Forecast for A/R: Dramatic Improvement in Efficiency

And the winner of the 2023 Gold Alexander Hamilton Award in Working Capital & Payments is ... Microsoft. Congratulations!

Reprints

Accounts receivable (A/R) at Microsoft is massive. The company’s Global Treasury & Financial Services (GTFS) group is responsible for A/R for almost every business unit throughout the global organization. That means the team handles receivables on millions of invoices, for thousands of customers, each month—and those numbers continue to grow. In fact, as Microsoft Azure and other innovative lines of business have taken off over the past five years, the company’s A/R volume has nearly doubled, and it shows no signs of slowing down.

The GTFS group outsources collections for all those receivables to Accenture. A couple of years ago, members of the Accenture team were preparing receivables forecasts using manual processes that were not sustainable considering Microsoft’s growth trajectory.

“The collectors were all managing their own forecasts in Excel, so it involved a lot of people and took away time from core collection activities,” explains Apra Bothra, senior finance program manager in Microsoft’s GTFS group. Each collector was assigned specific accounts. “They would go through their accounts on a weekly basis to determine which receivables were forecasted to come in and which of those were at risk. Then they would look at how their at-risk receivables might affect the ledgers, and they would prioritize accounts for their collection activities.”

It was a task that took about 1,000 hours a month across the Accenture team, and “the result was not 100 percent accurate” in terms of predicting which receivables would be paid late, says Helene Collignon, senior delivery lead for Accenture. “On average, the accuracy rate was about 90 percent globally, and sometimes a little lower regionally. And that would sometimes generate confusion about why we did not meet the forecast. We would end up performing root-cause analysis and discuss reworking the process.”

Another challenge, Bothra adds, is that the process relied on the expertise of each collector, making it difficult to transition the work to a colleague when needed. “Collectors would forecast risk for a receivable depending on their experience with that customer,” she says. “But if the key collector was not available, somebody would need to fill in, and they would have to manually go through that customer’s entire history. That was not efficient, and it was taking people away from their core activities.”

Collignon had previously worked with Accenture’s data science team to improve collections processes, so she approached Hongrui Gong, the firm’s principal director of data science, about the possibility of building a solution that would more efficiently predict whether a customer was likely to pay. “As soon as Helene explained the existing workflow, we thought we could help reduce the manual effort by building an intelligent collections forecast,” Gong says. Microsoft signed off on the project, and the data science team got to work.

“It was a challenge,” Gong says, “because at the same time we wanted to automate the process, we also wanted to make it as accurate as possible. If we produced an inaccurate A/R forecast, it might not be useful to the business. We decided to use multiple advanced machine learning technologies to achieve the highest accuracy level.”

Gong’s team determined that a single machine learning algorithm, working alone, would not solve the problem, so they began experimenting. “We tried a variety of different technologies to see which one produced the best result,” Gong says. “Then we combined the results together to achieve high accuracy.” The forecasting process they came up with is divided into two parts.

The first part focuses on Microsoft’s current outstanding invoices, relying on a custom-built machine learning solution to determine how many outstanding invoices the company has, and then predict which ones will be paid by the end of the month. It provides these forecasts at multiple levels of data—the individual customer level, the product level, the business unit level, the regional level, and the global level. Gong’s team began developing a machine learning model that would draw on customers’ payment behaviors over the past few years, and they experimented to determine which AI technologies should drive the model’s forecasts.

“We simultaneously trained the model using different machine learning technologies, including neural networks, time-series analysis, and LSTM [long short-term memory],” Gong explains. Neural networks use interconnected nodes, similar to human neurons, so that computers can process data in an approach fairly similar to the human brain. LSTM analyses combine long-term and short-term behaviors of the past to better understand and predict future behaviors. And time-series analyses, as the name implies, enable analysts to determine trends, seasonalities, and other patterns in activities.

As Gong’s team evaluated alternative artificial intelligence (AI) models, they compared the results with actual historic payments data “to see which one would give us the best result for each line of business, region, etc.,” Gong says. “We picked the best technology for each of the different levels, we optimized the parameters for each of them, and we combined them together to achieve the most accurate result possible.”

The other part of the forecast focuses the same AI tool on new billings. “We also forecast how many new billings will be created over the course of the month, how many of those will be paid by month-end, and how many will remain open. We forecast  that at the business level and regional level, and run these forecasts four times a month,” Gong says.

“Once we have both parts of the forecast, we can determine which strategy the collections team should use to try to get at-risk customers to pay on time,” he adds. The combined forecast is displayed for collectors and management alike through easy-to-use PowerBI dashboards. Gong’s team continues to tweak the scoring engine to improve its accuracy, but this approach has already significantly improved Accenture’s ability to predict which accounts will pay on time—and which will not. “In most of our internal studies of the new process, we have achieved 97 percent to 98 percent accuracy,” Gong says.

That, in turn, dramatically improves the effectiveness of collections activities. Accenture collectors use the output of the AI solution to prioritize and plan their daily activities. “The A/R forecasting tool is saving our collectors a huge amount of time that they would otherwise spend putting forecasts together,” Collignon says. “Our collectors are absolutely delighted with the solution; it’s changed their lives. And from a strategy standpoint, it helps us identify which accounts we need support on, whether that is assistance from Microsoft GTFS, a sales team intervention, or something else.”


See also:


 

This project has reduced Microsoft’s dunning efforts, because collectors have better information, earlier, about which customers they should focus on, and because the collections team has more time to focus on their core competency. As a result, it has accelerated payments by almost $300 million of A/R annually for Microsoft.

In addition, the tool gives the Accenture team an earlier warning about accounts that may end up needing to be written off as bad debt. “If the forecast predicts that an account will absolutely not come in, then we can identify other opportunities to make up for that loss of revenue,” Collignon says.

The initiative has dramatically improved the collections team’s effectiveness, and Bothra sees the development of the AI tool as Microsoft’s most efficient means of achieving that objective. “We had multiple options for enabling machine learning capabilities,” she says. “We could have plugged these capabilities into our ERP system, but then the internal engineering team who support our ERP system would have had to develop that functionality. Another route we could have taken would have been to create templates or macros in Excel. But that would have perpetuated some of the challenges we were facing because it would have required a lot of manual interpretation.

“Microsoft crossed the threshold of $200 billion in revenues last year, so we needed a technology that could scale,” Bothra continues. “The AI solution enables us to grow our receivables without adding more A/R staff. They can focus on their core job, while the tool performs the rote tasks for them.”

She adds that removing the human element from forecasting also has compliance implications: “Because the machine learning program is taking data directly from our subledger, we are not relying on individuals to manually feed in data. The data-ingestion process, data analysis using Power BI, and other elements of our A/R forecasting are automated, so we’ve removed the risk of human error in the entire forecasting process,” she says.

“Everybody’s talking about AI these days, and this project demonstrates how AI technologies enable you to do more with less,” Bothra concludes. “We didn’t have to invest in huge and complex systems, but we were able to leverage the power of machine learning to build A/R capabilities that are far more efficient than what we had before.”