How to develop truly conversational bots
Have you ever had a pleasant conversation with a bot? Chances are high that you have, as chatbots are getting smarter and better at helping us deal with the products we’re dealing with everyday.
Barcelona startup Caravelo is at the moment developing six chatbots for some of the biggest airlines in the world. The chatbots they build help millions of travelers to book and reschedule flights, provide customer service, act as personal concierges and much more.
Even though all their bots use NLU (natural language understanding), they’re not building their own solution, but are using existing NLU solutions to cope with all the different languages airlines need to speak with their customer, according to co-founder JoseLuis Vilar:
“If we would try to build an NLU solution for all the languages our client’s needs, we would be dead.”
Not replacing apps
“We won’t build bots for everything, only where it’s natural to have one.”
After being live with several chatbots for nearly 3 months, collecting thousands of interactions, they’ve already learned a lot of valuable lessons, the biggest being not ask the same questions over and over again.
According to the startup the distance between success and failure is quite short, so you need to get things right the first time around. And according to Vilar, even though the risk for failure is high, the reward for the customer when things go right is much higher.
Work on building a solid knowledge base
After tracking their bots conversations the latest months, Caravelo has found that 20 percent of the inventory of intents (Q&A’s) makes up around 80 percent of the total value. So you need to focus on that part first, building a really strong inventory to start all conversations, the right username and contact, pictures, etc.
The next 15 percent of value is based on the inventory of questions and answer (intents) where you need to build a solid knowledge base for your specific industry. In Caravelo’s case, they’ve built a database of 1000 FAQ’s related to the airline and travel industry, and this is based on interactions the airlines have had with their customers over the years.
The last five percent of value are from the questions a bot cannot answer, and the idea is this is the place where we get the human take-over, and a customer service agent will serve any remaining problem.
The most important learning Caravelo has done the last months is to avoid user loops, like shown in the pictures above. There are few things as annoying for a user to go through the same questions over and over again.
So far Caravelo’s solution to bugs like this one has been to build a small fix where the bot only can ask the same question a certain amount of times.
Another key learning is to use, but not abuse the NLU (natural language understanding). So for example, today they have some answers that go through the back-end from their database, and some answers that go through the NLU, but they classify the easy answers, like affirmations, to not go through the NLU.
Just as with any product, the on-boarding of the user is crucial to keep people talking to the bots. In the case of chatbots, you need to tell them what you’ll be doing for them, and give the user clear options.
The last take-away the Caravelo team has learned over the course of using bots in real life, is to not take too much advantage of bot trainers (external services), as they’re not building a natural language understanding themselves, and it’s easy to get too dependent on them.
To get the full value of Caravelo’s learnings, take a look at the video at the top.
If you want to learn more, take a look at our latest podcast about the European VC industry: