An AI workshop at PwC’s Brussels offices recently also featured practical use cases for AI in treasury from expert presenters from SAP and PwC who shared details of corporate projects they have worked on. After the event, Journeys To Treasury caught up with the speakers to obtain further details on these initiatives and to gather ideas as to how treasurers beginning their AI journey should approach working with the technology.
USE CASE 1: CONSOLIDATING AND FORECASTING FX EXPOSURES
Consolidating and forecasting FX exposures can be a challenge for treasurers, but it is one with which AI can assist. A presentation from PwC provided two use cases for AI in this topic. The first looked at how AI could be used more broadly with predictive modelling and hyperautomation to embed policies and automate the whole exposure management process. A second use case focused on AI’s large language modelling (LLM) part, which is the base layer used with generative AI (GenAI) and other tools.
The first use case under discussion featured a PwC client with a well-thought-out process for FX risk management. Despite this, most of the process took place on Excel spreadsheets, which led to considerable manual work.
Aniket Kulkarni, Partner, PwC, recalls: “They used to capture exposures and then check treasury policies, limits, and thresholds, and then make a deal request on their trading platform. It was a disconnected and extremely time-consuming process.”
This is where PwC saw an opportunity to use AI for hyperautomation, which could swiftly capture the FX exposures and embed the corporate client’s treasury policies in the process.
“If the exposure is above a certain threshold, it automatically generates a hedge request to the marketplace,” explains Kulkarni. “But if the exposure is below that threshold, then the model knows no action is required.”
This then linked to using predictive analytics for cash flow forecasting, as the corporate used its forecast for FX hedging rather than just managing balance sheet exposures. PwC used time series analysis with historical data and then built predictive models based on the company’s historical trends to help them forecast the data.
“One of the great benefits of this was, because their business is predictable, they could generate forecasts with an accuracy of around 90%,” Kulkarni reveals.
This project highlighted some challenges that any treasurer might encounter during their AI journey. The first fundamental challenge is the quality of the underlying data.
“Technology can help with the automation, but if the underlying data is not structured or in a proper format, then it is always a challenge,” emphasises Kulkarni. “A large part of our project was building the data foundation to feed the tool.”
A second learning related to the volume of data. Balance sheet exposures cover a huge amount of data, such as a company’s payables and receivables. All of these flows create a significant challenge when considering how to handle them more efficiently.
Fortunately, this project was able to answer the problem statement that the corporate started with, which was how they could automate and use technology to gain better insights into their data.
“We took the client completely out of Excel,” enthuses Kulkarni. “The previous process to monitor FX exposures took them half a day – and they ran the process daily. With this automation, it now takes them 15 minutes to complete the whole process.”
Saving time, speaking the right language
The second project in this presentation involved PwC working with a large treasury that had several people using the treasury system, including stakeholders from other departments.
“The client had their own GenAI installation, and the idea for this project was to use that specifically for treasury,” outlines Kulkarni. “We would use the existing infrastructure and train the model on treasury policies and standard operating procedures.”
The concept behind this was to create a ChatGPT-style tool that could answer users’ FX-related questions. For example, someone in treasury could ask how many currencies are in scope for a specific process, and the tool would give an answer of the permissible currencies in that process. Or, a user in another department could ask about the FX hedging frequency, and the tool could provide the answer, as long as the user has the correct authorisation.
“When dealing with thousands of users, this technology is useful because it saves a significant amount of time,” Kulkarni notes. “Otherwise, the treasurer would have to access a policy document or a standard operating procedure and try to find the right section to obtain this information.”
It could take minutes to find the correct answer manually, so on such a large scale, that time adds up. But with GenAI, the answer could be delivered within seconds.
“The greatest challenge to enable change in the transformation project was training everyone to use the new tool and technology, particularly as the process was running across multiple countries and regions,” recalls Kulkarni.
Another significant issue involved ensuring people could access the correct information at the right time.
“There was no knowledge repository or any particular tool to manage or distribute this information, so our solution was to use a large language model,” Kulkarni adds.
USE CASE 2: CASH FLOW FORECASTING
Predictive forecasting using ML and deep learning is a strong AI use case with the twin benefits of improving forecast accuracy while utilising fewer people on the treasury team.
Didier Vandenhaute, Partner, PwC, explains: “Cash flow forecasting uses extensive datasets and, historically, treasurers have used banking data for this. The transformative piece of AI is being able to use the transaction data along with that banking data.”
