Lingerie manufacturer had just updated its distribution system and started to sell online using online retail shop. Results of several weeks of advertising and selling made them think that leads were not converting because of some barriers that I had to identify and get rid of.

I used webvisor by Yandex Metrics for observing website visitors browsing history, click maps and traffic reports in order to find entry and exit points and their patterns. This analytics led me to a hypothesis that visitors need quick consultations on the website to ask for some questions about new to the market product offered on the website. Such functionality had to be added to the website.

Before starting the project I gathered the requirements and constraints in order to choose one of the schemes of interaction with website visitors. Key requirement was to provide user with maximum relevant information for decision making in minimum possible timeframe.

For this purpose we decided not to use such instruments as Jivosite, which in fact transfered leads to managers as this option had proven to have insufficient efficiency. Product’s specific features required that accurate and relevant information for the visitor should be provided without human interaction.

I started to look to consultant bots which immediately provided information after some interraction with them and which compliment website’s interface. Initially I looked at Telegram bots but due to possible blocks of this service in Russia we decided to use another B2C chatbot.

We stopped at cloud service Pandorabots which could be quickly set on manufacturer’s website using footer’s script. Thus I could spend most of the time on content generating, collecting frequently asked questions and correct answers which was good because of limited resources for this project.

Choice between scripted and machine-learning bots was easy because the former one was much cheaper and faster to implement. Also scripted chatbot suited better for specific offers on website. Thanks to analysis of questions asked by visitors of the website in previous several months, I could elicitate blocks of questions which covered almost all types of visitors' questions.

While populating chatbot with information and scripts I decided to emphasize on ready-made questions instead of typing random text to the chatbot as it drastically increased the speed of interacting with the chatbot. In result, we shaped business requirements to create dialog bot with user-friendly interface and convenient backend for the administrator. Pandorabots perfectly fit our task and offered flexible plans paying only for interaction with the chatbot.

One of the most important part of the project was elicitation of users' questions and inquiries which came from website and partners. This information formed the basis for specification document for website interface, content adjustments and advertising campaign as well as SEO.

Incoming information was not structured and was stored in different formats: from emails to notes in manager’s notebooks written after telephone conversations with clients. I was heading to collect all this information, structure it which was possible only by laborious and hard work. Analysis was made to categorize questions and identify frequency of these questions.

After elilcitation I had to create the structure of questions to simplify access of users to the required answers. I had to create answers tree which was a base for AIML (markup for AI). AIML is a XML dialect and it is easy to read and use.

This tree quickly received very wide branches and it became inefficient to store information for the chatbot backend. I decided to use this tree only to store information about markup files and connections between them and information for AIML files was saved in XLS file as it was easy to search information in it, add rows and copy data from it.

Thanks to user-friendly interface the process of AIML files creation and their linking was not difficult but I had to have accurate linking tree which had been created beforehand using UML. Thanks to this diagram, numerous connections of routes of answers followed to correct AIML files.

AIML files were created using tags and information about them was provided by the developer in chatbot instructions. I experienced several difficulties writing a deployment code but thanks to IT support and rule of thumbs the chatbot worked according to the expected algorythm.

The distinct feature of this project was a method of interaction between the developer and the client. In fact business analyst received a task which evolved to the pattern of interaction when business analyst and developer became the same party. This situation happened mainly due to the lean approach to goal achievement when cookie-cutter instruments were used as much as possible.

Usually enterprises order chatbots from certain companies or freelancers and in such cases business requirements document and specification document are obligatory. But this project I could complete by my own efforts in shortest timeframes and I decided not to pay much attention to the business requirements document for my own use.

The main document was a project implementation plan with deadlines for different stages of it. These tight deadlines helped me be in bright-eyed and bushy-tailed.