4 Strategies for Building Bots that Don’t Suck

4 Strategies for Building Bots that Don’t Suck

We’ve all had them: bad chatbot experiences. But, what can you do as a company to avoid them? At Sinch’s Conversational AI Fest, Ivan Westerhof, Ivan Westerhof, COO at Campfire AI, shared four strategies for building bots that don’t suck. 

What’s the most frustrating chatbot experience you’ve ever had? A bot that repeatedly told you how it didn’t understand you? A bot that gave you answers that didn’t have anything to do with your question? Or a chatbot that used technical terms without being able to explain them to you?

Unfortunately, we’ve all had way too many bad bot experiences. The good news is: you can avoid all of them!

At the Conversational AI Fest, Ivan Westerhof, COO at the Conversational AI Implementation Partner, Campfire AI, shared four lessons on building a bot that doesn’t suck.

In this article, you will read:

Building engaging bots: from hype to quality

In 2019, there was a huge buzz around chatbots. Big jumps in conversation design and artificial intelligence (AI) led to a new generation of high-quality conversational AI chatbots that were delivering more competent and more user-friendly experiences.

A few years later, these chatbots have become a standard solution for many companies, and the technology has moved from its hype stage into a phase where designers and companies can focus more on improving the bot quality.

"Chatbots have left the hype phase. Now, we can focus on the quality of AI."
Ivan Westerhof
COO, Campfire AI

So, what does it take to build a good chatbot? At Sinch’s Conversational AI Fest, Ivan Westerhof shared his insights from building more than 50 bots on 5 continents. From his vast experience, Westerhof, COO at the Conversational AI Implementation Partner, Campfire AI, has learned that the following four aspects play a major role in building a successful bot:

  1. Don’t treat a bot project like an IT project.
  2. Get your user the most specific answer as quickly as possible.
  3. Expect unusual expressions.
  4. Remove the answer, “sorry, I don’t know”.

How can you build a successful bot? Learn from successul companies! 

Don't treat a bot project like an IT project

One of the biggest mistakes companies can make in launching a chatbot project, is to treat them like an IT project. While it seems to make sense at first—they are both tech projects, after all—they are inherently different.

IT projects, such as developing a new app, for example, are linear. You design the app, test it, and deploy it. The functionalities and the user design of app itself will then let the users know what they can and can’t do.

Chatbot projects, however, are more complex. That’s because a chatbot is an open-ended interface, and users could potentially ask your chatbot anything. Furthermore, unlike with an app, it’s impossible for users to know exactly what your AI bot can and cannot answer.

scope NLP chatbot
Think about what your bot can and cannot do. (Source: Campfire AI)

And that’s where some bad bot experiences start: a user asks a bot something that it can’t answer (out of scope) and is left with answers that aren’t helpful.

How can you solve this problem? Ivan Westerhof recommends not only thinking about what your bot can do, but also figuring out solutions for out-of-scope interactions.

Get users the most specific answer as quickly as possible

One of the main reasons users love chatbots is because they can answer their questions quickly. If this expectation isn’t met, users will perceive this as a bad bot experience.

Consider the following example. A user tells a bank chatbot that they have lost their Visa credit card. The bot recognizes the term “lost” and moves into the lost flow with the first question: “I am sorry that you have lost your card. Is this a debit or a credit card?”

This is not necessarily an incorrect response, but it’s unnecessary, as the user has already told the bot that their Visa was a credit card. The user is now asked to explain something that they have already said. From a user perspective, this is annoying and time-consuming.

In order to avoid this, and give users the most specific answer as quickly as possible, Westerhof recommends developing a knowledge hierarchy for your bot.

knowledge hierarchy buildin bots
Hierarchies assure that users get help faster. (Source: Campfire AI)

If a user already provides detailed information, a knowledge hierarchy allows the bot to skip some general steps in the flow and jump ahead to the following more specific question in the flow hierarchy.

This will save the user time, get an answer faster, and provide an overall better user experience.

Expect unusual expressions

As mentioned before, a chatbot project is not as linear as an IT project. It is, in fact, more like a zigzag. You teach your bot new expressions, you test them. Then you go back and implement the learnings for the test–and test again, and so on.

testing an AI chatbot
Building good bots requires testing, testing, and more testing. (Source: Campfire AI)

Testing and re-testing are at the core of building a high-quality bot. However, no matter how many expressions you teach your bot, there will always something you didn’t think of, says Ivan Westerhof.

The solution? More testing! Westerhof recommends three testing stages for chatbots:

  1. Testing internally
  2. Testing with the client
  3. Testing with customers in the real world

Westerhof’s tip: continue the testing until you reach a 90 percent (in-scope) recognition rate.

Remove the answer, "sorry, I don't know"!

High-quality AI chatbots can solve up to 80 percent of incoming customer queries. However, there will always be questions that are out of scope for the chatbot. In these cases, companies should have a set-up that is as helpful as possible for customers. And it shouldn’t include the answer: “sorry, I don’t know”, says Westerhof.

This answer implies that the bot can’t help the customer and leaves the users with more problems than they started with. For example, if the customer tells the bot that their Wi-Fi doesn’t work, and the bot says: “Sorry, I don’t understand”, the customer now has two issues, a Wi-Fi that doesn’t work, and a bot that doesn’t understand them.

In Westerhof’s experience, there are better ways to react. For example:

  • Chatbots can use the information the customer provides and pair it with existing knowledge to provide a similar answer.
  • Direct the users towards a place where they can find an answer. This could be a webpage or a human agent.

The most important part, however, is that bots shouldn’t apologize. First off, nobody should be sorry for what they don’t know, and second, it’s more helpful for the user to focus on a solution instead.

Read how the European Commission used a chatbot to answer 80% of customer queries. 

Building bots that don't suck: what to consider

Summing up, if you want to build a high-quality chatbot, it’s important to realize that a chatbot project is not the same as an IT project, to answer users as fast and as precisely as possible, to always expect the unexpected, and to focus on solutions rather than on problems.

That’s why it’s crucial to work with an experienced chatbot partner if you want to launch a chatbot that your customers will love. For Ivan Westerhof, Chatlayer by Sinch is such a partner.

“Chatlayer by Sinch is working hard to make AI experiences better.”
Ivan Westerhof
COO, Campfire AI

Ready to improve your chatbot?

More interesting chatbot insights