Additionally, employing AI in this process removes the human element in the middle of having to analyse and crunch the data. This frees up the treasury team to use and apply the outputs to their work.
PwC developed a proof of concept for AI cash flow forecasting in one specific case with a corporate client. This would prove the use case in the client’s environment with their data to show that it works.
Vandenhaute reflects: “We had to build predictive models in six weeks, and they tested those against their upcoming quarter actual cash flows, which we had not seen, to test our accuracy and prove that capability. We were predicting revenue and came in at 97% accuracy, compared with around 80% from their previous forecasting process.”
This corporate went all in on adopting AI, having hired a chief data officer, with teams of data engineers and data scientists building this capability out across the company. And yet, one of the hurdles in this particular project was around the change management process.
“Even though the model is more accurate – 97% compared with 80% accuracy previously – getting people bought in is incredibly hard,” outlines Vandenhaute. “Bringing people to that comfort level is a huge learning curve. The amount of upskilling and education sessions needed to reach that point is more effort than building the models themselves.”
A cultural shift is required within corporates that want to embrace AI. Another example of this is on the building side and finding explainability with deep learning. “Data scientists can come in and create these models that generate great forecasts, but we want to explain why there’s a delta and ensure that treasurers can extract detail and insights from those models,” adds Vandenhaute.
Moving on from Excel as a source of truth
The data challenge from the first use case also applies here, particularly regarding how much relevant data lives outside of a TMS and how that is managed in an automation project such as this.
Vandenhaute explains: “We’re fixing lots of data on the back end to develop these models and adhere to our timelines, moving with pace. However, we also have to fix many upstream processes in the business to ensure that the data is getting put into systems and flowing to us. Our goal is not to have our source of truth in an Excel file.”
This use case highlights how AI could be used in cash flow forecasting to save treasurers significant time and effort by automating previously manual processes. The improved accuracy and the timeliness of the forecasts also put treasurers in a far better position to manage their cash in any economic environment.
Today, the combination of technology and process is starting to get to the point where GenAI can be hooked up to the data, and the cash forecasts enable users to extract and interrogate that information. That is transformative, as is the connectivity of that data back into the organisation.
“The role of the treasurer continues to be elevated and now, not only do they have the data, but they can start to have explainability about these results back to the business partners,” underlines Vandenhaute.
USE CASE 3: EMBEDDING AI INTO BUSINESS APPLICATIONS
One expert AI use case presentation at the AI workshop provided a technology vendor’s perspective. SAP delivered a presentation outlining how the software provider is striving to embed AI into its business applications. Data is vital for treasury use cases, where the more data that can be brought into play from a business process, the better the outcome.
Christian Mnich, Vice President, Head of Solution Management, Treasury and Working Capital Management, SAP, explains: “Business data related to treasury is not necessarily owned by treasury, but is needed to deliver useful results in AI scenarios. In a typical cash flow cycle, you have to deal with major data points related to invoice to pay and invoice to cash.”
The data from these processes is not owned by treasury, but it is essential for interpreting the cash flow in terms of where the cash is coming from, the payment behaviours of clients, and whether the company has any influence over whether it can make specific payments at a later point in time.
“The foundation of data provided by the ERP is fundamental,” Mnich asserts. “A good ERP captures every business transaction, and that information can be brought into context. Conversely, a typical cash management system tends to show only the cash flow plus and minus, with a reference to an underlying business event.”
Without visibility of the underlying business event behind a payment, treasurers’ ability to take action based on the information presented to them will always be limited. The fact that AI embedded into treasury technology can provide this is a huge step forward.
“It all starts by having an idea of what information is relevant and how reliable it is,” notes Mnich. “AI ethics are implicit here as well, as we want to ensure that all the compliance and regulatory topics are being captured when it comes to issues such as data protection.”
For example, if a technology provider were to hand over its system to a large Belgian corporate, and that company has all the accounts payable and customer and vendor data in the system, the vendor couldn’t use this data to make AI use cases without special legal approvals. The non-disclosure agreement (NDA) would prohibit them from using that data. Instead, a specific agreement between the tech company and the corporate is required and it must spell out this has been requested.
“This is an essential point, but there are so many players out there that currently ignore it,” Mnich stresses.
