How do you train a chatbot?

Heike Gnaiger

Director, Travel Channels Customer Service Excellence, Amadeus

This content is only available in this language.

By now, you’ve probably heard of Amanda, one of our customer service team’s most well-known agents. On an average day she answers more than 1000 questions from travel agents, or about 120 conversations an hour!

But have you ever wondered how Amanda came to be? Or more specifically, how to train a chatbot like Amanda to understand an array of questions and know the answers?

Here’s a quick look at how our engineers trained Amanda to answer all of our travel agency queries in just three steps.

Step 1: Identify the “Intent”

Amanda’s mission in life is to understand what a travel agency customer wants. So, imagine for a moment that a travel agent asks her “How do I split a PNR?”.  Amanda’s first goal will be to figure out what the travel agent wants. Or, in chatbot terms, what is their “intent”? In this example, the travel agent’s intent is “to split”.

Our chatbot trainers are constantly adding “intents” into Amanda’s system. These intents are based on frequently asked questions identified by our help desk agents; frequently asked questions from the Amadeus Service Hub (our customer service portal); and popular search terms from online “how to” articles.

Step 2: Identify the “Entity”

Once, Amanda has identified that our travel agent wants to “split” something. What now? The next thing she will look for is the “entity”, or what the user is referring to. This could be a location, a physical thing, a relationship, an object and more. In our example it is the “PNR” that the travel agent wants to “split”.

Our engineers make sure Amanda has many entities to choose from by adding all sorts of potential locations, things, relationships and objects, as well as their synonyms, into the database.

Step 3: Dialog

Now Amanda knows the travel agents wants to split a PNR. With this in mind, she will start building a decision tree. If the intent is A, and entity #1 applies, then Amanda knows the answer will be X. If not, Amanda goes to entity #2 and so on. She also calculates a confidence factor. Amanda will go through all the entities until she finds the correct response.

In addition, Amanda is also able to find similar questions in her database and share an answer when the confidence level is high.


Our chatbot Amanda can understand and provide answers in the correct grammar and syntax of 13 languages, which will be rolled out over time with 6 languages already available during 2019! So, she´s pretty smart.

However, nobody´s perfect, not even a robot. If the confidence factor is too low or if Amanda is not able to answer, she will refer the travel agent to Amadeus Service Hub, the one-stop shop for all knowledge and tools regarding Amadeus solutions...

Our trainers are constantly helping Amanda to learn and improve from one day to the next by correcting any wrong answers, or those that our customers rated as “not helpful”.

Next time you have a nagging question you need to figure out, give her a try. You can find her on the bottom right hand corner of Amadeus Selling Platform Connect. Chances are high that she might just know exactly what you need.

Are you interested to find out more about Amanda and what some of our travel agency customers have to say about her? Check out this blog post here.


Travel sellers, Live Travel Space, Chatbots