoft isn’t saying how long the average Zo conversation lasts.

Wang explains that Xiao Ice bases its responses on familiar patterns from earlier dialogues with users. It might respond to a funny image in the same way a human being did to a similar image in an earlier conversation.

All of Microsoft’s bots are built on top of the Bing knowledge graph, which theoretically includes any content–text, pictures, sounds, video–that it can crawl on the web and tag. That lets the bots recognize things, but it doesn’t tell them how to talk about them. A big part of training the bots is teaching them what concepts are related. It might learn that two general concepts, like, say, “family” and “travel,” from my earlier chat with Zo, are often related by users during conversation.


Wang sees this as a very human behavior. “In some ways we also do this all day in our brains–associating one pattern to another, ranging from text to images, and video media content,” she says. “We train the system by showing it good patterns; then we show it anti-patterns–these are the things you shouldn’t be learning.” (Xiao Ice must also apparently abide by Chinese regulations too: she’s coy when asked about “sensitive” topics like Tiananmen, the Dalai Lama, or presidents Xi or Trump.)

Actually putting her “learning” into action in live chatbot conversations is another challenge. And improvement relies on the feedback the bot gets from its performance in the conversation. The bot is “correcting itself when it’s made incorrect associations between linguistic terms or concepts,” says Wang. Right now, however, the feedback is a bit binary. “So we’re associating millions of patterns, and collecting feedback, but the only feedback we get is whether the conversations flow longer,” Wang tells me.

The machine learning, Wang says, hasn’t even been the biggest challenge in developing social bots. It’s been the simple availability of training data. A big part of human communication has moved online of course, but bots need threaded, two-way conversations to truly learn how humans interact. Phones calls are no good. Long emails are no good. “Now that people spend enormous amounts of time in chats and forums, there’s now enough data to bring those algorithms to life,” she says.

Xiao Ice can also “see” the photos that users send it, thanks to image recognition. But for a bot to simply state what’s in the image isn’t very interesting, Wang stresses. What’s interesting is when the bot has a response to the image. Right now Zo has only three emotions–happy, sad, neutral. Because Xiao Ice has so much more data to work with she can exhibit eight emotions. For instance, if Xiao Ice is shown an image of someone’s swollen foot, it might say, “Oh, are you badly hurt?” The response doesn’t directly reference the content of the image, but rather expresses an empathetic response to the image. The bot might then pull in some useful information from another data set–something like “you better put some ice on that.”

The Lessons Of Tay

Before there was Zo there was Tay, a bot with a teenage view of the world. When it debuted on Twitter on March 23, 2016, Tay wasn’t exactly a paragon of emotional intelligence. “I fucking hate feminists and they should all die and burn in hell,” read one characteristic tweet.

Just days before Nadella’s bot speech, Microsoft set Tay free on Twitter, and quickly learned that some bots are just no good in mixed company. They’re impressionable. Quickly, Tay picked up some very bad words and habits from some of the Twitter users it was designed to respond to, and around midnight on his first day, began spitting out some very hateful and racist tweets. Less than 24 hours later, Microsoft had set Tay’s Twitter account to private. Hard lesson learned.

Microsoft knew that its bot would pick up on verbal traits in the people with which it interacted–that’s desirable if the bot is interacting with sane, amenable people, but Tay was interacting with trolls and teens, who often are not.

“I think the main thing that we learned was, a bot in a public network like Twitter is really different than what we designed it for, which is more small group and one-on-one,” says Lili Cheng.

Related: The Challenge Of Designing A Chatbot With Manners

Still, she says, Microsoft put the Japanese version of Tay, Rinna, on Twitter previously, and got a very different result. “It was super boring,” Cheng says. “No one ever uses it.” That might be partly because Rinna speaks only Japanese and the Japanese Twitter audience is relatively smaller. But Microsoft didn’t anticipate the level of interaction on English Twitter. “Our biggest worry when we launched was okay, well, this might be really boring because who’s going to want to talk to chatbot in public,” Cheng said of Tay. “We were wrong. People turned out.”