Adding a copilot to the team
SAP gave four scenarios relevant to corporate treasury where the software firm has embedded AI into its business processes. The first relates to process automation, which is extremely effective in taking data from A to B, consolidating that data, and enabling treasurers to overcome challenges regarding certain non-automated processes, such as manual bank statements.
“I met a client recently who reflected that we have switched off manual bank statement processing at SAP,” Mnich recalls. “They said they want it back because now they have 100 people in Asia doing nothing else except typing in manual bank statements. This is something that can be easily automated with AI or RPA.”
Another process where AI can provide valuable assistance is with early warning indicators (EWIs). Mnich gives the example of a treasurer expecting information back from a high-value payment issued to a banking partner. If there is no response within a specific time window, the treasurer can receive a warning alert and take action accordingly.
SAP is also combining an aspect of ML with a rule-based framework. This can be used in processes where rules are applied and, over time, the machine takes over and uses intelligence to elevate those activities.
“Treasurers can achieve a better reconciliation rate based on machine learning, for example,” Mnich enthuses. “The tool can identify if there is a mistake somewhere in the process, fix it, and let the treasurer know what it has fixed.”
The final use case involves using LLMs embedded in business applications. Treasurers can actively discuss issues with the system and ask for a response. SAP has created its own AI copilot, called Joule, which can assist in various dimensions. It can provide business context for certain challenges on which the system user requires clarification.
“A treasurer could ask Joule how to manage a financial spin-off, and the copilot would guide them through the elements they have to complete,” explains Mnich. “Or perhaps a treasurer is looking for capabilities in the system to solve a business issue, for example they need to run a cost centre report. The system will respond by showing exactly where they can find it and how to run it, and could even provide the report as requested by the user.”
This also opens up the possibility for treasurers to create reports by asking the copilot questions. This would generate a report and enable the user to put it into perspective.
“The generated report could be about sales data, but to match it to the target audience the user could ask for information about the population in Switzerland, and then obtain usage data around that, and can combine external data with internal data as well,” Mnich adds. “This makes a huge difference.”
Keep calm and avoid ‘big bangs’
The practical use cases presented at the recent AI workshop underline what AI can do for treasurers already using the technology. After the event, the speakers also shared some suggestions for treasurers contemplating the start of their AI journey.
For PwC’s Kulkarni, establishing the treasury data foundation and knowledge repository are two early steps to success.
“It often comes as a surprise how fragmented and distributed the various documents are within the treasury set-up,” he reveals. “If treasurers invest time in their own data warehouse, that would be a good starting point for employing AI and other technologies.”
Vandenhaute highlighted investment in digital upskilling as a vital step for treasurers to consider early on their AI journey. He also emphasises the importance of being comfortable running some AI proof of concepts.
“Treasurers could do a proof of concept in their environment, or outside in a partner’s environment, to analyse their datasets and carry out an ‘art of the possible’ brainstorming on use cases. That demonstrates the value and they can start to become comfortable with the technology, which opens the door to even more issues that they could tackle.”
Brown agrees that the proof of concept is excellent for gaining stakeholder buy-in across the organisation.
“A proof of concept is relatively quick in a timeline perspective, it’s relatively cheap to carry out, and it helps calm the nerves around the reality of AI,” Vandenhaute enthuses.
SAP’s Mnich drew attention to the fact that business AI requires a solid combination of reliable data with the connected business processes.
“Without that, the user may get an output which is not good enough to be used to trigger decisions,” he warns. “GenAI models can be manipulated, so if you are not careful enough or provide the wrong environment, it could lead to issues.”
Combining the business data with underlying processes requires treasurers to have an accurate overview of where they are in their finance operations and how the system supports them. At that point, they can consider where AI can be applied.
“We have been working for more than six years on use cases to improve cash flow forecasting by applying predictive models,” Mnich outlines. “Based on the usage data we receive from clients, we know that to have a useful and reliable output, tons of data is required – particularly in certain industries and geographies – so as not to make the wrong decisions.”
One point stressed by all of the presenters at the AI workshop was that treasurers do not need to worry about suddenly inserting AI across the breadth of treasury processes in a ‘big bang’ style of implementation. It is possible to start with simple use cases, considering where the most manual interactions or time-consuming efforts exist that could be optimised by applying AI to a small project.
“Don’t try to reinvent the wheel but rather have practical use cases and examples that will be quick wins for the organisation when applied,” concludes Mnich. “Talk to your peers to learn about such scenarios.”
