AFP Executive Guide to Emerging Technologies Part 2: Artificial Intelligence and Machine Learning

Emerging technologies are changing finance functions from the top down. In particular, treasury and finance executives must evaluate how treasury technology can enhance their overall operations.

Read part two of AFP’s Executive Guide to Emerging Technologies, underwritten by Kyriba, to learn how treasury and finance are deploying artificial intelligence (AI) and machine learning (ML) in their processes. While these technologies have many uses in cash and treasury management, this eBook explores treasury use cases for:

• Actual reconciliation
• Streamlining collections
• Data consolidation
• Fraud detection

Be sure to also explore part one of this series, AFP Executive Guide to Emerging Technologies: Robotic Process Automation.

Emerging Technologies Are Changing Finance Functions

Emerging technologies are changing finance functions from the top down. In particular, treasury and finance executives are being forced to evaluate how robotic process automation (RPA), artificial intelligence (AI) and machine learning (ML) can enhance their operations overall.

This two-part Executive Guide, underwritten by Kyriba, explores why and how these technologies are reshaping finance. In this section, we will look at how treasury and finance are deploying AI and ML to improve their processes.

Artificial Intelligence and Machine Learning

A key limitation of RPA is that it is not actually “intelligent.” RPA does what it is told. AI, in contrast, uses machine learning so that it can essentially think for itself. The software learns, without human intervention, by analyzing data. It can be used to develop new rules, instantly discover exceptions and build forecasts.

RPA is like an Excel macro; it is automation that mimics what the user tells it to mimic, noted Bob Stark, vice president of strategy for Kyriba. Treasury management systems (TMS) and other types of financial platforms don’t typically rely on RPA within their product, but rather support their customers’ use of robotic process to automate the interaction with other systems. “Machine learning, on the other hand, learns from the data that it receives within the treasury system so has a natural role within a TMS,” he said.

But although AI has incredible potential to improve many processes for treasury and finance, it has yet to catch on, noted Jason Dobbs, senior manager, and Kyle Olovson, CTP, senior consultant, both with Actualize Consulting. They see this as irrational, particularly since humans have the ability to only recognize a few patterns. Machines, meanwhile, can pick up on patterns indefinitely. However, recognizing patterns does depend on human design, which means that AI can contain biases that can skew the outcomes.

Dr. Rana el Kaliouby, CEO and co-founder of Affectiva and MindShift keynote speaker at AFP 2019, sees bias as the true “threat” that AI poses. She noted that while many people are apprehensive about the technology, what we should really be wary of is AI that is not ethical—systems that perpetuate biases existing today. Therefore, companies that utilize this technology should take care to steer clear of biases whenever they adopt these systems.

In an interview with the AFP Conversations podcast, Dr. el Kaliouby explained that the key to avoiding data and algorithmic bias is diversity. “I think the way we get there is to ensure that the teams that are designing and deploying these AI systems are as diverse and inclusive as possible,” she said. “At Affectiva, at my company, we make sure that we bring in people from different age groups, different genders, different ethnic groups, and different perspectives. We have our machine learning scientists, our data scientists and our computer vision folks around the table, but we also bring artists into the equation or psychologists because I believe they have a very different perspective on how we think about this, and that’s really important.”

But for the treasury and finance space specifically, Stark is less concerned about bias affecting AI and ML. Unlike RPA, AI and ML aren’t based on what the user thinks, but on what the artificially-intelligent process is able to glean from the data. “I think in a treasury use case, introducing bias is more likely when using RPA because the ‘bot is automating the steps and finding the data that the user thinks is correct,” he said. “In treasury, at least, machine learning programs allow the data to drive the decision, and not you as a programmer. Machine learning can help you overcome your own bias by potentially finding a very different, objective conclusion.”

Stark noted that if, in the initial setup of an AI/ML program, an organization sets certain parameters or thresholds before it has adequate data, then it’s possible that biases could impact the outcomes. “But if you do it right, data should drive your decisions rather than the other way around,” he said.

Treasury Use Cases

There are many treasury use cases for AI, and even more will likely be revealed as practitioners familiarize themselves with the technology and what it can do.

Actual Reconciliation

AI and ML have incredible potential for cash management and forecasting, particularly when reconciling prior day bank files with yesterday’s expected cash position. “This is one of the first cash management processes performed each day,” Stark said. “And for some organizations, the volume of transactions is so big that it can take hours and multiple people to do that reconciliation.”

ML can be used to identify and resolve those discrepancies on its own. “In the simple scenario where the prior day file reports a $1 million wire and we thought it was going to be $900,000, the cash manager will know through their experience what explains that $100,000 difference and what to do about it,” Stark said. “Machine learning will learn from the user’s manual reconciliation, so next time it will reconcile those transactions without human intervention.”

Added Stark: “The tests that we did with our clients reconciled 99% of previously unmatched transactions, just by looking at the past six months of data. With more data, the accuracy approaches 100%. The key, however, is the amount of data. Machine learning cannot ‘learn’ from one or two scenarios a day; it needs a large sample size of information in order to find meaningful conclusions.”

But AI and ML can do more than detect anomalies—they can recognize when an exception isn’t actually a problem. For example, your company might make a regular monthly payment to a supplier of approximately $10,000. However, a recent payment made at the end of the current month is $15,000. With rules-based automation or even RPA, you likely have a payment control that flags that 50% variance from the normal monthly amount, quarantining that payment for further review. But ML can recognize that this particular payment is part of a larger pattern where the last monthly payment in each quarter is substantially higher than the average. “ML looks at historic data to determine what the true pattern really is to reduce potential false positives,” Stark said. “But the AI process needs to have access to all historic payment patterns in order to make that determination, much like an individual would use their own experience to make that assessment.”

So you do not have a ‘one-size-fits-all’ solution for disputes, credits or cash application. You really have to use machine learning algorithms to identify each customer’s different payment patterns and deductions behaviors that they’ve had in the past.”

Streamlining Collections

AI and machine learning can also be used in the actual AR process as well. Shankar Bellam, manager, solution engineering for HighRadius, noted that one of the biggest challenges for AR teams is the huge amount of customer transactions that they have to manage. “In the B2B space, where you have hundreds of thousands of customers, you have millions of invoices and you really have to go after every transaction,” he said. “So that’s where a lot of time is being spent by having AR teams manually going after invoices and trying to get paid or trying to figure out why the customer has not paid on something.”

While RPA allows AR to automate the repetitive, manual tasks, machine learning algorithms can be used for more intricate work, like identifying patterns in transactions. This can help your organization in its decision-making and really tailor the behavioral change that you want to bring about in your customers, Bellam explained. “Maybe you want to change the way you collect from the customer, or maybe change the amount of manpower that you assign to a certain segment of customers,” he said.

HighRadius works with a number of top-tier companies, like Nike, Adidas, Johnson & Johnson, Walmart, and Proctor & Gamble. These companies each manage a massive amount of transactions, going across currencies in different regions. “So you do not have a ‘one-sizefits-all’ solution for disputes, credits or cash application,” Bellam said. “You really have to use machine learning algorithms to identify each customer’s different payment patterns and deductions behaviors that they’ve had in the past.”

Improving efficiencies in the collections space can have a positive effect on cash flows. “Because of the sheer volume of these organizations and the magnitude of the size—even if they’re bringing in one day of [days sales outstanding (DSO)] improvement—it’s huge in terms of dollar values for these companies,” he said.

We try to analyze the data historically to come up with insights; if you augment on top of the data you already have, you can make this smarter.”

Data Consolidation

AI can also help treasury as it consolidates copious amounts of data from ERP systems, TMS and other bespoke sources when doing cash forecasting. “Everyone cares about the accuracy in the end, but the process to get there is quite cumbersome in most cases—to get your hands on the data, then to mix all the data, and to make something out of it that makes sense,” said Nicolas Christiaen, CEO and co-founder of Cashforce, who discussed AI and ML in an AFP 2019 session. “And then there are the people involved; you don’t want to have 60 people tripping over each other to make that forecast.”

