Chatbots have been in the news a lot this year. First there was Tay, Microsoft’s racist love child, who just couldn’t keep herself under control. Then came Facebook with its Messenger bots, which, by most accounts, didn’t start well. And now there’s news coming from Viv, which at first glance seems like Siri 2.0. Not surprisingly, it was built by the founder of Siri.
What’s really surprising is that, despite all this hoopla, it’s hard to find a single chatbot that’s actually a really good product. Of course, we can wrangle over the definition of what makes a good product, but in its simplest terms, a great product would have three traits: (1) It’s simple and easy to use; (2) It works well 99 percent of the time; and (3) It removes or reduces friction in whatever it was you want it to do.
So back to the question — why do chatbots suck? I mean, there’s not a single chatbot I can think of that has won wide praise, or, more importantly, that has proven that chatbots are actually easier to use than an app. So why is this? Let’s take a closer look.
Chatbots that try to do too much usually fail
Viv and Siri both suffer from this; they aim to cover everything, and in the process compromise on quality. To understand why a broad scope is problematic, we need to first understand how chatbots work.
Chatbots basically have two parts to them: the “brain” and the “body.”
The first part is the “brain.” To understand what this brain is, think of a car. Inside a car you have an engine that makes the car move. And attached to that engine you have a body that you can customise to your needs — you could have a four-wheeler, a sporty roadster or just a normal sedan. The same engine could be used in many different ways depending on the body you attach to it.
That’s kind of how chatbots work, too. The “brain” is like this giant nervous system that takes words that humans say and converts them into actionable code. In a very narrow-scope chatbot, there’s only so much that a human can say.
For example, if someone built a chatbot to help you book flights, you can imagine about 100-200 things a human could possibly say to the chatbot. So, it’s quite easy to build a brain for this need. I can do it in a few weeks and be pretty sure that I’ll get 99 percent accuracy in responding to the human. But now imagine a chatbot whose scope is much wider — like Siri. You could say literally anything to it.
For Siri to correctly understand everything is incredibly difficult. It requires an immense engineering effort, and takes time. If it takes a human baby years to learn a language, it will surely take humans years to artificially build one. And then too, they will fail, because they can never quite replicate the irrational aspects of human behavior. Just like you can never predict what a six-year-old will say. But you can predict what a 40-year-old will say.
The second part of chatbots is the “body.” As I said, this is the “easy” part of it. It needs to be built and customized, but it’s literally just a vast body of knowledge. Compiling it can take time, depending on the industry. For something like travel bookings via a chatbot, the body of knowledge you need is not immense. I imagine you could literally plug your chatbot into a TripAdvisor API and deliver a pretty robust product.
It’s hard to find a single chatbot that’s actually a really good product.
But for anything in financial services, which is the world in which I operate, the barriers and the time it takes to compile this body of knowledge is immense. Most financial services websites aren’t accustomed to talking like a real human. Chatbots, by nature, need to talk like humans. They need to be simple and easy to understand. Translating from banking lingo/websites to chatbot-friendly lingo is time-consuming and hard. Hard because financial products are by default complex, and describing them in simple, layman’s terms is therefore quite hard.
It’s really hard to build a truly intelligent chatbot
As strange as it may sound, most chatbots aren’t actually intelligent. Well, at least not the ones that are used. How do you know if a chatbot is intelligent? Here are two defining traits: (1) It is able to “take care of itself” and get better with time; and (2) It lets you say or ask literally anything you want to.
Let’s take Tay — the Microsoft chatbot — as an example. Yes, it was intelligent. It tried to learn as people spoke to it. And, in theory, it should have gotten better with time — better in that it starts seeing patterns without having to be “told” about them. What about an unintelligent app? Most Facebook Messenger bots, for example the CNN chatbot, are not intelligent. Why? Because you can only go down one of the many pre-set paths on them.
The apps that are most likely to get it right — to actually do what they’re supposed to do without screwing up — are the unintelligent ones, because the number of “paths” you can go down is finite.
Chatbots look cool in demos, but the UI is actually quite cumbersome
This is an interesting one. Most chatbots you see being demo-ed, for example Viv, completely gloss over this minute but important detail. The thing is, with most good apps, you actually type very little. It’s swipe swipe, touch touch, and you’re done. With most chatbots, this is not the case.
Chatbots are great for explaining stuff because they can keep asking you if you get it.
Most chatbots don’t support voice yet, which means you literally have to type everything. Not surprisingly, this means more time to complete the task — so it’s actually less efficient to use a chatbot than to use an app! Which is totally counterintuitive. And when you factor in the delays caused by the chatbot not understanding what you are communicating, the efficiency becomes ever worse — because now you have to re-type your question and hope the chatbot gets it.
So are chatbots doomed to fail?
Not at all. But it does suggest that it’s still very early days for chatbots. Ten to 15 years back, when apps started taking over the world, the benefits were crystal clear. It was just quicker and easier to use an app on your phone to do stuff than to try to use the mobile web.
Today, the same benefits — speed and efficiency — aren’t apparent when I use chatbots versus the mobile web. And even when they are, the lack of accuracy means I’m never quite sure if I can really trust the chatbot to get the job done.
Personally, I believe that un-intelligent chatbots will be the first excellent use of chatbots. Using them for customer service — playing back answers to FAQs to customers. And using them to actually explain how complex things work. Like bank products, for example. Imagine a chatbot that could help you — a first-time buyer — understand anything and everything you need to know about buying your first house. How valuable would that be? It would be amazing and immense.
Chatbots are great for explaining stuff because they can keep asking you if you get it, and if you don’t they can keep going deeper and deeper into explanations. Chatbots are also, I think, a great substitute for long forms. They’re much more human and palatable than forms. So there are great places to stratify if you’re an entrepreneur looking to build a chatbot.
A final note on Viv
Some people are a bit confused about what Viv is. To be clear, it’s not a Siri. It’s not a complete chatbot. It’s just a brain. Viv is a brain. It relies on input from third-party APIs to actually deliver content for you. What Viv is trying to do is become the default operating system or the network on which all other chatbots or chatbot APIs ride.
Viv, if it works, could become like the App Store. So in that sense, they will compete with Facebook Messenger or Slack or social networks that are also trying to become the default platform for chatbots.
Is this what the future looks like? Personally, I’m not sure. I’m not sure there is value in actually having a centralized chatbot operating system like Viv. If chatbots are the new apps, then wouldn’t we expect to find them in the app store, like we find our apps today? Maybe, maybe not. Only time will tell.