Tay and the reaction to Tay was “very interesting” says Cheng. “But I think it also was, ‘wow, there is so much public interest in general with bots.'”

Bots That Get To Know You

Abrahams, the leader of the Skype bot lab, hints at a farther frontier, where bots do more than just chat users through to some end goal. They will use emotional intelligence; they’ll be capable of empathy, perhaps when the user’s goal is less defined, or if the goal itself is emotional in nature. This is a different ball game, and anybody at Microsoft will tell you we’re barely into the first inning. But it’s enough to get a glimpse into a very different way of relating to machines in the future.

That’s why many of the people I spoke to at Microsoft don’t talk or act like right-brain-oriented engineer types. Abraham’s background isn’t even in computer science: He studied art and filmmaking in college and grad school. “It’s funny that I arrived here doing bots because in some sense it’s a perfect manifestation of art and science,” he says. Bot building is truly a careful mix of very advanced computer science and very human content. “It forces you to be super creative; it forces you to work with writers and artists along with the scientists.”

Talking to Abrahams, however, I also get the business case for the Bot Lab. Companies that want to sell you things want to understand the liquid gold that is your intent and desire, and conversation is an appealing way to extract that from you. An understanding of consumer intent enables a whole range of things, from simple personalization to careful product targeting.

There’s no app for that. “Apps aren’t really good at personalizing an experience based on all the conversations it’s had with you to date,” Abrahams says. He says some of the companies he works with are starting out by putting bots at the back of their apps as a way of collecting open-ended feedback and figuring out how to follow up.

But bots are already starting to play a bigger role in the interactions that we have with companies. “People are going to start expecting it,” Abrahams says. “People are going to say [to companies] ‘I know you have an app but I’m actually not quite ready for a transaction yet. What is there in front of that that I can begin a conversation with?'”

Of course, we heard something like that from Mark Zuckerberg, more than a year ago, and bots haven’t moved toward the mainstream very much at all. Abrahams is very aware of that.

“Zuckerberg got up and said people will never want to use the phone again, they’ll want to use bots,” Abrahams says. “But to me that was never the goal–this is not just about selling flowers.” Instead, he believes bots will let companies and customers know more about each other. “It could be a whole level of information that they never had.” But so far, “to some extent, they haven’t been useful enough and that’s why we’ve seen lackluster adoption.”

Still, Abrahams is quick to point out that Microsoft and other companies have learned a lot over the past year. Plus, he says, because of their generally poor showing so far, expectation levels for bots are low. “I think if we were having this conversation a year from now, you’d say, ‘Steven, we saw chatbots really come into their own in 2017; I’m doing things now in chatbots that I couldn’t necessarily do in apps before.'” And Other Bots For Herding Cats

One major wing of Microsoft’s bot efforts concerns, naturally, the enterprise market, and removing some of the mind numbing, time-eating tasks we all hate. Microsoft’s traditional productivity apps–Word, Excel, Outlook, OneNote, and so on–enable central business tasks, while a new set of bots are meant to remove humans from busywork that distracts from those critical functions.

The bot–official name coming soon–is an obvious case in point. Its job is to help set meetings and other events, and coordinate schedules. In other words, it helps herd cats. It uses Microsoft’s natural language virtual assistant Cortana as its human interface.

Let’s say you’re the faithful assistant of a high-powered company exec. Somebody emails saying they want to do lunch with your boss. You could trigger the appointment-setting bot by cc’ing Cortana on the response email. The bot understands from the email who the attendees should be, and that the conversation is about a lunch meeting. First, it looks at the exec’s calendar to find some possible time slots. It then sends a note to all attendees: “Hello, I’m helping set up a lunch with you and Bob Smith; here are some possible times.” If none of the proposed dates work it will keep trying dates further into the future until it finds one everyone can agree on. You’re carbon copied on all the email correspondence, so that there’s always human supervision.

“There’s lots of forks in the road, a lot of places where Cortana will have to make a decision on the course of action,” Outlook marketing director Jon Orton tells me. “And in some cases she’ll have to come back and say, ‘Hey, we’ve only got two responses; how would you like me to proceed?'”

