Table of Contents
The Holy Grail of marketing is offering the right product (or information) to the right people at the right time, and personalization has been playing a key role in that for quite a while. And as technology advances, so are personalization tools, with smart chatbots being among the most popular ones. Smart chatbots are an example of generative AI – a subset of Artificial Intelligence which focuses on content generation. By using such bots, companies can significantly cut down response time and speed of request processing while boosting customer service and user experience.
SoftTeco implemented an AI-based smart chatbot in the new company website and is now using it to attract potential customers and provide relevant and up-to-date information to website visitors. The use of this chatbot also allows us to better understand our clients’ interests and align our marketing efforts correspondingly.
The problems of standard bots and how AI helps
First, let’s discuss how AI bots differ from traditional ones and what challenges they help resolve.
A standard bot is a simple robot that can answer basic questions and provide limited information. Such bots normally don’t go beyond preconfigured answers and the accuracy of their responses varies from medium to low. Therefore, standard bots do not provide sufficient customer experience as they cannot immediately provide accurate and detailed information.
However, modern customers expect bots to answer immediately and in detail. And since at SoftTeco we aim to stay ahead of the curve in Gen AI marketing paradigm, we came up with the decision to create, train, and implement a fully automated AI assistant to serve our customers on site. One of the biggest requirements for this bot was that it understands the specifics of the IT industry and relies only on trusted sources when providing information.
A quick review of SoftTeco’s smart AI chat
SoftTeco’s AI bot functions on the base of GPT-4 and uses our own data as the primary source of information. By that we mean the data that we used for bot training, such as company information, internal resources, etc. One more distinctive feature of the bot is the manner of its replies – we trained it to resemble the company representative and to maintain a professional and helpful attitude.
The bot provides the requested information almost instantly and can be easily updated by a marketing specialist whenever a new relevant information appears on the website. Regular updates help ensure that the responses are always accurate and relevant. Overall, the bot now serves as a valuable AI-based assistant and can be used not only by our clients but by employees too.
Main challenges that we faced during the bot implementation
Before explaining the process of bot training and implementation, we’d first like to highlight the main challenges that we faced during the processes:
- Multilingual support: it was critical for us that the bot understands various languages. For that, we used two external libraries, langdetect and FastText by Meta.
- Creativity and determination: we wanted the bot to go beyond standardized answers so we utilized prompts, system messages, and various configurations of the RAG system itself to tweak the responses and their style and depth.
- Priority level: our experts meticulously defined the priorities for the distribution thus ensuring the accuracy of responses and proper information sources to be used by the bot.
As you can see, the biggest issue with AI bot training is that it should sound natural, push beyond standard chatbot boundaries, and rely on specific information, not just on random Internet webpages. The good news is that once you resolve these challenges, you get a powerful and intelligent tool, perfect for leveraging your marketing efforts.
Expert Opinion
We have tested lots of various methods during the development process: from classic indexlinking or Spacy to innovative RAG systems. As expected, the latest technologies like RAG prove themselves better for such complex tasks as AI bot fine-tuning, which is why we decided on them
How SoftTeco’s AI bot processes the data
As in any RAG system, in order for the bot to “learn” the information from our documents, we first feed it to the LLM. The LLM transforms the text into numerical values, which we then store in our internal database. Upon receiving a request, we transform it into numerical value too and compare to the values already stored in the database. When we find the most similar value, that means that the request and the found value share the same context or topic. So the next step is taking a bit of the data from the found value, the initial request and addressing the LLM with the prompt “answer this question based on that data”. And that’s how the training process is technically carried out.
To keep the bot updated with relevant information, we also had to find a way to add it to the bot’s internal knowledge base. Here is a step-by-step explanation of the process:
- SoftTeco’s specialist adds new pages on the website
- These pages are then added to an internal document with the knowledge base
- Once the bot requires an update, website administrators initiate it by sending the bot to the document with the knowledge base
- The bot checks the list of pages in the document against the ones that are not allowed for parsing and parses the remaining pages.
- After the parsing, the bot receives the information from all these pages, sends them to the LLM for conversion and completes the update.
Training process explained
There were several user testing stages in addition to technical testing completed by SoftTeco’s QA engineer:
Basic testing
During this phase, we tested how the bot perceives and processes the information and how accurate its responses are. For better evaluation of the bot’s performance, we have defined several categories of questions:
- Expertise
- Portfolio
- Job
- Price
- Contact
For each category, we wrote down a list of questions and tested how well the bot answers them. In case of an incorrect or too vague answer, we added extra rules to the bot configuration and tweaked its creativity level to achieve the perfect accuracy of responses.
Multilingual testing
As an international company, SoftTeco works with clients from across the world so it was important that the chat understands multiple languages. However, as English is the primary communication language both for our company and for our clients, we configured the chat so it replies in English only. This feature also facilitates the collection of the data and the understanding of user behavior on the website.
A mini challenge that we faced was that the bot had issues recognizing Korean, Chinese, and similar languages as it perceived hieroglyphs as symbols and not as words. Our expert fixed it by writing additional rules and adding them to the bot configuration.
Accurate recommendations of relevant pages
The last stage of the user testing process was testing the quality of recommendations that the chat provides in relation to user’s questions. This was a bit challenging from the technical point of view – here is why.
The chat provides recommendations based on certain key words. The main issue was that the company’s blog has more key words than service pages. This led to the bot recommending mostly blog articles, even if they did not match the user’s requests. To fix that and to ensure that the bot recommends only relevant information (from both service pages and articles), we introduced the priority system which artificially lowered the reliability of the blog and increased the reliability of service pages. In addition, we excluded certain pages from the response list for the chat and, as a result, recommendations became more accurate and valuable for users.
What can the bot do? Core features that drive business value
While the core function of SoftTeco’s AI smart bot is customer service, it performs several other critical tasks that contribute to the growth of our company and constant improvement of our services:
Logging. We track and log all bot-user interactions to analyze requests and improve the bot’s performance. In addition, bot logs are integrated with analytics so the marketing team can easily access needed information about user behavior on the website.
Easy and flexible management. Marketing managers can independently set rules for the bot configuration and adjust its performance by adding priorities for various information sources. In this way, marketers can better manage the priority of various pages on the website and fine-tune the accuracy and relevance of responses.
Reporting. As mentioned above, the bot helps analyze what services interest potential customers of the company and how exactly users interact with our AI assistant. For better access and management of this information, there is a detailed admin panel that displays statistics on the bot, allows bot reloading, and helps monitor set trends and KPIs.
Improved user experience. The bot is able not only to accurately answer questions but also provides relevant recommendations and allows evaluating its responses by liking / disliking them. This feature also helps marketers better understand how valuable the responses are for users and whether the bot’s accuracy should be improved.
Automation of customer service. The main goal of AI-powered bot is to automate customer service to a certain extent and speed up the delivery of information to users. Hence, the introduction of the bot significantly reduced the load of the customer service department while maintaining user satisfaction and customer experience on a high level.
SoftTeco’s AI bot now: results and statistics
SoftTeco’s AI bot is a full-fledged virtual assistant that performs several important functions and helps our team deliver consistently good customer service. For marketers, it is a valuable tool for tracking online behavior of our users and an opportunity to independently see and manage statistics and bot’s performance. For website users and employees it serves as a primary source of relevant information.
In the future, we plan to further improve the bot and potentially add new features to it. And right now, you can visit our main page and see the bot in action – we’ll be happy to receive your feedback in the comments section!
Comments