How to make your chat and voicebots even smarter

Companies that use chatbots are generally convinced by the technology, but often struggle with the unsatisfactory results. Jörg Feldmann, founder of kompaktwerk and expert in digital customer dialog solutions, shares his experiences and talks about the improved options offered by enterprise search and the latest generative artificial intelligence.

Companies often see a need to optimize chatbots.

In our discussions with companies, we often notice that the performance of existing bots is often critically questioned. The usual feedback is as follows:

  • Our bot can respond appropriately to too few customer requests.
  • Customers do not use the bot as expected because it generates too few correct answers.
  • Our bot is not able to establish a natural flow of conversation.
  • Managing the bot causes considerable administrative work for the customer service teams.
  • The management and maintenance of content is decentralized and difficult to control.
  • There is a lack of a wealth of information to make the bot more intelligent.
  • If anything, the use of the bot has worsened the user experience.

This occurs almost regardless of whether the intelligent software package is provided by companies such as IBM Watson, Cognigy, RASA,, OMQ, Zendesk or other providers.
But what is really behind it?

The intelligence of chatbots

The intelligence of chatbots in companies is not always sufficient, as they are trained for specific questions and processes and are therefore unable to handle unpredictable or complex requests. In addition, insufficient data quality and quantity can contribute to chatbots not always being able to provide precise answers.

In addition, maintaining and updating the content is often time-consuming. It is crucial to provide the chatbot with accurate and relevant information so that it can provide appropriate responses to queries. This requires cooperation between different departments within the company to ensure that all relevant information is available. It is also necessary to regularly review and update the content to ensure that it remains current and accurate.

To overcome these challenges, companies must continuously train the chatbot and constantly check and update its database. However, our experience shows that this is precisely where the main problem lies: The chatbot was originally trained for simple questions and processes and therefore has difficulty processing complex requests. However, since such data is not available, the chatbot cannot provide precise answers to such questions.

How does the chatbot obtain its knowledge?

Improving the intelligence of a chatbot can be achieved by using an AI-supported knowledge management solution. This solution enables the chatbot to access a wide range of information and knowledge from your company to provide more accurate and appropriate answers to user questions. Such an AI-based search engine, also referred to as an enterprise search system, helps to increase the accuracy and relevance of answers by selecting and providing relevant information from the knowledge base and other connected data sources. This search system can easily access various systems in your company via so-called connectors, such as M365, Confluence, SharePoint, Jira, wikis and more.

In addition, an AI knowledge management solution also enables more efficient management and updating of knowledge. When new knowledge is added, updated or deleted in the connected data sources, the company search engine is automatically kept up to date.

Overall, a high-quality AI knowledge management solution can improve a chatbot’s intelligence by providing it with a comprehensive and up-to-date knowledge base, while giving it the ability to use this base effectively to provide relevant answers to user questions.

Alternatively, you also have the option of integrating generative models such as ChatGPT from OpenAI, BARD (PaLM2) from Google, or Luminous from Aleph Alpha (German manufacturer) into your question-and-answer communication instead of a specially trained chatbot. The advantage of combining an AI-based search engine such as iFinder with an integrated generative AI model is that only the data that is actually available to the user is used, whether based on their authorizations or on the trained model. This significantly minimizes the risk of “hallucinations” caused by the generative model.

How do you increase the intelligence of your chatbot?

We offer an effective approach to increasing the intelligence of your chatbot. Our approach to improving your chatbot includes a Natural Language Question Answering module (NLQA) in the form of a question-answering system. This powerful search extension has been specially developed for interactive question-and-answer chatbots. With this solution, your chatbot can access company-wide information quickly and efficiently, which in turn offers the opportunity to optimize your chatbot dialogues and significantly increase user satisfaction.


The NLQA module ensures that predefined answers are extracted from your company data by our system and seamlessly integrated into the chatbot dialog. The iFinder is an AI-based knowledge management solution that extracts terms, texts, keywords and corresponding questions from various data sources and delivers suitable results. This application not only searches your chatbot database, but also other data sources such as Confluence, SharePoint, other content management systems and, if necessary, voice or chat transcripts.

The maintenance effort is considerably reduced, as the iFinder already provides many concepts “out of the box”, which reduces the maintenance effort for the “intents” (recognition or understanding mechanisms) or questions. It also draws on existing knowledge from the available information. For example, the search can automatically add synonyms to the search query. It not only searches for “pension”, but also for terms such as “pension”, “old-age benefit”, “retirement pension”, “company pension”, “statutory pension”, “old-age security” and “old-age provision” to return corresponding hits. The textual preview of the search results already provides the context for the answer you are looking for. Ranking by relevance enables the best answer to be selected precisely from the available information.

We are convinced that question-answering systems (Natural Language Question Answering) offer numerous advantages compared to pure chatbot applications. Our main focus is therefore on implementing the question and answer system. Compared to a modeled chatbot, a question-and-answer system requires less initial project investment, shorter project runtimes and less maintenance effort.

It is important to emphasize that a chatbot is helpful when users do not know exactly what they are looking for. In addition, a chatbot can be useful in situations where the search function is not used, whether due to lack of awareness or lack of access to the data. Our experience shows that a search usually provides an answer more quickly than a guided dialog.

Our question and answer system also enables voice control of the search. The iFinder offers an optional Speech-to-Search function. Our solution can be provided both on-premises on the company’s own servers and as a cloud-based SaaS service.


We are convinced that chatbot applications offer the greatest added value when they directly complete the information transfer process and are actively involved in the handling of processes, for example when ordering a pizza or taking out insurance. If the main function of the application is to find and pass on information, then the concept of generating answers directly from the existing content is the optimal solution.

A generative language model such as ChatGPT cannot replace a search engine or knowledge database, as it does not have up-to-date and detailed specialist knowledge. The better option is a search engine that can integrate a generative model. A search engine can provide text passages from documents that are likely to contain answers to a question asked, and the language model can convert this information into a comprehensible answer that can be used in a chatbot.

Therefore, the ideal solution for companies is to combine an AI-based search engine with automated text extraction, such as iFinder, and a generative language model similar to ChatGPT. The generative AI takes over the human-like, automated dialog, while the iFinder extracts the relevant content from the company’s data sources.

We would be happy to discuss your individual application in person. As experts in the field of knowledge management, we have been working intensively with various large language models (LLM) for some time now.
We offer you various implementation proposals with advantages and disadvantages so that you have an optimal basis for making a decision for your project.