Can monotonous and time-consuming customer service tasks be entrusted to artificial intelligence? Together with our client, we have created a well-functioning solution relying on the latest generative AI technologies.
The Client
booked4.us is a highly customizable online booking system utilized throughout Europe, serving as the digital tool of choice for various industries from private tutors to franchise networks. Its versatility and user-friendly interface have made it an essential component for professionals seeking to streamline their scheduling processes and enhance client interaction.
The Challenge
The number of booked4.us clients has doubled rapidly in recent years. The tasks of customer service are diverse: billing administration, answering incoming emails, receiving incoming phone calls, and assisting customers. Additionally, a key activity is calling newly registered users who are trying out the free trial period, supporting them in customized system setup. This is an important customer acquisition strength: those who are contacted in this way are more likely to become subscribers.
At the same time, the success in growing the number of customers has significantly increased the volume of tasks in billing, subscription management, and customer support. As a result, customer service has had less and less time for customer acquisition and retention activities. Thus, booked4.us sought a solution to automate repetitive, time-consuming tasks. A survey conducted among their team clearly identified incoming email inquiries as a target area, suggesting that if they had to spend effort on this, more time would be available for value-creating activities.
In other words, a solution has begun to take shape that, relying on the appropriate backend calls, is capable of some preliminary processing of e-mails written in human language and searching for responses. It was obvious that the possibilities offered by generative AI should be leveraged for this.
The Solution
We worked together with booked4.us customer service, reviewed their processes, and began analyzing the incoming emails – all from the perspective of identifying where an AI-based solution could help and where the most manual work could be replaced with the least investment. From this, we concluded that collecting and formulating the appropriate responses for different categories of emails was likely the best area to target.
The first step in processing the emails was categorization: determining which email belongs to which case group, as this forms the basis for different logical operations later on. We asked the customer service colleagues to categorize a significant number of emails. Following this, we were able to fine-tune the OpenAI model to learn this task, achieving an 85-90% correct classification rate. The low-hanging fruit was answering simpler questions, for which we used the booked4.us user manual to build a vector DB-based knowledge base. The most effective method for improving accuracy was saving question-answer pairs. Our framework includes an admin interface for creating, editing, and modifying the knowledge base, where the solution could be tested and continuously refined through a chatbot after the initial upload.
At this point in the project, OpenAI released the function calling feature, which made our work easier. By integrating it into the framework, we were able to perform queries and operations in the booked4.us backend systems based on the content of the emails. We could instruct the AI model on the available services and provide a brief description in human language of what they entail. Additionally, we specified the format of the input for the queries, i.e., what data and in what format it should expect.
Thus, the model could determine, in response to a question/request received in an email, which service to call with which input and provide the response to formulate the reply. One of the most common requests was modifying a subscription (number of calendars, payment method, payment frequency). We familiarized the AI model with these services, enabling it to display on an interface which service to call with which input when a modification request was received. We found that it handles these cases very well: it can extract relevant information for modifications from completely unstructured text. Once this is done, the agent reviews the suggestion and, if appropriate, can make the modifications with a single click. After that, the AI solution is also capable of writing the response letter, for example, confirming that the client’s request has been fulfilled with the following content.
Using function calling, the knowledge base, and fine-tuning, we created a system that could draft responses to a significant portion of the emails with minimal or no intervention from the customer service agent (e.g., correcting a stylistic error). We achieved an over 70% correct email response rate.
The Results
- Automation has freed up significant time for value-creating tasks that require human interaction, leading to substantial improvements in customer conversion and retention rates.
- With a one-time investment, booked4.us acquired a continuously evolving (learning) system that centrally stores and maintains corporate knowledge.
- Due to the project, processes and the General Terms and Conditions were reviewed and improved.
- Employee satisfaction has improved as colleagues can spend much less time on repetitive tasks, allowing more room for creativity.
- Future development opportunities include the introduction of a chatbot and the use of the knowledge base as a training and examination platform for new colleagues.”