To create the forecast, treasury needs to consolidate the data correctly to make sure it is getting the right data sources from the TMS and ERP systems. To improve the forecasting process for its clients, Cashforce looks at historical GL (general ledger) data. “We try to analyze the data historically to come up with insights; if you augment on top of the data you already have, you can make this smarter,” Christiaen said.

He added, as an example, that attempting to forecast based on due dates for payments is pointless, as customers rarely pay precisely when you require them to. Hence why it is so important to look at the historical trends and patterns, if available.

The final step in Cashforce’s process is what is called its back-testing algorithm. Based again on historical data, Christiaen’s team looks at how good a system-based forecast measures itself against the actuals. “We’re trying to understand the historical deviation of your system’s forecastable data versus the actuals to come up with a segmented variance, and then inject that on top of the current forecast,” he said. “To summarize, I would say it’s using different smart data sources and different smart algorithms which will optimize the accuracy of your forecast. With the back-testing algorithm, we’ve just scratched the surface, in my opinion, of what we can potentially do.”

This works by the model having an understanding of what is ‘normal’ for each account or card and recognizing patterns based on past transactions and behaviors. For example, if 99% of the transactions for one account happen Monday through Friday, a transaction that occurs over the weekend will be seen as abnormal and flagged as such.”

Fighting Fraud

Financial institutions have gradually been adopting AI and ML solutions to protect customer accounts. AI can help banks and their corporate customers keep track of fraudulent activity and anomalies much faster than ever before, explained David Duan, data science stream lead and principal data scientist at Fraedom. “This works by the model having an understanding of what is ‘normal’ for each account or card and recognizing patterns based on past transactions and behaviors,” he explained. “For example, if 99% of the transactions for one account happen Monday through Friday, a transaction that occurs over the weekend will be seen as abnormal and flagged as such.”

Duan noted that anomalous transactions aren’t always fraud incidents. However, using AI to flag any transactions that are out of the ordinary is worth the hassle of slowing down unorthodox legitimate transactions. Earlier this year, Visa helped banks prevent approximately $25 billion in payments fraud using AI, reducing global fraud rates to 0.1%.

The Visa Advanced Authorization (VAA) is a risk management tool that monitors and evaluates transaction authorizations on VisaNet in real time to help banks identify and respond to emerging fraud patterns and trends. Visa processed more than 127 billion transactions between merchants and banks last year, using AI to analyze all of them in about one millisecond per transaction, allowing for legitimate transactions to be processed and fraud to be quickly rooted out.

Of course, there will always be bad actors that try to use innovative technology to their advantage, and AI is no exception. Micheal Reitblat, CEO of Forter, noted that criminals are using AI to mimic the way good users behave. “So instead of collecting static data like credit cards and addresses and getting access to physical locations, they are actually recording user behavior,” he said. “They understand that AI is being used to track fraudsters, so they need to fight back with a similar technology.”

Furthermore, the good actors are actually at a disadvantage. Although they have more data at their disposal now than ever before, they don’t have enough “bad” data, Reitblat explained. “If there’s a fraudulent transaction, we can’t collect a million of them just to learn a better model,” he said.

Additionally, using AI to make predictions on fraud isn’t like predicting normal customer behaviors, because fraudsters adapt. They are constantly fighting the efforts of the good actors; the moment they understand that AI is being used to identify and catch them, they change their behavior. Once that happens, the AI model needs to learn from scratch again.

KYC Compliance

Complying with know-your-customer (KYC) regulations is one of the biggest headaches for corporate treasury departments. AI may be able to relieve some of that burden.

In an article earlier this year, Eurofinance pointed out that AI has the ability to replace the current system of “scattered humans and fragmented legacy technology” used to ensure KYC compliance. Even smaller businesses may be required to evaluate a multitude of tax and legal updates on a weekly basis across an international network, so using RegTech solutions like natural language processing and machine learning can help immensely. “By treating regulations as data, software will dynamically help ensure compliance and bring compliance into the enterprise risk environment, enabling treasurers to take a genuinely risk-based view of regulatory compliance,” Eurofinance explained.

