How does an AI chatbot work?

Learn how an AI chatbot works by matching similar phrases, it doesn’t match to keywords.
James Donald
February 4, 2020

An artificially intelligent (AI) chatbot works by matching similar phrases, it doesn’t match to keywords although keywords are still important.  Yonder has built a framework of commonly asked questions in the tourism industry and each of these questions has been built with a massive set of questions real travellers have asked. 

For example, we've collected 100’s of phrases which all relate to someone asking about price, like "Prices please", “Are the prices per person”, “How much is it for 2 adults on Friday?”. 

Artificially intelligent bots take into account phrases and not just key words. A few example why:

  • Many words carry a different meaning when they're surrounded by certain words, let's take the word "pick-up"..... "What does the pick-up from the city cost?" or "Where do I pick-up my tickets from?"
  • The phrase might carry many keywords, such as pick-up and cost. "What does the pick-up from the city cost?"
  • The phrase may start with context and the true question is at the end, "I'm thinking of booking the trip for 4 people on 4th February, what does the pick-up from the city cost?"

Tenses also present their own challenges when relying on keywords, particularly when the person making enquiry is not skilled at English. For example the word book may mean different things in these scenarios "I'm looking to book", "I have a booking " "are there bookings available".  This means we need to look at the other words around it to understand what the person is asking.

How to build great answers

Since we're building question based on a set of similar phrases then responses need to be robust to a range of ways the question has been asked. That means responses need to be framed to respond to a question in the positive or negative tense which means we can’t answer “Yes” or “No”.  It’s also best practice to repeat the question in the response.  And it's good practice to add additional information to make it a comprehensive answer.

Here’s an example: 

Turn your chatbot into a sales assistant

A good chatbot doesn't just answer questions, it creates a conversation. Better still, we all know that the best assistant is a sales assistant, someone helping steer that customer towards bookings. This is how the Yonder chatbot has been designed. It’s great at keeping the conversation going, whether that be providing options to find more information (within the bot or on your website), directing users to book on your website, or leading them towards contacting staff.  This is when we add links.

We have designed Yonder around messaging protocols established by Facebook Messenger, a leading messaging app loved by billions of people around the world.  You can easily connect your chatbot to Facebook Messenger.  

  • Links are provided at the end of every response as ‘buttons’, not within messages.  This helps create conversation flow by using “calls to action”.
  • A maximum of three buttons per response.
  • A character limit of 20 characters per button.

Here’s an example from the Response Editor within Yonder, showing how buttons might appear, either linking to another response inside the bot (“Booking required”) or linking to your website (“Check availability”).

Does the Yonder chatbot learn ?

Just like your best staff member, the Yonder chatbot learns as it goes. As customers ask questions, Yonder adds this “question” to its data set, and over time, gets better at matching questions to things the chatbot knows. You don’t need to take any action to make Yonder “learn”. It does this by itself.

Adding new knowledge to the Yonder chatbot requires work by Yonder and the client. Let's take an example, you may see a theme of questions come through relating to "Do you have a family package deal". The chatbot needs to be set-up to give a nicely scripted response (provided by the client), and a range of phrases set-up so it can match future questions successfully (done by Yonder). Setting up just a few questions has a low likelihood of matching future questions, it's extremely rare for different people to ask a question with exactly the same words/phrase, for example the next person might ask "I've got a family of 4 with kids 2 and 6 years, are there any special prices for us?".

How hard is it to build an AI bot?

There are several ingredients to building an AI bot:

  • Lots and lots of data. This drives the performance of the Artificial Intelligence matching process. You need hundreds of phrases for every question to build effective matches.  This is why AI chatbots are typically in the realm of enterprise or large businesses with many customers. Yonder has aggregated data across lots of similar customers so they all benefit.
  • Skill set and time of people to train the chatbot with data. The 'learning' process of an AI bot is not done by a machine, rather the machine needs to be fed a set of 'training data' which is defined by people. Given the speciality of this domain such people are highly sought after and highly paid.
  • Journey design.  This is the process of designing a conversation flow that feels natural, it requires strong domain and customer knowledge to do well.  

It requires a specialised team to do an effective job building an AI bot that delivers great customer experience and outcomes. Yonder has taken care of that hard work for tourism businesses (and we're pretty good at what we do!), allowing you to focus on delivering great service and outcomes for customers.  

You might also like

Drive more online sales without more marketing spend