AddAI.Life (AddAI) (link resides outside of ibm.com) is a Czech-based startup, and an IBM Business Partner, established in 2019 with a vision to help companies improve their customer experience and efficiency by implementing AI assistants that are available 24x7 to answer a wide range of questions promptly and accurately. “When developing virtual assistants for our customers, our goal is to ensure each assistant sounds very natural and brings the customer and its end-users an amazing experience,” explains Jindrich Chromy, CEO and co-founder of AddAI.Life.

When it comes to generative AI, trust and transparency, as well as accuracy, are critical for AddAI and its customers. Each assistant needs to be specifically tailored to each customer and its unique use-cases: whether it will be answering billing questions for a user in telecommunications or resetting a digital account password for a banking customer. “Our customers include some of the top Czech companies as well as telecommunication corporations and banks. It’s important for our assistants to be able to deliver accurate and consistent answers. Moreover, the answers need to be curated or generated automatically on verified data. We also need to comply with bank-level security, which includes granted data exclusivity for the customer. We also need to be able to opt-out the training of the AI model on our user’s conversations,” adds Chromy.

In order to continue delivering an exceptional customer experience while also shortening the time it takes to implement a new AI assistant, AddAI has begun using the IBM^® watsonx.ai^™, IBM watsonx^™ Assistant and IBM Watson^® Discovery solutions.

In collaboration with the IBM Client Engineering and Technical Sales teams from the Czech Republic, AddAI built an AI assistant for a customer in the banking sector that they then piloted. The assistant utilized the retrieval-augmented generation (RAG) methodology, coupled with watsonx.ai, to deliver relevant and precise responses in the Czech language for frequently asked questions (FAQs) on topics ranging from account access to payments, pricing and account services. Through intensive optimization efforts, the implementation team ensured that the answers generated by large language models (LLMs) met a high standard of accuracy and language fluency. To optimize end-user experience, the team integrated watsonx.ai with watsonx Assistant, and was able to deliver instant replies on a user-friendly front end.

“We use watsonx.ai to access selected open-source large language models like Llama 2 to enhance and contextualize information from our databases and present it to the end customer through the watsonx Assistant chatbot in an accurate and conversational manner. In our case watsonx.ai generates its answers to users’ questions through RAG,” explains Klaudia Miezgova, watsonx developer at AddAI.Life.

Additionally, Add.AI included IBM Watson Discovery in this pilot for its banking customer. The team used publicly available bank documents such as contracts, terms and conditions and price lists and tuned Watson Discovery to understand domain-specific questions and expressions. “We were able to obtain very accurate answers, including snippets from the document used as a reference or basis of the answer. This happened automatically, without the need to build and set the dialog manually,” notes Chromy.

“We always track, measure and test new ideas to see how they work in real-life situations. That way, we can improve them quickly,” explains Chromy. “In this case, we took a set of testing queries, for which we had the correct answers pre-defined in our hand-built AI assistant which was powered by watsonx Assistant. We then fed those queries to the new assistant which was powered by both watsonx Assistant and watsonx.ai. That one had no pre-built dialogues and had to come up with the answers by itself, using the provided documentation and skill it had as a large language model. It responded with 85% accuracy compared with the pre-defined answers of the former version. With much less work input. So now we know that it is on a good track.”

Through this new assistant, AddAI also helped its customer achieve a 50% reduction in unanswered customer queries. Moreover, when answering customer questions, an AI assistant powered by watsonx.ai is designed to provide a more detailed and actionable answer than was previously possible.

“We have been able to tune the model to query complex data sets and then summarize the data it collected in a much more effective way. For example, the previously mentioned banking customer uses the assistant to deliver fee information to its customers. Rather than serving up raw data that requires the user to analyze the output and infer the next step, the assistant can deliver a clearly summarized answer with helpful next steps for the user to consider. This was not possible before watsonx.ai,” explains Miezgova.

This is just the beginning. Going forward, AddAI is confident it can reduce the implementation time on any new AI assistant by up to 30% given they can now use generative AI to help prepare the assistant to be questioned by end-users. The team expects this to significantly shorten the amount of time it takes to test and optimize an assistant before it is ready to be shared with end-users. AddAI also expects it can provide next-level, more knowledgeable assistants, even without the need of developing each dialogue (response) manually, using this new approach with watsonx.ai and watsonx Assistant.