And KYC repositories are also making use of AI. As noted in a previous Executive Guide, IHS Markit has implemented AI in its KYC Services solution. Using intelligent process automation (IPA), the system gathers and analyzes large amounts of complex information from multiple sources, such as corporate registries, exchanges and regulator sites, to build a visual picture of an entity. The solution, which was developed by encompass corporation, has enabled IHS to improve data quality and achieve a 30% reduction in the time taken to gather KYC data.

To date, KYC Services has developed more than 30,000 KYC profiles and has more than 160,000 entities represented on its counterparty platform.

Financial Planning and Analysis

AI and ML have also found a home on the FP&A side, and the technology has actually been there for some time. Dr. Liran Edelist, President of Jedox Inc., who discussed AI at AFP 2019, noted that although the technology is often seen as experimental, many of the algorithms used for budgeting, planning, and forecasting are already in use and were proven decades ago. “The innovation is around the availability of such technology,” he noted in a recent blog. “Just a few years ago, you had to purchase expensive hardware and software and hire a data scientist to build a model that today is most likely available out of the box.”

Dr. Edelist pointed out key areas where AI algorithms may help FP&A professionals in their planning and budgeting. One is time series forecasting and prediction.

“What we’re seeing is that many activities, both expenses and revenue-related in our organizations, can be automated by machine,” he explained. “The most common use case for AI is for revenue planning—both budgeting and forecasting. However, standard costs in an organization are predictable. Planners can ask the machine to preform, as an example, time series prediction. The system will do it, in most cases, in a better way than what human beings can do.”

Although Jedox does not collect statistics that can show how much AI solutions can improve these predictions, users generally see AI making better predictions than humans. One of those clients is Terminix (part of ServiceMaster Group) which has used the technology to greatly improve customer loyalty. “The information that the AI is giving them really allows them to improve customer loyalty,” Dr. Edelist said. “They’re able to see what enables them to keep loyal customers, and on the other hand, why customers don’t return. By nature, it improves the recurring revenue to the company.”

Another area where FP&A can use AI is in data cleansing. AI algorithms can indicate a potential abnormalities in data trends that manual methods typically won’t pick up on. “We’ve seen [abnormalities] happening a lot, especially when you need to consolidate data from different systems and integrate it between different systems,” Dr. Edelist said. “And so, for the user, it’s very difficult to go looking for each individual data point on your database, but with the AI algorithm, you can actually find some discrepancies.”

But much like in treasury, it’s not just anomalies that can be identified. The algorithm can also pinpoint trends that users might not be able to see. “Some of those discrepancies are really showing a change in trends,” Dr. Edelist said. “The algorithm can give a clearer picture of the variances, which makes the analysis easier.

Where RPA replaces mouse-clicks and Excel macros, AI actually helps the decisionmaking process start sooner. It replaces a lot of collection and aggregation of treasury data, which often consumes many peoples’ mornings, and allows the strategic part of your job to start at 8 a.m. instead of later in the afternoon. You get so much more done when you don’t have to spend your morning finding and reconciling data. It’s not only an improvement on manual processes; it’s an improvement on automation. And anyone that doesn’t recognize that isn’t going to be here in five years.”

Taking the Next Step

Artificial intelligence and machine learning have the potential to change the game for treasury and finance, in a way that RPA never can. This is because AI and ML software learns, without the need for human interaction. RPA is essentially process decision-making, where is AI/ML is data-driven decision-making, and the latter requires a mind shift for many treasury and finance departments to truly get on board with it.

Additionally, as is the case with RPA, treasury and finance professionals may be apprehensive about AI/ML because they fear that their jobs could ultimately be phased out. But it is much more likely their roles will simply need to evolve with the technology. The more familiar they become with it, the better off they’ll be.

“Where RPA replaces mouse-clicks and Excel macros, AI actually helps the decision-making process start sooner,” Stark said. “It replaces a lot of collection and aggregation of treasury data, which often consumes many peoples’ mornings, and allows the strategic part of your job to start at 8 a.m. instead of later in the afternoon. You get so much more done when you don’t have to spend your morning finding and reconciling data. It’s not only an improvement on manual processes; it’s an improvement on automation. And anyone that doesn’t recognize that isn’t going to be here in five years.”