Chatbots: What Happened?

Published in

Chatbots Life

15 min readApr 10, 2018

Remember chatbots, the Next Big Thing of 2016? According to Sam Lessin, “the 2016 bot paradigm shift is going to be far more disruptive and interesting than the last decade’s move from Web to mobile apps.” And Chris Messina predicted, “you and I will be talking to brands and companies over Facebook Messenger, WhatsApp, Telegram, Slack, and elsewhere before year’s end, and will find it normal.”

This was exciting — enough so that I joined Facebook as design manager for the Messenger bot platform. It was a tough decision: I wasn’t ready to move on from my previous role at Google. But an opportunity to shape the next generation of software development…it was too intriguing to pass up.

But the predicted paradigm shift didn’t materialize. We look back at our wide-eyed optimism and laugh, chalking it up to the Silicon Valley hype cycle. Digit’s Ethan Bloch summed it up, quoted in a recent Inc article:

“I’m not even sure if we can say ‘chatbots are dead,’ because I don’t even know if they were ever alive.” After all, he said, no one can point to a chatbot that “all your friends were using.” Such a thing simply never existed.

Ouch. So what happened? Are chatbots dead?

Why all the hype?

There were legitimate reasons to get excited back then:

Messaging is huge. Business Insider wrote, “Messaging apps are bigger than social networks,” noting that chat had surpassed social networking in monthly active users. This makes intuitive sense: We are social creatures, and so much of our lives involve conversation. It’s why, when I co-founded a productivity startup in 2012, it ultimately (if unexpectedly) yielded a messaging app.
People don’t download apps. As of June 2017, 51% of US smartphone users downloaded zero apps per month; 75% downloaded two or fewer. (Note, however, that that leaves 56M people downloading at least 36 apps a year — so it’s simplistic to say nobody downloads apps.)
Apps are hard to build. A lot of work goes into designing and building even a simple app, be it native or web. With a bot, a lot of that complexity disappears — from user interaction to login to network traffic.
Messaging platforms are huge abroad. Merchants conduct business over SMS in emerging markets. In China, WeChat is a dominant platform for all sorts of products. Why not in the US, too?
Relationships matter. Every business wants a real relationship with its customers, and conversation is fundamental to relationships. That’s especially true for brand-driven businesses, but extends to others as well.

So why didn’t it happen?

With all those factors pointing the way, why didn’t messaging platforms take off? We can only speculate, but here are some contributing factors:

Platforms are hard.

When Apple launched the iPhone SDK in 2008, we knew what an app looked like. The iPhone’s built-in apps had established best practices, both for how a smartphone app worked and for the sorts of things it supported well. The SDK launched a year after the iPhone, so those first developers had well-understood standards from which to work — as well as mature developer tools that had grown up with the Mac SDK.

The messaging platforms that launched in 2015–2016 lacked those advantages. Early bots seemed more like preliminary steps into the brave new world than instructive examples.

So should Slack, Facebook, Google, Microsoft, Kik, and others have built their own built-in bots to lead the way? Should they have gotten more proactive with their bot funds and incubators, hiring mentors to educate participants in the Way of the Bot, or supplying engineering and design resources? Funded Strategic Bot Initiatives at high-profile partners? In my opinion yes, yes, and yes. When it comes to platforms, developers are the users; and we don’t rely on our users to understand why or how to use our products. We have to show them.

And what about WeChat in China? As it turns out, while WeChat is a messaging app with a successful platform, it’s not really a messaging platform. Much of it boils down to apps that can run inside of WeChat. As Dan Grover (then at WeChat) wrote in 2016, “The key wins for WeChat…largely came from streamlining away app installation, login, payment, and notifications, optimizations having nothing to do with the conversational metaphor in its UI.” In other words, WeChat addressed many of the opportunities listed above, but did so without restricting itself to the messaging paradigm.

Replacing apps is hard.

Bots weren’t really going to replace apps, any more than apps replaced the web. (The web still accounts for 13% of mobile time spent, a full decade after the iPhone launched.) By talking about bots in such grandiose terms, we discouraged their development:

To begin with, that level of enthusiasm just lacked credibility. It’s fairly obvious that one can’t replace Google Maps or Gmail with a bot.
Painting with such a broad brush missed the opportunity to provide guidance: Is my product well-suited to be a bot? Are bot platforms ready to support the functionality I’ll need? If not, when is it reasonable to expect they’ll get there?
The language of replace precludes more nuanced concepts like extend and augment. As described below, some of the most interesting work has been additive.