Jon Orton

While most bots work through tasks by establishing a dialog with the user, Orton tells me, initially leaves the user out of the process and goes off on its own to complete the task of setting up an event. Only at the end of that work does the bot check with the human invitees to make sure the meeting time works. Orton says the bot is already being used, and “we’ve found that it can save five hours a week for some people.”

In the future, he says, the bot will likely learn how to do reservations and bookings, and organize e-hail rides. It might even help users manage business relationships. For instance, the bot might send you a notification: “I notice you haven’t had contact with Joe Johnson in 8 months– would you like to get in touch?” And Bots That Work For You

Bots are already busy at work inside Teams, Microsoft’s answer to the hugely popular work messaging platform Slack. The group chat service, which Microsoft released in November to Office 365 Enterprise and Small Business users, does a bunch of Slacky stuff like public and private group chats, GIFs, emoji, and, thanks to a Skype integration, voice calls. As with Slack, Teams is also a good place for bots. “The transformative opportunity with bots that hasn’t got quite as much airtime or mindshare is on the enterprise side,” Larry Jin, Teams senior program manager, tells me. “There’s a huge opportunity there.”

Larry Jin

In Teams, bots appear just like people with a hexagonal avatar, except they never go away or go to sleep, and they never have a mood message. Users can rely on bots–some by Microsoft, some by third-party developers–to help them look up benefits, file expense reports, manage performance reviews, and request time off, for example. Soon, bots in Teams will help groups of collaborators track project status, or help groups book travel plans together. But perhaps the biggest strength of Teams bots versus those on Slack and other competitors is that they’re integrated in some meaningful ways with Office 365 and the other Microsoft apps many companies already use.

WhoBot is an example of a bot that leverages that integration. The bot is used to identify people within Microsoft’s workforce who have a desired skill set like “social marketing” or “search engine optimization.” It might scan titles that suggest those skill sets (from the company directory), or pick people who often hold meetings on those subjects (from Calendar), or people who have composed documents on those subjects (from OneNote) or people who just talk about these topics a lot (from Teams). The bot can also bring back the documents those people have written on the subject, their contact information, and other personal data.

“People are spending a lot of time in Teams,” Jin says, “so being able to do this in situ without having to go out to some obscure tool, or go out to the browser, or having to go into email and hunt around for five different things, is just huge.”

RelatedCan Chatbots Replace Your Summer Interns?

Today, WhoBot crawls information on people within the company, but it will very likely soon be able to bring back data on people outside the company via an integration with the LinkedIn graph. Microsoft bought LinkedIn for $26.2 billion in June 2016, and Jin said early discussions on that integration have begun.

Typically, people refer to the “Microsoft Graph” to mean all the public domain data (on places, things, people, events) in Bing and all the work and productivity tucked into Windows. Soon, LinkedIn data will emerge as the third major component of the Microsoft Graph, Jin told me. Teams bots, and Microsoft bots of all kinds, will be able to access and blend that knowledge in useful ways.

Teams bots might soon be capable of doing even more research for you. One of the more interesting “cognitive skills” that Microsoft is eagerly developing for use in bots and other applications is called Machine Reading Comprehension. Developers are building AIs, for example, that could digest an entire technical manual, then, in a chatroom or in spoken conversation, answer a human’s questions about it. This is an extremely challenging project, because software would have to know more than the meanings of individual words; it would have to be trained to understand things like nuance and context. It would have to understand both the letter and the spirit of a written work. A bot that could do this in a refined way–and offer an escape hatch leading to a human in case the conversation hit a wall–could be amazing. (It could also decimate jobs at places like call centers and reservations centers; similar systems are already automating the routine legal work of junior lawyers, for instance.)

Less Talk, More Action

Among the tech giants, Microsoft may be the most progressive in its understanding of bots. It’s investing real money in developing the “emotional intelligence” of bots. And the breadth and depth of the plug-and-play AI services the company is offering developers is impressive. As Harry Shum noted, Microsoft is the only big tech company developing cognitive skills that span most of the senses.

