Streamlined c-class item sourcing
Struggles with Manual Data Management
When GIS customers provide extensive Excel files containing descriptions of C-class items they wish to purchase, sourcing specialists must manually clean, review, and update these lists. This labor-intensive process is not only time-consuming but also prone to human error. Proper categorization of items is crucial for approaching the right vendors, consolidating demand, and negotiating better deals. The manual process hampers GIS's ability to efficiently serve their growing customer base.
The Potential of Data Task Automization
By automating item categorization and enhancing data quality, GIS can significantly reduce onboarding time for customers and improve product matching with suppliers' offerings. This will lead to increased customer satisfaction, streamlined procurement processes, and the capacity to handle a larger volume of clients without compromising on efficiency or quality.
The Need for an Effective Solution
An effective solution must automate the task of item categorization with high accuracy, minimize manual workload, and continuously improve the performance of the models over time. This will enable GIS to efficiently manage an increasing number of clients and suppliers, ultimately contributing to the company's growth and success.
Faktion's IDQO - An AutoML Platform to Automate Data Quality Optimization
Faktion and GIS International have been trusted partners in AI for multiple years now, and have developed a shared understanding on GIS’ data challenges and the way AI projects are typically engineered and rolled out.
At the end of last year, Faktion and GIS enrolled in a new project to speed up GIS’ customer onboarding time by automating item classification for vendor selection, using the latest state-of-the-art AI tools from Faktion’s ‘Intelligent Data Quality Optimization’ toolbox (IDQO). For GIS, data quality of C-level items has been both the long-term challenge and the direct driver of scaling its integrated supply/business model.
Faktion's IDQO is an AutoML platform containing reusable tools and techniques that can be configured, trained and deployed to automate manual data quality tasks, such as product data classification, tagging, matching, search, data enrichment, and data quality validation.
During the project, the tools have been configured specifically for GIS 12x12 classifcation system, building towards the required coverage and accuracy. The IDQO service provides GIS's own data specialists and sourcing team more than 15 functionalities via an API endpoint and via an Excel integration, enabling them to expand and configure the system beyond the in-house 12x12, for example with vendors’ classification systems. Since last month the service is fully operational and already delivering business value by automating the tasks in a reliable way. As a side-effect, the solution helps the data team develop a thorough understanding on how data quality has a direct impact on the AI system’s performance along the way.
In the next phase, Faktion will activate, train, deploy and evaluate additional features in its IDQO solution, such as item matching on OEM level, brand recognition, and vendor identification. By leveraging this powerful solution, GIS International will be a first mover within its competitive landscape, leveraging AI tools and the Faktion partnership to streamline its operations and provide best-of-class service to its clients and suppliers.
Faktion's Approach to the Solution
The typical approach we use at Faktion consists of four steps:
Proof Of Concept (POC)
Proof Of Value (POV)
However, in the case of GIS, Faktion opted for a slightly different approach.
In the first phase, a POC on a subset of the data was conducted. Here, the POC showed that there was lots of semantic overlap between different categories. It was decided that the semantic overlap should be mitigated instead of building a sub-optimal classifier. Thus, we opted to use different classification systems in parallel, combined with additional techniques like annotation, to mitigate the problem and increase the accuracy, ultimately resulting in more business value creation.
Doing a POC shows its usefulness, revealing how the used classification model works, and serves as sanity check before moving further in the project.
When the POV was delivered by our engineers in phase two, GIS was already able to use it on a static version of the data. Certain business users were able to use the classification system guided or unguided. Moreover, a production user-friendly API was developed, together with the Excel plug-in, so GIS could already make use of the software.
Combining the second phase with operationalizing and scaling has some benefits, such as:
Having user validation, which can show room for improvement.
Increased adoption of the system.
Increased accuracy of the system.
The benefits mentioned above ultimately lead to a more easy functional adoption.
As in every project, some problems emerged during this phase as well. When upscaling from the subset to the full data set, the data did not appear to be representative for the full data. As a result, Faktion made use of brute force techniques to make the data representative.
In the next phase of the project (phase three), Faktion is currently activating new features within the platform, namely item matching, brand recognition, and vendor identification. On top of that, we keep close ties with GIS to see which additional features should be added as well.
As a result of the collaboration between Faktion and GIS to automate item categorization, a baseline accuracy of 88% for 12 main categories and their respective sub-categories was achieved. To further increase accuracy by 5%, input features were augmented with data enrichment using both the Bing search API for relevant keywords (used when abstract item codes are encountered) and a 'GPT in the loop' feature for text summaries (applied when items have a lot of unstructured metadata). Human annotators participated in the training phase through active learning, significantly reducing labeling time. A co-pilot of the ML models was also built, providing a trust score during inference with configurable thresholds.
Optimized Procurement as a Result
The implementation of the automated item categorization solution has dramatically improved efficiency and accuracy. Domain experts manually validate tasks that do not meet the set threshold, further enhancing the models' performance over time. This collaboration has helped GIS International's operations a lot, providing better customer experiences, more effective supplier matching, and the ability to handle a greater volume of clients without sacrificing quality or speed.
Moreover, this growing partnership helps to build GIS's AI capabilities. They are believers of our data-centric philosophy, and also notice that the better the data, the better the results. In the future, GIS can incorporate other data such as product data, transactional data, and vendor data for additional performance and enhancing the current performance.
Expanding the Partnership to Tackle New Data Quality Challenges and Opportunities
As GIS International and Faktion continue to build on the success of their collaboration, they aim to address other data quality challenges and opportunities in the procurement process. By leveraging the IDQO platform and AI, they plan to tackle issues such as item matching, supply chain data, and more. This ongoing partnership will ensure that GIS stays at the forefront of procurement innovation, continually optimizing and streamlining their processes to better serve their clients and suppliers. The collaboration is poised to set new industry standards and reshape the future of C-class item procurement.