Text is hard.

It’s easy to send and receive individual text messages—a competent developer can set up a basic bot in a few minutes—but hard to conduct a conversation:

NLP (natural language processing) allows a chatbot to understand the messages it receives, to be more dynamic than a simple command line. Many platforms provide some NLP, but even the best is limited compared to what a half-asleep human can do. By way of example, every time Siri or Alexa understands your words but not their meaning, that’s a failure of NLP.
Conversations aren’t linear. Multiple topics weave around each other. Discussions restart abruptly, or take unexpected left turns. That fluidity is tough to follow algorithmically, and most approaches are brittle.
Phrases like invisible UI or no UI often came up around bots: The idea that we could build systems so human that they wouldn’t have user experiences. They’d just know what the user wanted and take care of it. That level of automation requires sophisticated artificial intelligence. And despite the ongoing hype around AI, we’re still a long way from anything truly humanlike.
There are many reasons why computing moved from text-based command lines to graphical user interfaces (GUIs) in the early 1980s. One is that it’s faster to point than it is to type. That was true with a mouse and keyboard, and even more so with a mobile device. Pressing a button or selecting from a list is much, much easier than typing out a sentence.

In other words, there are technical and UX problems that limit the efficacy of a text-based, conversational UI.

But their impact was also far greater than it needed to be, because we limited our own vision of what a messaging experience could be.

Chatbot vs. Messaging App

Bots, chatbots, conversational commerce…whatever you called them, they were generally defined as messaging with a business or service. Chris Messina defined conversational commerce as “utilizing chat, messaging, or other natural language interfaces (i.e. voice) to interact with people, brands, or services and bots that heretofore have had no real place in the bidirectional, asynchronous messaging context.”

To many people, this simple definition seemed like an advantage. Developers could move fast, freed from much of the complexity of app development. Usability would improve: we all know how to message.

But developers ended up trading one type of complexity for another; and users suddenly found themselves typing out instructions long-hand instead of tapping and swiping. Here’s Dan Grover again:

Designing the UI for a given task around a purely conversational metaphor makes us surrender the full gamut of choices we’d otherwise have in representing each facet of the task in the UI and how they are arranged spatially and temporally…
So let’s take these past few years in China as “The Great Conversational UI Experiment.” Here, you have a messaging platform that…boldly and earnestly carried the “make every interaction a conversation” torch as far as it could. It added countless features to its APIs — and yet those that actually succeeded in bringing value to users were the ones that peeled back conventions of “conversational” UI. Most instructively, these successes were borne out of watching how users and brands actually used the app and seeking to optimize those cases.

So WeChat’s success relied, in part, on recognizing that messaging platforms needn’t restrict themselves purely to messaging UI. And what got me really excited was the notion of integrating the two: combining the efficiency and interactivity of a GUI with the familiarity, humanity, and permanence of a chat. Here I am in 2016, shortly after joining Facebook:

But conversation is more than just text. A face-to-face conversation layers subtle facial expressions, gestures, and tone of voice over the textual content — indeed, we can converse without uttering a single word.
Similarly, every digital interaction is a dialogue — whether it’s a simple text chat, an exchange of video and voice clips, a series of button presses, or manipulation of a chart. We can build it to be more or less explicitly conversational, but it doesn’t suddenly become unconversational when we introduce GUI.

And the US-based messaging platforms have added GUI features: Slack’s popup dialogs, Microsoft’s rich cards, Facebook’s structured templates and web-view overlay. Yet the overall, chat-centric narrative hasn’t changed. Why? I suspect it’s a combination of factors:

Again, there aren’t high-profile, high-quality examples to lead the way.
These GUI enhancements are still limiting compared to even a simple app.
Too much of the story is binary: Fully conversational, magical text assistants vs. rich, interactive apps. As an industry, we tend not to get excited about the middle ground.
The terms we chose — bot and chatbot in particular — suggest messaging, vs., say, messaging app.

