Allegro.pl is the most popular shopping platform in Poland with one of the largest e-commerce websites in Europe. Its marketplace model has over 164 million offers from all key categories. Each month, 20 million customers visit the online platform, and their needs are attended by 2,200 employees working in five Polish locations. The Findwise & Allegro digital workplace team started working together at the end of 2017. As a part of the knowledge management program, joint projects from the outset were related to the accessibility of information.
Allegro are developing dynamically with the number of employees constantly increasing. It's not surprising that finding relevant information can be challenging, especially when searching across several data repositories such as Google Workspace, Confluence and JIRA. It’s typically very time consuming and employees often don’t know where or even how to search for relevant information. Outdated content plus not reusing existing information assets causes frustration and increases costs.
The first natural step was to deploy an enterprise search solution. For this purpose Findwise i3 along with Apache Solr was used. The platform was integrated with Google Workspace and the intranet homepage. However, that wasn't the end of the journey! The Allegro team understood quickly that discovering corporate knowledge and accessing information is a multi-dimensional problem.
One of the remaining challenges was meeting the ever-increasing user expectations of using multiple information sources as well as the desires for different types of interactions with these systems. This is how the idea of creating a chatbot was born.
The chatbot’s interface provides natural language support and a more human-style type of communication. Furthermore, it adds a great potential to automate certain business processes within areas such as HR and IT support. The Findwise and Allegro team took a step-by-step approach which included planning, designing (including a workshop dedicated to gathering requirements), proof of concept and then a pilot. From early on in the process the virtual assistant learned how to answer questions related to:
- Employees, their phone or room numbers, presence in the office, roles, teams etc.
- HR applications like vacation form, sick leave, business trips etc.
- IT services, their business and technical owners
- Employee meetings in Google Workspace Calendar
- Triggering enterprise full-text search interface
- Employees' IT assets and devices
Open-source Rasa was chosen as the conversation engine and it was integrated to the existing i3 installation. Having both the enterprise search & chatbot as independent applications running on the same platform made it possible to reuse mechanisms, data flows and helped to reduce implementation and future maintenance costs. The solution also saw an integration to Slack, which is the company’s main communication tool.
A key aspect of this project is understanding the meaning of user input. The solution doesn’t depend on triggering phrases or static rules. It goes beyond just simple keyword matching. The embedded machine learning model ensures NLU/NLP capabilities. Depending on the user story, we observe accuracy in solving users’ requests from 75% to over 90%. It’s not easy to understand the intention. Language traps, such as rich declination or surnames, tend to muddy the waters. Another challenge is to build a proven chatbot’s fallback loop. It's also worth to have a long-term strategy plan, but start from narrow scope and clear business case.
We managed to overcome these obstacles. Our prize is high efficiency and thus high conversion on a platform that tends to support everyday tasks of employees. The solution was neatly woven in one common information layer for both search and conversational solution, which optimizes the return on investment and ensures consistency in digital channels.