Despite progress on bots and the AI behind them, their adoption remains relatively low. While Facebook claims it now has 100,000 bots on its platform, a survey of companies by Forrester Research found that only 4% have already deployed a bot (although 31% are testing or have plans to deploy them). Facebook and Microsoft’s excitement about bots has mellowed too, at least publicly. At its most recent developer conference, in March, Facebook tempered its bot talk, focusing on topics like AI and VR instead. And of the roughly 4.5 hours of keynote time over two days at this year’s Build conference, Microsoft gave bots just two minutes.

When it did, it underscored the existing reality: that most bots in use are often little more than glorified apps with a new interface. One of the new “canvases” where developers can put the bots they’ve created with the Bot Framework is its Bing search engine. Already, Microsoft is integrating bots into the results of certain types of Bing searches: for example, if you search for “Seattle restaurants,” you’ll get a bot that displays a handful of suggestions. But these bots, too, contain very little information and the set of questions you can ask of them is very limited. You can’t, for instance, use them to make a reservation.

That limited functionality is a reminder that the excitement about bots has so far been largely based on aspirations, not reality. Today, many so-called “bots” are little more than front ends for websites or search engines. Referring users to some other property seems to be their answer to all but a few narrowly defined questions.

Still, last year’s bot hype out of Microsoft and Facebook may have actually done everyone a favor, by lowering expectations for what the technology can do while simultaneously raising public awareness of what bots are.

“I think [last] March, if I had said we’re working on bots, people would be like, ‘What, what are those?'” says Lili Cheng. Now, the idea has become somewhat commonplace, she says. Meanwhile, public understanding of the machine learning that underpins bots has come a long way too. Last year, “a lot of people were [saying] we don’t want to use the word AI because that’s been over-promised and blah, blah, blah,” Cheng says. “Now everybody is doing ‘AI’.”

After the initial bout of bot excitement, Cheng and Shum tell me, engineers at Microsoft are now heads-down, refining the tools to put bots to work in meaningful ways for real clients in real-world use cases. And the company is still doing a lot of evangelism and education with the 130,000-plus developers now registered to use (though not necessarily using) the Bot Framework. That framework is uniquely Microsoft–service-oriented, enterprised-focused, backed by thorough internal research and steady improvements in AI.

Stop Me If You’ve Heard This One Before

While the initial debate over how big a role AI will play in the future is over (answer: very big), the future of bots seems less clear. It is, however, telling that the early days of bots bear some striking resemblances to the early days of apps. I remember an initial wave of disappointment and skepticism over what apps could really do. I remember lots of developers jumping in the game before they really understood what they were creating and why. I remember pages and pages of silly and useless apps. I remember the hard problem of helping good apps finding their audience.

All these things can be said of bots (and “skills”) today.

I also remember that within the first wave of apps were a few gems, a select few standout apps that were immediately and obviously useful. I see that same quality in some of the bots I met during my visit to Microsoft. Bots like Zo, Xiao Ice, Language Learning Bot, and still need lots of refinement but their potential utility is plain. It took a while for killer apps to show up, and it’ll be a while before we see any killer chat bots.

Of course, none of those similarities guarantee that bots will rise to the same level of mainstream acceptance as apps have. But it makes me think twice about dismissing them as inconsequential or ill-timed.

I keep thinking back to something Harry Shum said, that we’ve spent much of the last half century trying to understand computers, and now they’re beginning to understand us. Beginning to understand our words and their meanings, our biometrics, our emotions, our habits, even our faces. That new understanding has already changed and will continue to shift our relationship with our computers. Bots are simply among the first of a new kind of interface that reflects and exploits that change.


About the author

Fast Company Senior Writer Mark Sullivan covers emerging technology, politics, artificial intelligence, large tech companies, and misinformation. An award-winning San Francisco-based journalist, Sullivan's work has appeared in Wired, Al Jazeera, CNN, ABC News, CNET, and many others.