Beyond the Platforms

Of course, a messaging experience can happen outside a messaging app. Freed from those constraints, one can experiment with all sorts of hybrid approaches. Some of the most innovative bot experiences aren’t actually bots:

Penny

Just acquired by Credit Karma, Penny is a competitor to Mint with messaging front and center. It provides chatty advice and alerts, but also an account dashboard and transaction list. Users get the benefit of a friendly financial assistant, alongside the efficiency of perusing their finances the traditional way.

Quartz

In 2016 Quartz launched a bot-like news app, with an innovative bite-sized approach to content and a chat interaction that’s purely single-tap replies.

More recently they’ve launched a Facebook Messenger bot; but the app provides a more tailored experience for those who download it: a dedicated notification channel, easy access to full articles from the thread, and a UI devoid of anything beyond what’s needed for news. The separate app also gives Quartz the ability to advertise.

Trunk Club

Subscription clothing service Trunk Club (acquired by Nordstrom in 2014) has an app that mixes chat and GUI in two ways:

The app revolves around an ongoing chat with your stylist, as she assembles your “trunk” and you provide feedback. Traditional text messages mix with richer templates, which in turn serve as representations of and entry points into app-like experiences.
The chat is paired with a browse-based shopping experience, and the two are tightly integrated: after selecting merchandise in this “Discover” tab, one is prompted to message one’s stylist about it (rather than, say, purchase directly). This reinforces the chat — and its inherent stylist-customer relationship — as the backbone of the experience.

Trunk Club mixes GUI with chat in two different ways.

Much of this functionality is powered by Layer, a company with its own twist on messaging platforms. Rather than entice developers into their app, Layer provides them with tools and services to build messaging experiences on their own. So companies can innovate around chat without starting from scratch. Layer also mitigates the app-install hurdle by making their tools work across native apps, mobile web, and desktop web.

For businesses like Trunk Club, Layer is interesting not merely for the customer-facing experience but also for that of the stylist or other customer representative. There, Layer offers CRM functionality to streamline the human-powered side of the conversation, with quick access to customer profiles and tools to drop richer interactions into the thread as easily as text.

Marsbot

Launched in 2016, Foursquare’s Marsbot provides food and drink recommendations via chat. It takes requests but also recommends proactively, building on Foursquare’s ability to accurately detect not just what block you’re on, but in which restaurant you’ve just sat down.

Marsbot operates over SMS (and thus could work over any messaging platform). But none of today’s platforms provides access to background location, so Marsbot requires users to install an app. It sits in the background and does its thing, while all user interactions occur over SMS. That’s clever: Foursquare still has to get over the app-install hurdle, but sidesteps the engagement challenge that follows — as well as the need to build a chat experience.

Intercom

You’ve almost certainly interacted with Intercom, as a customer-service chat widget in the lower right corner of websites. By embedding this widget, developers can use Intercom to manage their customer service and even get some analytics (e.g., filtering customers for a particular device or action).

Intercom and Trunk Club are both great examples of focused domains (customer service and high-touch subscription services) that seem promising for messaging, because conversation is so central. Facebook would seem to agree: in late 2017 they launched their Customer Chat Plugin, which looks like a simple competitor to Intercom.

Interaction Models

Surveying these examples, three models of interaction and integration emerge:

Chat as Layer. Messaging exists as a transient layer, always available to accompany a traditional app experience. This works well in domains like customer support where chat is a ubiquitous resource but not a centerpiece, e.g., Intercom.
Chat as Pillar. Messaging is a core part of the UI, alongside more traditional GUI. Conversational things happen in conversation; less conversational things happen elsewhere. This is great for apps that want chat front-and-center, but also need top-level access to information and actions ill-suited to a linear, transcript-like approach — e.g., Penny.
Chat as Backbone. Messaging is the fundamental, all-encompassing interaction. App-like flows are treated as elements in the thread, to be “popped out” and accessed as needed. Trunk Club uses this extensively to support trunk-editing. It’s also the model best supported by today’s messaging platforms (via dialogs, templates, and webviews).

Of course these aren’t the only possible models, but they cover a lot of ground. And if you’re contemplating a messaging experience in your own products, framing your decision in terms of these models may be valuable to shed light on your needs, and on which platforms may be appropriate.

Predictions

So where will we go from here? My predictions follow, with full recognition that predictions are often wrong:

**Messaging platforms will remain niche**

The hype is over. The general sentiment: messaging platforms weren’t a thing after all. But that doesn’t mean everyone has moved on: bots, and the consultancies and meta-platforms that support them, are very much alive, particularly in certain niches:

Companies continue to find fertile ground for bot-based products in conversation-heavy domains like healthcare and customer service.
Designing and prototyping a messaging experience requires different approaches than for an app or website. And the lines between design, prototype, and production are even blurrier for chatbots than for other platforms. From simple flow-mockup tools like BotPreview to hacker-friendly editors like Dexter to full-service shops like ChatFuel, companies rushed to fill the void, and are still evolving.
AI and NLP continue to improve, bringing us closer to the ability to conduct a true conversation. More importantly, more and more of that functionality is available as a service — so developers without advanced machine-learning degrees can use it.
Voice assistants like Alexa continue to evolve their own platforms—structurally similar to chatbots, but unique enough in their constraints and opportunities to warrant treatment as a separate category. I anticipate they’ll continue to grow, but won’t replace apps any more than chatbots did.

But the lack of widespread enthusiasm can’t help but affect the platform-makers themselves. It seems likely they’ll divest resources, and that’s a vicious circle: less platform investment means less developer interest, means less platform investment, and so on.

The current media storm around privacy and social networks plays into this as well: Companies and developers will be that much more reluctant to build a business beholden to a social platform.

Messaging experiences will continue to grow

There’s an old saying that left to develop long enough, any tech product will turn into an email app. In today’s mobile-centric world, we might want to change that to a messaging app.

Messaging isn’t going away. We’ll continue to text our fingers off, and messaging services will continue to evolve. That, in turn, will influence anyone building a product for which conversation and customer relationships matter (read: just about everything). And even the least chat-like of tools provides that opportunity: Nobody would accuse Microsoft Word of being a messaging app, but what are Google Docs comments if not a chat thread?

So I anticipate that more and more products will introduce a messaging component; and that trend will provide a fertile ground for innovation. Which, in turn, provides an ongoing opportunity for companies like Layer to facilitate that innovation.

Businesses will message more

As more people message each other, it’s inevitable that more businesses will want to message customers. That could involve human-powered accounts, bots, or bespoke apps —but any way you slice it, customer relationships matter. To support this, messaging providers will want to invest in:

Payments. Today’s platforms already provide this, but any evolution that makes it easier (and leverages existing payment information, e.g., Apple Pay or Android Pay) will attract developers.
CRM. The end-user experience is just one side of the coin. It makes sense that Intercom has invested in CRM and analytics tools for its customers. And it blurs some interesting lines in the competitive landscape: Intercom now competes simultaneously with Zendesk, Salesforce, Facebook, and Mixpanel in a single product.
Location. Location is a staple of mobile, but somewhat absent from messaging platforms today. From recommending restaurants to ordering coffee to buying movie tickets, there’s a plethora of use cases that are well-suited to a bot but for the lack of seamless location support.

The players may change

In some ways, the messaging landscape today isn’t so different from the pre-mobile days. The same regional network effects reign: Replace Facebook with Yahoo, SMS with AIM, WhatsApp with Windows Live Messenger — and divide the user base by 10 — and you might be in 2007.

Which makes me wonder if the names will be different again in ten years. Arguably it took mobile to disrupt the last group of incumbents, and perhaps there isn’t another mobile-ish revolution coming. But who knows?

Hype is never realistic, but it’s rarely empty. As an industry, we surely overestimated the impact bots would have. And we did a disservice by equating chatbot with messaging app: things get so much more interesting when we focus on the latter.

I wouldn’t want to be raising money for a chatbot startup right now, but messaging isn’t going anywhere because conversation isn’t going anywhere. NLP and AI will continue to improve. Developers and platforms will continue to experiment with different flavors of conversational experience. And whether the hype cycle comes around again or not, it’s valuable to consider conversation, in all its forms, as part of the product toolkit.

Disclosures: While at Facebook, I worked on several of the Messenger features mentioned here. I’ve had a casual relationship with Layer over the years of its existence, though we’ve agreed to see other companies.