Conversational Interfaces — Beyond the Hype

What makes a good conversational app?

Published in

Chatbots Magazine

11 min readMay 3, 2016

The buzz around bots is incredible —there’s a new article on the topic every day. On the one hand, it’s great because A) smart people are sharing their thoughts and experiences and we are collectively building insights, UX patterns, and so on B) it drives more smart, interested people to the field, creating a virtuous cycle.

On the other hand, however, I feel like the media mixes two trends, the rise of messaging platforms, and AI and natural language processing (NLP). Obviously, the two trends are linked, but they are different. It seems to me that the rise of messaging is a fundamental shift, while NLP is at best a nice-to-have add-on, at worst a distraction.

— — — — —

Hi! I’m Matthieu, maker of Smooz, the Slack-to-Slack communication channel that helps agencies work with their clients. This article is the expanded version of a talk I had the opportunity to give at Le Wagon Paris.

— — — — —

Why Do We Care?

The Rise of Messaging Platforms

In order to avoid confusing the fire with the smoke, we can first recognize that we are excited by messaging apps because of the incredible rise of messaging platforms. As developers, we get salivating by the prospect to reach millions, or even billions, of users. Hey look, a new playground!

We’ll get into some details later, but basically any platform with such high usage, additional context like (depending on the platform) friends, business colleagues, location, likes, purchase history as well as features like payment, is attractive.

What Makes a Conversational App?

As messaging platforms open to developers, applications for which interaction takes place in conversation threads can be developed.

I think the name of “bots” is misleading and too narrow. It’s an anthropomorphic bias to consider you’ll have to chat with the app to get something done. There are plenty of other interaction patterns — buttons, clicks, commands, images, and so on. Bots and chat-like interaction is only a subset. So how about we call the general category “conversational apps” ?

In any case, conversational apps are interesting from a UX point of view by how constrained the UI is. The jump from visual-based UX to text-based UX will be at least as disconcerting as the change from desktop web to mobile apps.

Concrete promises

Opening messaging platforms to apps promises benefits to all key actors.

Users

Users will have less clutter. We, as users, download very few new apps and use even fewer of the ones we’ve installed.

Less clutter: Since starting a chat, or adding a bot to a chat room, is all it takes to “install” an app, we’ll save on smartphone storage and homescreen clutter.
Central place: Not having to navigate between apps will reduce both the cognitive load for our brains, and CPU load for our phones. We’ll be able to do more, from the familiar environment where we already spend our time.
Convenience: Messages are asynchronous. We are used to have some unread messages that we check later. We know that replies may not arrive immediately. This will allow us to handle several threads of conversations and apps in parallel, in a way not possible with current apps. This will also make notification bots quite convenient.

Marketers

Obviously, new, not-yet-crowded app stores is attractive. But it’s not going to last long. Some features of messaging platforms will remain, though

Place: Messaging takes the largest share of our smartphone usage time. As a marketer, you want to be present in the environment your users spend their time in.
Re-engagement may become easier. A notification message can put an old conversation back on top — even if platforms will obviously set up boundaries to protect their users from spam. The conversational apps that provide value to their users with infrequent usage will be less likely to fall into the oblivion current long-tail mobile apps experience.
Group/social discovery will be an incredible growth tool. Now we just need Facebook Messenger to open that feature.

Developers

Conversational apps present some challenges to developers — there is very little in terms of established UX patterns, for example — but they are full of promises. Some of these promises are, by the way, very reminiscent of Web apps development.

Native: Your sitting on top of platforms that have many engineers working on iOS, Android and Web clients. Your app works natively across all these environments.
Easy to port: the different platforms will have different specs, but nothing as drastically different as different programing languages for iOS and Android.
API: You can code in your favorite backend language.

The flip side of the incredibly constrained UI are also additional benefits:

Cheaper to develop (less designer time) and faster to iterate (no app update to download).
Messages are very suitable for low-bandwith countries or low-connectivity situations.

What Works, What Doesn’t Work

In this section I try to present some thoughts about several fields. However this is a very new field, in constant evolution, so please disagree in the comments! For example, I thought this format was not suited for news, and instead I found myself using the Techcrunch bot a lot.

Conversational Commerce

I don’t think we have, collectively, a clear idea of what conversational commerce will look like. WeChat, of course, is already a very active commerce platform. I don’t pretend to perfectly understand what’s going on there, but it seems that shops on WeChat cater to very specific segments, with limited product ranges and great interactions with their audience.

Featured CBM: China, Wechat, and the Origins of Chatbots

My gut feeling here, is that conversational commerce will not turn to out to be an “Amazon in chat”, but more like Instagram, with influencers and trend-setters. Focused value propositions around advices and expertise may also emerge.

11 Examples of Conversational Commerce

Customer Service

Customer service is probably where the lowest-hanging fruits are for chat platforms, and Facebook is clearly focusing on it. Chat is the perfect blend of immediacy (you get your answer much quicker than, for example, through emailing support) and asynchronicity (you don’t have to wait 20min on the phone). It is also enriched by having context (your ID, conversation history, purchase receipts) so that you don’t have to repeat yourself everytime.

