Hotailors is an enterprise SaaS platform that helps companies organize and supervise business travels of their contractors and employees. Employees can easily book flights, hotels, rent cars, buy insurance, and transfers while an employer receives one invoice to cover these expenses. With Hotailors companies can automate travel policy compliance checks, avoid the hassle of managing hundreds of travel invoices and travel cost reimbursement for employees.
We wanted to validate the hypothesis that AI could provide better hotel options for users looking to book accommodation for their next business trip. We aimed to find out if the previously used algorithm could be improved to offer users their best hotel choice higher in the hotel listing with hundreds of hotels available at once.
With the client’s model, the results took time to display, keeping a user waiting. During the AI Sprint, we wanted to find out if the model could be improved to speed up this process.
During the AI Sprint, we set on track to build several models that would help verify the initial hypothesis of the AI Sprint. In the end, the best-performing model was selected for further improvement and testing. After receiving the final model, the client would feed in more data in the future to make it more efficient to later implement it within the product.
The client’s system was not working properly – it was slow and often caused users to contact customer support. In the long run, this would lead to increased costs of serving customers by customer support specialists.
Another goal was to allow users to spend less time on choosing the best hotel option that, at the same time, would be compliant with their company’s travel policy.
To find out if it was worth implementing AI for Hotailors (and investing bigger budgets), we conducted an AI Sprint starting with a 2-day workshop. After the workshop, we built a Proof-of-Concept that took us 3 weeks and helped validate if AI can solve the problem the client approached us with.
Provided with the user data, we performed data cleansing. The process included, for example, removing empty fields and applying mathematical calculations to substitute them. Cleaning up data contributed to increasing the efficiency of the final model.
We made the extraction of new hotel features possible with feature engineering which means creating new features applying math calculations to existing ones. For hotels, these features included, for example, amenities such as pool or concierge services and much more. By being fed with more features, the model was able to produce more precise search results.
During the AI Sprint, we built a set of prototype models predicting the possibility of a user to choose a hotel from a group of hotels matching search criteria. As the number of available hotels varied we wanted to make sure our model could provide the best choice for a user higher in the listings. That is why we set up the criteria of hotels appearing within top 3%, top 10% and top 20% of all available hotels. For example, if there were 100 hotels available building a model providing a user’s top choice within the first 20 positions would be a way to go! Or if there were just 10 hotels available, showing a user a hotel on position number 3 would be considered a success. With this model, we validated this was possible.
As a result of the AI Sprint, we have identified that AI can provide users with better recommendations and that the client should proceed with testing feeding new data of user searches.
We have validated that with AI, it is possible to show search results faster for users. After further tests and training the model, the client will be able to provide users with better user experience.
The AI-powered model was able to learn from the past history of user choices and offer a hotel that would meet user requirements. We have validated the hypothesis that our model could provide better hotel recommendations appearing in the top 20% of available search results. This has proved that if there were 100 hotels available, our model would be more efficient in displaying the top user choice within the first 20 positions. Having achieved the success criteria in this regard, the client decided to polish the model, feeding in more data and train it further.