While most enterprise businesses are running Generative AI pilot projects, typically they are taking a cautious approach to customer facing applications due to the reputational, financial and brand implications should something go wrong.

Booking.com are an example of a business that have already rolled out a customer facing product with Generative AI at its core, the AI Trip Planner. The planner incorporates the benefits of LLM’s with the value of their internal data, to deliver a differentiated product in the travel market.

In this interview with Charlotte Munro, Global Product Marketing – Generative AI at Booking.com she covers the business case behind the product, how they build the internal team that worked on it, the risks they faced and how they managed them, and the early feedback from their customers.

Why did Booking.com decide to integrate Generative AI into its products?

• While the AI Trip Planner is a new product that we developed and launched in July of this year, AI and ML has always been an integral part of Booking.com, recommending destinations and options to millions of travellers every day on the platform. We have been using AI extensively for years in order to optimise interactions with our customers who are both travellers and partners - from laser focussing / contextualising / tailoring recommendations and the user experience, through analysing the content and pictures provided by our partners and customers. It is the strong foundations of AI within Booking.com’s ecosystem that allowed us to develop a robust AI Trip Planner so quickly.
• We believed then, and we continue to believe now, that we can build a more compelling and differentiated offering if we can leverage AI technology to deliver a more tailored and relevant booking experience, a connected trip, that would be more responsive to a booker's needs and help manage the different aspects/components of their trips.
• While we make some use of OpenAI’s ChatGPT to partially power the new conversational experience, Booking.com’s existing machine learning models enrich the information presented to customers with tailored destinations and accommodation options.

How did you assemble the team to create your first customer facing Generative AI products? What job functions were represented?

• We assembled a virtual task force team of colleagues that represented all functions of our business. This included our ML experts, Product, UX and UI as well as our research, product marketing, legal and PR teams.
• To compliment the technical skills and requirements, and ensure we created a product that resonates seamlessly with our customers, we then focused on ensuring we had the right customer insights, legal protections, and impactful key messaging. This meant we not only had a great MVP product, but also one which should have longevity build in

How are you ensuring that customers get a differentiated experience when using Generative AI?

• We know Booking.com is really good at helping our users book their trip, especially when they know where they want to go and have a fair idea of when they want to travel. However, we know that before travellers get to that point there is a lot of discovery and inspiration that happens. The AITP allows users to start their search and discovery based on their trip intent - whether it be a beach holiday with their family or a hiking adventure with their friends or simply exploring a range of potential destinations and options. This means we are interacting with customers throughout their entire trip journey – from ideation through to bookings.
• The AITP allows for a more unconstrained search experience where users can quite simply tell us what they want, and we show them options that suit their needs. Travellers can ask the AI Trip Planner general travel-related questions, as well as more specific queries to support any stage of their trip planning process. This includes scoping out potential destinations and accommodation options, providing travel inspiration based on the individual traveller’s needs and requirements, as well as creating itineraries for a particular city, country or region.
• The AITP also really helps travellers by saving them time and effort - doing the hard work for them.
• We are always looking for new ways to make customer interaction smoother and richer and our recommendations have a seamless visual UX which really helps our travellers explore and get an idea of a potential destination or accommodation for their next trip.

How did you manage the risks associated with the technology, such as privacy and security?

• Our customers' rights, privacy and security is always of the utmost importance to us, which is why we engaged our legal experts from day one of the build. The AI Trip Planner went through a robust product review process and has been developed according to rigorous standards to ensure a safe and inclusive environment for travel exploration. We used the same robust legal protections that we use with all of our products and this includes custom AI moderation layers for the AITP, which detect and block harmful content, and ensure our customer’s data remains safe and compliant with all regulations.
• We also carry out robust data protection impact assessments to identify any areas of potential risk and to ensure compliance with the applicable data protection laws
• With all our travel experiences we want to ensure that travel is a safe and inclusive environment for everyone, and this traveller experience carries through to how customers interact with the AI Trip Planner.

How do you protect personal information and prevent hallucinations?

• Booking.com has built a custom moderation layer to structure the collaboration between the LLM and Booking.com’s machine learning models, as well as guardrails to detect and remove certain personal information and/or non-travel related content. This also works together with OpenAI’s moderation to help identify and block potentially harmful content. The aim is to remove personal data when it’s unnecessary to the search and travel planning process.
• To minimise and mitigate any inappropriate and inaccurate responses being surfaced to travellers, we have introduced moderation content filters which remove any offensive questions and informs the user very clearly that the tool cannot and will not respond to such questions.
• We have made some good strides in moderating and filtering out inappropriate content and will continue to innovate and develop within this space to optimise further and incorporate learnings to ensure accuracy and relevancy of the content that is surfaced. We have also carried out robust data protection impact assessments to identify any areas of potential risk and to ensure compliance with the applicable data protection laws.

How do the LLM and Booking.com work together to answer a question?

• The AI Trip Planner is a combination of third-party LLM technology and Booking.com’s own machine learning recommendation systems. The LLM technology that we are currently leveraging from OpenAI enables the AI Trip Planner to quickly read and understand questions that are being posed in the tool, as well as to provide responses in a conversational, natural way. The LLM also references a wealth of other data to provide certain kinds of recommendations, for example itineraries with suggestions of great things to do in a specific destination.
• Where the Booking.com machine learning models take over are when the data needed to respond to the question is fully in Booking.com’s realm of expertise, for example with specific destination and property recommendations. All of the visual elements in the tool, including pricing information are from Booking.com.
• To make it even more concrete/simple, conversational text and answers to general travel questions are provided by ChatGPT. For example, what’s the weather like in Bali in December? Or what’s a good itinerary to visit the north of Italy for a week? The visual links for specific destinations and property recommendations, as well as the moderation layers that help keep content travel-related and safe come from Booking.com.

Do you have any initial market feedback on how the product has been received?

• The objective of rolling out the MVP slowly and market by market allows us to continue to develop and optimise the product based on how our customers are interacting with the product and what they are using it for/type of broad queries that they are asking. We are seeing travellers explore popular destinations as well as themes like beaches and family trips. Itineraries are also popular. What would you do differently if you ran the project again?

What do you wish you had known before developing this product?

• I think the nature of AI and this type of emerging technology means that you are never going to have the perfect test and launch conditions. Which is why developing an MVP that we gradually roll out to our customers and new markets, allows us to improve the product and the experience. As we can get early signals from our customers - how are they interacting with the product, what kind of themes they are exploring and how are they using it? We will continue to evolve the AITP and ensure that the product meets the needs of our customers.
  

Return to Home