Procurement control tower: Proof of concept through machine learning and natural language processing

Editor’s Note: The SCM thesis Procurement Control Tower: Proof of Concept Through Machine Learning and Natural Language Processing was authored by Bishwajit Kumar and Pablo Barros Gomez, and supervised by Dr. Elenna Dugundji ([email protected]) and Dr. Thomas Koch ([email protected]). For more information on the research, please contact the thesis supervisors.

Our capstone project sponsor, a global pharmaceutical company, faces significant challenges in its procurement processes due to its large procurement spending, diverse product needs, extensive supplier base, and divergent software insights. The company recognizes that to remain competitive in today’s volatile, uncertain, complex, and ambiguous (VUCA) market, it will be imperative to gain insights faster, enhance decision-making capabilities, and optimize exception management. As a potential solution, the company wants to explore the value proposition of implementing a procurement “control tower:” a centralized platform that offers end-to-end visibility and control over procurement processes.

To address our sponsor’s objectives, our study aimed to answer two key questions:

1. Would a procurement control tower create measurable value for the sponsor’s procurement functions?
2. Could we demonstrate the proof of value of a procurement control tower by creating a prototype of one of its use cases?

Twofold research: qualitative study and quantitative analysis

Our research followed a two-step process. First, we conducted a qualitative study to define the scope, value proposition, and deployment strategy of the control tower. We interviewed subject-matter experts from various procurement processes to understand their existing challenges. Additionally, we investigated industry best practices and aligned with our sponsor on the specific use cases that the procurement control tower would cover, such as spend analytics, contract management, risk management, and supplier management.

In our qualitative research, we proposed the overarching architecture of the procurement control tower and outlined its value proposition to our sponsor. The first crucial step in implementing the control tower is to consolidate data from various data sources into a common data layer. This convergence ensures a single version of truth (SVOT) of data, serving as the foundation for the control tower. Unified data and information retrieval becomes easier and eliminates discrepancies arising from differences in source data. The unified data enables enhanced data analysis from a single source, empowering the procurement control tower to generate valuable business insights.

Second, we performed a quantitative study to develop a prototype (proof of concept) focusing on the spend analytics use case, specifically spend categorization of materials. This use case holds immense value for the sponsor, since approximately $250 million worth of spend data remains unclassified in terms of accurate category or subcategory classification. This lack of accurate classification of spend hampers business analysis based on spending categories, thereby increasing the potential for inaccuracies.

For our quantitative study, we compared multiple machine-learning algorithms, including logistic regression, decision trees, random forest, and XGBoost, using our data to predict the right categorization of materials for unmapped spend. After careful evaluation, we selected Random Forest as the best-performing algorithm in terms of accuracy. To further enhance the algorithm’s predictive power, we preprocessed the data using natural language processing (NLP), a computational technique designed to mimic a human-like understanding of text. The final algorithm achieved a 94% classification accuracy at the category level and 90% at the subcategory level for the unclassified spend data.

Advanced insights and benefits of implementation

Implementing the procurement control tower will provide advanced insights to our sponsor, bringing them end-to-end visibility, enhanced exception management, improved decision-making, improved risk management, cost savings, and more. The categorization of the unmapped spend of materials by the machine-learning algorithm will have a positive impact on our sponsor’s business in various ways. Specifically, it opens up opportunities for supplier renegotiations, improves budgeting accuracy, and reduces the man-hours required for manual categorization. Given that our sponsor adds thousands of new SKUs each year, which translates to tens of thousands of spend data records, our proposed solution becomes highly valuable as it offers an ongoing, periodic categorization of spend data. Our proposed solution has been accepted by our sponsor, and its implementation is underway, marking a significant step toward optimizing procurement processes and achieving competitive advantages in the VUCA market.

Every year, approximately 80 students in the MIT Center for Transportation & Logistics’s (MIT CTL) Master of Supply Chain Management (SCM) program complete approximately 45 one-year research projects.

These students are early-career business professionals from multiple countries, with two to 10 years of experience in the industry. Most of the research projects are chosen, sponsored by, and carried out in collaboration with multinational corporations. Joint teams that include MIT SCM students and MIT CTL faculty work on real-world problems. In this series, they summarize a selection of the latest SCM research.

SC
MR