|
This content has been archived, and while it was correct at time of publication, it may no longer be accurate or reflect the current situation at Microsoft.
When the treasury team at Microsoft wanted to streamline the collection process for revenue transactions, Microsoft Digital created a solution built on Microsoft Azure Machine Learning to predict late payments. Using a third-party algorithm, XGBoost, we spotted trends in five years of historical payment data. Aligned with our mission of digital transformation, these insights join data, technology, processes, and people in new ways—helping the collections team to optimize operations by focusing on customers who are likely to pay late.
Every year, Microsoft collects more than $100 billion in revenue around the world. About 99 percent of financial transactions between customers and Microsoft involve some form of credit. The company’s treasury team manages credit and collections for these transactions. The collection process involves all payments—not just late ones—so streamlining and refining a process of this scope is important to our success.
Speeding up collections has a big financial payoff. Considering the amount of revenue, you can safely assume that even small improvements in collection efficiency translate to millions of dollars. And the quicker we collect payments, the quicker we can use that money for activities like extending credit to new customers.
To detect who’s likely to pay or not pay—and adjust collection efforts accordingly—Microsoft Digital partnered with the treasury and finance teams at Microsoft. We brainstormed scenarios, questions, and solutions. We asked things like:
To help with these and other questions, we use data science and Microsoft Azure Machine Learning as the backbone of our solution. Azure Machine Learning is a cloud-based service that detects patterns in processing large amounts of data, to predict what will happen when you process new data. In other words, it helps us do predictive analytics.
Within two months, we easily set up a predictive model with Azure Machine Learning that helps the collections team prioritize contacts and actions. We use the eXtreme gradient boosting (XGBoost) algorithm—a machine learning method—to create decision trees that answer questions like who’s likely to pay versus who isn’t.
Data Science for Beginners compares an algorithm to a recipe, and your data to the ingredients. To train and refine the model, we overlay it with five years of historical payment data from our internal database. Output from the model, based on this data, helps us predict with over 80 percent accuracy whether customers are likely to pay late. We also get a valuable understanding of the factors or tendencies linked with customers who’ve paid versus those who haven’t.
Why is this understanding important? The insights we get fit into a broader vision of digital transformation—where we bring together people, data, technology, and processes in new ways to engage customers, empower employees, optimize operations, and transform business solutions.
We use past data and predictive insights from the model to:
The insights that we get help us to better understand our markets and to classify customer behavior in those markets. When we onboard new customers, we can correlate certain trends to them quite accurately, based on what we’ve seen with other customers. Examples include:
Table 1 shows what we used to do, compared to what we do now that we’re using Azure Machine Learning, for improving our credit and collections processes.
Table 1. Before versus Now
We weren’t as predictive or proactive.
These are the technologies and components that we’re using for our solution:
Azure Machine Learning Studio makes it easy to connect the data to the machine-learning algorithms. Figure 1 below shows the model that we built. We collect data from a variety of data sources and store it in our internal data warehouse called Karnak. Karnak contains historical information from SAP, Microsoft Dynamics CRM Online, MS Sales, our credit-management tool, and external credit bureaus.
We take this data and determine if there are other features that we need to build out of the data to improve the success of the model. This is called feature engineering, and we used this approach to create feature variables such as type of customer, customer tenure, purchase amount, and purchase complexity (products per order).
We then combine the data and engineered features into the machine-learning algorithm called XGBoost to get the late-payment prediction.
Figure 1 quickly summarizes our solution.
Different skill sets are used within Microsoft Digital to build out our machine-learning models. Figure 2 shows the iterative process that we use and the different roles employed at each stage.
The process works as follows:
Note: The decision tree in Figure 2 is for illustrative purposes only. We have more than 1,000 trees. The largest tree has 100 levels.
This produces one of two outcomes:
Azure Machine Learning also gives us a risk percentage score of how likely the customer is to pay on time.
Credit and collections team members often come across the same questions over and over. To speed up the process of answering these recurring questions, we built a chatbot. There are thousands of questions in emails, but there wasn’t a real tracking system.
The following steps, as shown in Figure 3, show how the chatbot works:
Now, field sales, operations, and collectors can see the latest information about customers they interact with and detect issues. For example, they easily see what the customer credit limit is, the overdue amount, whether a customer has exceeded the credit limit and is temporarily blocked, and answers to other questions.
We used Bot Framework and Azure App Service. The chatbot talks to App Service, and App Service talks to Karnak. Karnak data goes into Azure SQL Database, and App Service connects to SQL Database to answer the bot’s questions.