For more on this, check out: Can Chatbots Help Reduce Customer Service Costs By 30%?

Business Apps

Business apps, typically on Slack, are a slightly different breed. Useful pro apps will fluidify team workflow or take advantage of the whole team being present on one platform.

I wrote an article specifically on this topic if you are interested: what makes a good Slack app.

For more: Secret Invasion of Business Bots

Long Tail

The combination of lower development cost and lower threshold to install and use, should be an opportunity for “long tail” apps. A local shop would never build an iOS app to let customers know the opening hours or collect feedback, it would be too expensive and nobody would download that. On the other hand, it looks feasible with a bot that would appear as a Messenger conversation.

Bots: the Media Burns What They Worshiped

The current hype around bots is fueled by the confluence of the trends around messaging platforms growth — which we just discussed — and progresses in AI and NLP. The issue is that a semi-sentient app makes a more grabbing headline, so the media focused on the latter. The hype bubble, however, burst quickly when Facebook Messenger released with great publicity several bots for which the interaction was not great.

My opinion is that the NLP-powered-chatting-bot-with-a-personality has great potential but is, for now, mostly a distraction from well-designed and useful messaging-based apps.

The Order-a-Pizza Fallacy

Many providers of bot frameworks, NLP-as-a-service engines etc… use the example of purchasing a pizza as an example of how conversational interfaces could replace standard interactions. However this is a rather poor example, because it’s typically a case where it would be more cumbersome to type your request instead of clicking on buttons, looking at a list of images, and clicking on the one you want. This “Order-a-pizza” fallacy is based on the wrong premise that interactions in conversational apps need to happen through chat. Indeed if you hold this assumption, you need a powerful NLP engine to make the interaction less insufferable.

One way to solve that problem is to recognize that interactions don’t need to be limited to chat. Most platforms provides action buttons, images, etc… and WeChat even offers functional Web views. These interfaces allow the user to shorten its path to purchase completion. Suddenly, the task of the UX designer is no longer to prepare a witty, smart bot to chat with, but to focus on the simplest way for the user to achieve its goal.

NLP As a Nice-to-Have Feature

The corollary of the hype around bot and the Order-a-pizza fallacy, is that a project of conversational app needs to involve a developer fluent in NLP libraries, or use a NLP SaaS provider. This is a risky path towards a techno-centric approach, that ships late and frustrate users.

The main issue with NLP is that even if it gets 80% of the sentences right, it will fail in the remaining 20% in a very stupid, non-human way. Even if your error messages are witty and fun, the users are frustrated.

I often advise people who are considering starting with NLP in their conversational projects, to pause and re-consider. Here is what Guillermo Gette, maker of the popular /todo Slack app, has to say about this topic:

I was putting so much effort on making it “conversational” and actually a few helps and hints was much easier and effective than spending hours on getting the models right.

There are very, very few exceptions for which NLP is absolutely central to the value proposition of project. Most projects fall into one of these two categories:

The value is in human expertise, recommendations and so on. In that case, start with real humans, doing 100% of the work. They may or may not pretend they are bots, by the way (I think that’s a lousy idea, but people can disagree). Introduce NLP later to automatize the 10% most frequent questions, or just the onboarding. Then later, much later, when you have a trove of dialog data, you can move up the automatization %.
The value is in an automated task. In that case, start with a command-line style of rigid syntax. You can have nice, helpful reaction messages to guide the user when they enter an incorrect format. We humans are OK with robots being dumb and requiring robot-like syntax. That’s different from an NLP service pretending to be a human and failing at it — that’s frustrating. If you start with rigid syntax, then later you can introduce NLP to be more flexible in the sentences you accept, or handling the errors more gracefully.

An example of NLP abuse is the Poncho bot that shipped with the Facebook Messenger platform. This Medium article by Greg Leuch who worked on Poncho, leaves me the impression he misidentified the source of the problem. To me, the issue with Poncho was not the quality of the NLP or the wittiness of dialog, it was that the focus on NLP and dialog totally distracted the product from doing its core job, letting the user know the weather. I’ve just given the service another try — it’s still lacking as a weather info service — but at least now it supports the ‘help’ command.

Conclusion

Some Good Practices

The field of conversational interfaces is very early, and we are only starting to see a few good practices and UX patterns emerge. In particular, the Slack API platform team has started to share some good advices, which should be applicable on all platforms. Generally speaking, developers and designers should pay attention to:

Providing guidance and examples of what to ask, including from the initial welcome message. Letting the users wonder what to ask is the worst idea possible.
Supporting basic commands that “lost” users will default to: “help”, at least, and perhaps “menu”, “stop”, “cancel”, “options” and so on.
Proactively suggest actions that are likely to follow-up after the current interaction.

— — — — —

Thanks for reading! If you have any feedback, thanks for sharing it in the comments. And if you liked the article, remember to click the green ❤ below so that others can discover it ;)

— — — — —