How we saved customer support dozens of hours with an integrated chatbot

The e-shop of MISURA - a Czech brand of premium electronics - has long struggled with a growing number of inquiries directed to customer support. Given the company's focus on innovative products such as specialised portable monitors, the increased number of customer queries was to be expected - people naturally want to ask about technical details, compatibility, availability or order status. Questions were coming in both by email and phone, and employees were spending a lot of time each day processing them.

MISURA wanted to divert these routine inquiries away from their team without having to hire a new employee or use an outside call center. The goal was to find an efficient and scalable solution that wouldn't require major changes to the way the e-shop operated.

How to choose a chatbot?

After a joint analysis, we decided to train and deploy a custom AI chatbot that would be integrated directly into the e-shop. During our research, we came across a large number of tools that pretended to be smart assistants, but after deeper testing, it turned out that in most cases they were just very simple systems with limited ability to respond to anything outside of a predefined list of questions and answers.

We therefore set the following requirements:

• The ability to train the chatbot on its own data
  In order for the answers to match reality and not be based on public sources, we needed the chatbot to be able to process its own product feeds, articles and internal documents such as terms and conditions or complaints policies.
  
• Easy integration into the UpGates platform
  The e-shop runs on the UpGates system, where we wanted to maintain full compatibility and easy connection without the need for complex code interventions.
  
• Ability to dynamically retrieve data from product feeds
  Prices, availability and specifications change regularly, so the chatbot needs to be able to work with and dynamically retrieve up-to-date data.

For these reasons, we finally chose the Chaterimo platform, which allows to link custom data with OpenAI language models. Once the API key is loaded, the chatbot runs directly on the ChatGPT model, but with the ability to train accurately on the company data.

Training data for the chatbot

In order to achieve the most accurate answers, we uploaded the following:

• e-shop product feeds, from which the chatbot continuously updates price and availability information
• articles from the brand page, which often answer typical questions and serve as a good basis for answers
• complete terms and conditions, complaints policy and service descriptions

In total, we have trained the chatbot on more than 20 million characters and more than 5,000 datasets in all languages in which the MISURA e-shop operates.

At the same time, we had to define specific constraints during training to prevent misuse of the system, such as:

• prohibition of providing information about discount coupons (inspired by a case from the airline industry where the chatbot randomly provided employee codes),
• prohibiting the provision of margin information or internal data
• preventing the use of the model for purposes other than customer service (e.g. help with school assignments, etc.)

The Chaterimo interface itself provides great configuration possibilities. In addition to the choice of the language model, there is also, for example, the possibility to adjust the so-called "temperature" of the chatbot, which marks the control of the balance between creativity and determinism in the answers.

The chatbot also allows visitors to the e-shop to leave a contact, thus becoming a secondary tool for collecting leads and further work with potential customers, for example in the form of mailing.

After a series of internal tests, we deployed the chatbot directly on the e-shop. Using the Chaterimo interface, we modified its appearance so that it would not be distracting on the website and would fit into the overall design. The reactions were positive from both customers and the MISURA team itself. Due to the high satisfaction with the solution, we subsequently integrated the chatbot into Facebook, where it now answers questions directed to the page's private messages.

The results after more than a month of chatbot operation

The result is a significant saving in customer support time, as well as higher quality and consistency of responses to customers. Within the first month, more than 50 customers used the chatbot on the e-shop and 95% of these queries were resolved through the chatbot. The load on the customer service line was reduced by almost 70% compared to previous months, allowing MISURA staff to focus more on developing new innovative products and supporting sales of existing ones.