We use the XGBoost algorithm to create decision trees that look at features. For example, suppose an invoice is due on Saturday, or a customer in a particular country/region tends to pay late, and the average invoice is, say, $2,000. After we have the forest of trees that explain the historical data, we put new data in different trees. If most of the trees predict that an invoice will be late, we mark it accordingly. Otherwise, we mark it as unlikely to be late.
For customers with invoices that are due soon, the model shows which customers to prioritize. It puts their names at the top of a list for the collectors, so that they can contact these customers earlier in the process.
Here are some of the challenges that we initially had, but that we overcame:
To have the right data to put into an algorithm, you should have someone who understands the business processes and has good business insights. In our case, we had people with this knowledge and five years of historical data. We knew what business factors were important. But say you’re starting from scratch. If you don’t have someone who understands the business scenarios, and you don’t have much historical data, it’s harder.
Solving the machine learning problem itself took us only about two months, but deploying it took longer. There were lots of reviews and test cycles to demonstrate the accuracy and the high level of security that we have. If you’re doing something similar, build in extra time to allow for these cycles.
It’s unreasonable to assume you’ll get it perfect the first time. To get expected, consistent results, keep iterating.
We keep learning all the time as we iterate. Down the road, we plan to build on what we’re doing now. For example, we have integrated insights into several of our collection processes and some systems, but not all of them. Our approach is to incorporate changes to get the best return, and we’re still working on deploying these AI-based insights to everything we do.
We have also started to expand our scenarios into areas that are adjacent to credit and collections: sales and supply-chain features. We plan to add additional scenarios, use cases, data sources, and data-science resources for even more insights.
Also on our feature list is macroeconomic data, such as gross domestic product, inflation, and foreign exchange, to make our predictions even better.
Even small improvements in collections efficiency add up to millions of dollars. In an age of digital transformation, data and predictive insights are key assets that help us tailor our strategies and focus our efforts on what’s most important. With data science, Azure Machine Learning, and predictive analytics, we improve customer satisfaction, empower our collections team, optimize the efficiency and speed of our collection operations, and we’re more predictive and proactive.
|
Imagine having a personal assistant that helps you navigate your daily tasks effortlessly. Microsoft 365 Copilot offers just that, allowing you to work smarter, not harder. And the best part? You don’t need to be…
|
Microsoft 365 Copilot is a first-of-its-kind technology and demands a first-of-its-kind adoption process. We’re using new approaches to learning and development, community engagement and, of course, measuring success. Microsoft Digital, the company’s IT organization, has…
|
What this guide contains Intro: Deploying Microsoft 365 Copilot Chapter 1: Getting governance right Chapter 2: Implementation with intention Chapter 3: Driving adoption to capture value Chapter 4: Building a foundation for support Appendix: Downloadable…
|
Editor’s note: This is the fifth video in our ongoing series that showcases our in-house experts using real-world scenarios to explore the transformative capabilities of Microsoft 365 Copilot. Robert Gates likens Microsoft 365 Copilot to…
Imagine having a personal assistant that helps you navigate your daily tasks effortlessly. Microsoft 365 Copilot offers just that, allowing you to work smarter, not harder. And the best part? You don’t need to be a prompt engineer to use it. Here in Microsoft Digital, the company’s IT organization, we found 10 scenarios for using… Read more
Microsoft 365 Copilot is a first-of-its-kind technology and demands a first-of-its-kind adoption process. We’re using new approaches to learning and development, community engagement and, of course, measuring success. Microsoft Digital, the company’s IT organization, has been leading the charge on Copilot adoption; and measuring the value Copilot is driving for internal users has been an… Read more
This content has been archived, and while it was correct at time of publication, it may no longer be accurate or reflect the current situation at Microsoft. Data is the new currency of digital transformation. Whether it’s providing new insights, improving decision making, or driving better business outcomes, enthusiasm for unlocking the power of data… Read more
This content has been archived, and while it was correct at time of publication, it may no longer be accurate or reflect the current situation at Microsoft. It was time for a fresh approach to data analysis at Microsoft, one that would make it easier to track sales and operations activities across regions and roles.… Read more
This content has been archived, and while it was correct at time of publication, it may no longer be accurate or reflect the current situation at Microsoft. For almost 20 years, our Microsoft’s Commerce Transaction Platform (CTP) processed online payments through an on-premises environment, verifying that all transactions had been processed, sales had been finalized,… Read more