AI: Reshaping the Treasurer’s Role

AI captured the zeitgeist of the past 12 months, generating endless headlines worldwide. For treasurers, AI has the potential to offer a new level of automation and to be a new co-worker on the team. Still, treasurers must also navigate challenges around data management and privacy.

While AI has been on the treasury technology radar for many years, nascent generative AI tools such as ChatGPT have opened up a new range of potential use cases. But for treasurers looking at taking advantage of AI in their daily tasks, there is much preparation work to do.

Bob Stark, Global Head of Market Strategy, Kyriba, comments: “The core issue for treasurers to understand is that you don’t have an AI strategy until you have a data strategy. You can’t plug in this tool without having access to a broad range of information to train the model.”

A key piece of technology that goes hand in hand with AI in this situation is the API, which plays a vital role in unifying data across the enterprise. Whether it’s a data warehouse, data lake or TMS, treasury teams must utilise APIs to ensure access to all relevant data.

“Treasurers looking to use AI to improve their cash forecast, for example, need APIs to bring together all the sources of information,” adds Stark. “This structured data then enables them to learn what has happened in reality to train the AI to make better predictions for the forecast.”

Understanding where the data that an AI tool is going to use will come from – and which elements of the data are reliable or not – puts treasurers in a position to start training the intelligent capabilities of AI.

Mining hidden data sources

Mario Del Natale, Treasury Director, Global Digital Treasury, Johnson Controls, has been refining the company’s data strategy since 2008. “We have huge quantities of data in our database from various treasury systems, so the first question we asked ourselves was what we wanted to get out of that,” Del Natale reveals. “That’s not only the data that the user can see on the screen, such as a payment notification or a bank statement, but also hidden data. We need to include this when thinking about what data can make treasury life easier from a reporting and a dashboard point of view.”

Thinking about how to structure the company’s data, the first task that Del Natale and his team carried out was to create a naming convention. This provided a structure of code for any data that was generated.

“Then we also created additional data linked to that data, enabling us to enrich our existing or standard data sets to improve our reporting even further,” Del Natale reflects. “We explained this all in detail to the entire treasury team because, back in 2008, people didn’t understand why we needed to have a naming convention.”

At Johnson Controls, the data was structured in a ‘bottom up and top down’ approach, providing detailed data for the analyst and aggregated data or country data for the manager. The solution enables users to drill down to the most granular levels of the static data in a real-time approach.

“We have strict policies that people must follow to create the static data and an approval process to ensure that it is created correctly,” adds Del Natale. “We also have some annual objectives and KPIs for those in charge of the static data for the same reason.”

The whole approach implemented by Del Natale stems from the idea of streamlining treasury processes by automating and integrating data coming from multiple sources. This includes structuring the data and paying close attention to and establishing strict policies around the static data setup.

“That approach has been fundamental for our digital landscape today,” Del Natale shares. “Treasurers should view static data like the basement of their house – without a good one, everything will collapse.”

Identifying AI use cases

When looking to deploy AI in treasury, a vital first step is for the treasurer to identify practical use cases for their working environment. For example, there could be a manual step in a specific process where a member of the treasury team has to look at certain data and then take action based on that data. That process could be a candidate for applying AI.

Jan-Willem Attevelt, Co-Founder, Automation Boutique, details: “A treasurer may be reading bank statements every day, looking at the individual cash flows and trying to categorise them. This process could be used to train an AI to classify the data for the treasurer.”

To get that process right, the AI model would need to be fine-tuned by giving it examples of expected outcomes when the data looks a certain way. This is something that AI can already do today for treasurers.

“To bring it to the next level, treasurers can think about fine-tuning an existing AI model, such as a large language model [LLM], and feeding it with their company data,” adds Attevelt. “This would need the corporate to create a training set of data, which means collecting data for the AI model to learn from and applying it when the treasurer is going to feed it with fresh data.”

Much careful consideration is required around what kind of data to put into the AI model in this case as treasurers think about what it is they want to accomplish from the project.

Stark reveals: “Probably the most popular question we hear from treasurers is regarding how they can improve their cash forecasting with AI. I’ll always ask them what they are trying to achieve. Are they trying to improve the predictions that they already have or are they aiming to generate brand-new data that doesn’t exist?”

The treasury team at Johnson Controls has implemented a Power BI application that includes AI tools to assist with cash management. The application provides a real-time liquidity position for about 3,000 bank accounts with connectivity to 250 banks and more than 150 ERP connections.

Del Natale recalls: “We deployed a logic-based technical solution that we found hard to acquire from a legacy software provider. As a treasury, we are fully agnostic from a software provider perspective regarding the coding and building of our dashboard.”

Suppose a brand-new entity or bank account is created somewhere within Johnson Controls. In that case, the solution does not require additional setup or new coding because the source static data is already categorised. This enables the company’s global AP and AR processes to be automated.

“We’ve taken an AI approach in our coding, meaning that we have created the script ourselves – it’s a staggered approach of SQL [structured query language] script,” explains Del Natale. “This means we can define whatever we want and even categorise items on the fly.”

This approach also enables the treasury to create non-existing flows and non-existing static data to aid with understanding of their positions. