We analysed Booking.com’s customer service conversations and realised that 20% of them followed a similar pattern: The agent would refer the customer to a self-help page that provided the answer to the question the customer had. We set out to automate this process.
But we had to do better than automate, because the customer experience was awful. The self-help articles were, on average, 800 words long, and after reading the page, customers would often still need to navigate further to other self-help pages to get their answer. The answers to many questions lay within them, but it was hard to get them out.
We fed the 1800 self-help articles into a custom vector database and built a ChatGPT-powered chatbot that would try to address users’ queries and escalate to the existing call center if it couldn’t. We ran 500 past queries against it and found that it answered satisfactorily in 91% of cases.
The precision of the answers was impressive. Whereas previously users would have to sift through 1000+ words or self-help article, ChatGPT would now answer their specific question with a few concise sentences - in under a second and 24 hours a day. An unexpected benefit was its ability to speak all languages even though it was only trained in English. This is a huge cost saving, as Booking.com usually pays professional translators to translate their content into 9 languages.
