Chat bots were supposed to be the next breakthrough: what went wrong?

News publications spread like this: “In 2016, the transition to the paradigm of bots will bring far more far-reaching and interesting consequences than the transition from the web to applications that occurred in the past decade.”

Chat bots were on the queue in the trend.


Our hopes soared in transcendental heights. The industry was expecting a new era of innovation with glowing eyes and a tail pipe: the time has come to communicate with the machines. It seemed logical. All signs pointed to a mad success. Message exchange has reached the peak! Interactive marketing has become a new term! And WeChat! And China!

In addition, it became obvious that in the market for these stupid applications with their time-consuming development, supply greatly exceeds demand.

At Mobile World Congress 2017, chat bots covered all the front pages. The conference organizers referred to “universal agreement that in the future the attention of brands and corporations will inevitably shift to chat bots”. In fact, the only serious question was who exactly monopolized the market; The fact that chat bots will “fly up” was not in doubt:

“Will there be a platform that will dominate the ecosystem of chat bots and personal assistants?”

A year later, we received an answer to this question: no. Because there was not even the ecosystem itself in which to dominate.

Deceptive HYIP cycle

Chat bots are not the first technology that was extolled to the skies and which subsequently significantly sank. Ancient as the world, the HYIP cycle has retained its usual course here.

• Reverent articles on TechCrunch
• Looking to the future, opinion leaders like Chris Messina have their say
• In the Silicon Valley, everyone was drooling at the looming opportunity to chat with smart computers.
• Messenger flooded bots
• Slack went into rapid growth and even founded a special fund for the investment in the development of chat bots

Expectations grew and grew ... and then everything was somehow blown away at once. The paradigm shift that everyone had predicted never happened. And applications, which is typical, are still alive and well. Looking back at our choking optimism, we look at each other with some bewilderment: “Is that all? Was this the very promised revolution? ”

Ethan Bloch of Digit expressed the general sentiment as follows:

“I don’t even know if it is possible to talk about the death of chatbots, because I’m not sure if they ever were alive at all.”

According to Dave Feldman, vice director of the product design department at Heap, chat bots did not just take on a complex task and didn’t pull it out; they took on several difficult tasks at once and did not cope with any of them.

Text vs. Speech vs. GUI: A Brief History

Bots can interact with users in different ways. The main opposition here is text and speech. At the beginning (era of computer interfaces) was the (printed) word. Users had to drive teams into the machine manually in order to achieve anything from it.

Then came the graphical interfaces and saved them all. They charmed us with windows, mouse clicks and icons. Why, after a while it all even became colored!

Meanwhile, an army of researchers worked tirelessly to integrate natural language into interfaces so that people no longer had to learn the puzzling language of database queries. Another army developed speech recognition software so that you could just talk to your computer without having to type. It turned out to be much more complicated than it all seemed at first. A lot of accents, people who tarator or stretch words, gibberish, all sorts of "mmmm" and "eeeeee" ... How-how did you say, "recognize speech" or "just break the furnace"?

The next item on the plan was to maintain a two-way dialogue with the machine. Here is an example of a possible dialogue (originally from the 90s) with the VCR tuning system :

User: Good afternoon.
Agent: Good afternoon, what is your name?
User: Candy.
Agent: Hello Candy!
User: Let's adjust the clock.
Agent: What time is it?
User: It is 11 o'clock in the morning.
Agent sets the time: 11 hours.

Really cool? The system works with the user bilaterally and cleverly reads what is required of it. But she was carefully adapted to the conversations related to the work of the VCR, and functioned only within these well-defined frameworks.

Modern chat bots, both text and voice, face all the listed difficulties, and, moreover, should be easily scaled and retain their effectiveness on a wide range of platforms. In fact, we are still trying to achieve all the same innovations as thirty years ago. As it seems to me, the snag is as follows.

Contrasting applications and bots

There was a bold assumption that "the time of applications has passed," and they will be replaced by bots . By confronting these two significantly different concepts (instead of treating them as two separate entities created for different purposes), we ourselves created an obstacle for the development of bots.

Perhaps you still remember a similar battle cry that rang out ten years ago, when applications just began to enter the market. But do you remember how applications have supplanted the Internet?

It is said that a new product or technology must satisfy two of the three requirements: to be better, cheaper or faster. Are chat bots faster than apps? No - at least not at the moment. Is it possible to consider them “better” - a moot point, but, in my opinion, to date, the highest-quality chat bot is inferior to the highest-quality application.

In addition, no one seriously thinks that everything in Lyft is too complicated or that it is very difficult to order food or buy clothes through the app. What is really difficult is to try to perform these actions with the help of a bot, if the bot does not understand you.



Clear. Add more to your shopping cart?
add 4 apples
i.e. 4 bananas
Well, I added 4 apples and 4 bananas to the basket.

From the first-class bot is now as much confusion as the average application. If we talk about applications with rich, multi-layered, intricate functionality, they have nothing to oppose.

The reason lies in the fact that computers give us access to extensive and complex information systems, and the first systems with a graphical interface have become a revolutionary step in the development of these systems. Modern applications are built on the foundation of years of research and experimentation. Why give up on all this?

But if you put the word “add” instead of the word “replace”, then everything becomes much more interesting. Today, the most successful experience is given to those bots that adhere to a hybrid approach, embedding chat in a broader strategy covering traditional elements.



Penny talks to the user at ease, sending tips and notifications, but also provides a standard-looking dashboard and transaction list.

HubSpot Conversations combines Facebook Messenger, local chat, social platform, email and other messaging systems - everything comes in one box.

Layer gives developers the tools to create a personalized messaging experience on mobile and desktop web applications and in native solutions.

The next wave will be applications with several modes : with them you will be able to voice your request (as with Siri) and get an answer in different formats - in the form of a map, text or sounding speech.

Bots for bots

Do I need a bot in my product? Can existing platforms support its functionality? Will I have enough patience to create a bot that will cope with the tasks that I put in front of him? Another negative aspect of general enthusiasm is that, under its influence, such questions are usually bypassed.

For many companies, bots are simply not the right solution. Over the past two years there have been many cases where they were used at random where they are not needed at all. If you make a bot just “to be”, give it complete freedom of action and hope for the best, it’s definitely not going to end with anything good:



Chat bot Maroon 5 in action; sure he was badly needed

The overwhelming majority of bots use decision-based logic based on the assumption that the bot will issue one of the prepared responses based on the keywords it reveals in the entered text. The advantage of this approach is that it becomes quite simple to compile a complete list of scenarios for which the bot is designed. And this is also its main disadvantage.

This is because a bot is nothing more than a reflection of the abilities, diligence and patience of the person who created it, and how many user needs and possible replicas he managed to capture. Problems begin when reality refuses to fit into a given framework.


Please select an item from the list.
Show the main menu
Please select an item from the list.
Show previous versions
Please select an item from the list.
Back to all options
Please select an item from the list.
Dumb bot

According to the latest reports , 70% of the total number of bots on Facebook Messenger (more than 100,000) can not fulfill even the most simple user requests. This is partly due to the fact that the developers did not choose one narrow field of application for bots in order to focus on it and strengthen it in every way. We, working on the creation of GrowhBot , decided to divert it to the needs of sales and marketing instead of the “on all issues” model, although the temptation to succumb to excitement and use the potential opportunities to the full was great.

Do not forget: a bot that does something ONE, but well, at times more useful than a bot that performs many tasks, but somehow.

Low availability

Why do I need to vote?
I will clarify. Can I help with something else?
Why do I need to vote?
I'll look for an answer. Do you have any other questions?
Who is Hillary Clinton?
Looking for information. Tell you something else?

A competent developer can create the most primitive chat bot in minutes - but speaking of one that could keep up the conversation ... That's another story. No matter how much they raise the hype around AI, in fact, we are still far from something even remotely resembling human thinking.

In ideal reality, natural language processing technologies should allow the bot to understand the messages that are transmitted to it. But they are still in their infancy, they are just leaving the research laboratories. Some platforms provide processing in some volume, but even the best samples are still presented in a very rudimentary form (as an example, you can recall Siri, which understands individual words, but not the general meaning that is embedded in them).

Matt Ezey describes another related problem: they do not excite the developers desire to invest in them work and creative thinking:

“We had no chance to excite the interest of consumers until artificial intelligence had reached the human level. The interest of users will be attracted by those AI technologies that will make bots truly suitable for communication. ”

Conversations have a complex structure, they are not linear. Topics layered one on another, change direction sharply, suddenly end and re-emerge after a while. Modern conversational systems based on rulebooks are too fragile to cope with such unpredictability, and statistical approaches using machine learning also have serious limitations. The level of AI needed to imitate a conversation with a real person is now out of reach.

At the same time, there are already some examples of high-quality innovative bots that define the development path for the rest. As Dave Feldman noted:

“Did Slack, Facebook, Google, Microsoft, Kik and others have to create their own built-in bots that developers could target? Should they be more active in establishing foundations and incubators, hiring mentors who could pass on knowledge to participants who entered the Bots Way, providing engineering and design resources? Should they invest in Bots Strategic Initiatives and bring together researchers with well-known partners?

My answer is yes, yes and yes. If we talk about platforms, developers - in fact, the same users. And we do not provide users themselves to understand how and why to use our product. They need to demonstrate this. ”

Do not neglect graphical interfaces.

Once upon a time, communicating with computers was the only way to type cryptic commands in the terminal. Visual interfaces that use windows, icons or a mouse have become a real revolution in working with information. Computer science has moved from text-based to computer interfaces with good reason. If we talk about entering data, click faster than typing. Pressing a button or choosing from the list is clearly preferable to typing text manually, even with auto-completion (which often fails). If we talk about reading information, then one time to see, as a rule, is really better than a hundred times to read.

We like the visual presentation of information, because in principle we are very sharpened by visual perception. Children do not accidentally love touchscreens . The pioneers who came to the idea of ​​graphical interfaces were inspired by cognitive psychology and research on how our brain builds communication.

Interactive interfaces according to the idea should imitate the type of communication that is closest to people, but in reality they only increase the cognitive load. In fact, we are changing something very simple to a more complex alternative. Of course, we can express some ideas only in natural language (“show me all the options, how to get to the museum to collect 2000 steps, but at the same time meet 35 minutes”), but most of the tasks in the graphical user interface can be performed more efficiently and intuitiveness than using dialog.

People prefer to communicate with people.

The desire to add a human aspect to business communication is quite logical. What is so bad about sales and marketing is the sense of human presence: brands hide behind numbered requests, feedback windows, and emails marked “do not respond to this email”, automated distribution of answers and forms that send messages “ one way".

Facebook set its goal to create bots that would pass the so-called Turing test, that is, they would be indistinguishable from the person in the dialogue. But a bot is still not a person, and it never will be.

Communication involves much more than just words. People can read between terms, extract information from the context, and decipher multi-level messages — for example, read sarcasm. The bots quickly forget what they are talking about - as a result, one gets the feeling that you are trying to build a conversation with a person who suffers from serious violations of short-term memory.

The HubSpot team made an accurate comment about this:

“Bots provide a way to communicate with customers individually, which can be easily applied on a large scale. However, they are not able to give people that pleasant, productive experience, which brings the complex in its semantic structure of conversations with other people in messengers. "

People are not so easy to fool, so any attempt to give a bot to a person will have a bad effect on profits (not to mention the fact that it is generally undesirable to deceive users). Even the few bots that are built on the most advanced technologies, do not reach this bar.



What weather will be on the weekend?
Are you in the boat? And then I can not find information on this location.
What will the weather be like in the weekend in Brooklyn?
In Brooklyn, New York is now 46 degrees Fahrenheit, clearly.
And on the weekend?
Sorry
ON THE WEEKEND
Sorry, dozed off. What did you say?

And one moment. Dialog interfaces seek to imitate the patterns that people prefer to follow in communication ... with other people. But do they want to build interaction with machines in the same way? Is not a fact.

Ultimately, no witty remarks and imitation of human behavior will save the bot from communicative failures.

Where should we go next?

In some ways, the early followers were still right. People really require Google Home to put their favorite song on them, order pizza from the Domino bot and learn how to apply make-up under the guidance of the Sephora bot. However, from the point of view of consumer reaction and developer activity, chatbots did not live up to the enormous expectations that they placed on them in 2015-2016. The real picture is not even close.

Computers are good at performing their tasks: look for information, make calculations, analyze opinions and present all this data in a compact form. But computers are not capable of understanding human emotions. With the current state of processing technologies, they do not even understand what we want from, let alone what we feel.

Therefore, it is still impossible to imagine an effective user support without the participation of a person who would bring empathy and emotional intelligence. Today bots can only help us with monotonous, mechanistic, low-level tasks, perform the role of gears in a larger and more complex system. We hurt ourselves and them, demanding too much and not giving time.

But this is not the end of the story. Yes, our industry greatly overestimated the effect that bots will give initially. However, the key word here is “initially.” As Bill Gates once said:

“We always overestimate the scale of the changes that will occur in the next two years, and underestimate the scale of the changes that will occur in the next ten years. Do not let the lull deceive you, be active. "

The hype has passed, and it's for the best. Now we can move from the stage of feverish exaggerations and black-and-white thinking to a moderate approach and careful study of the gray zone.

I believe that now we are approaching the stage of rapid growth . For technologies that bring significant changes, such a shift in climax is quite typical . Instant messaging will continue to gain popularity . Chat bots are not going anywhere . AI and natural language processing technologies are evolving every day . Developers, applications and platforms will continue to experiment with and invest in interactive marketing. And personally, I can not wait to see what this all happens .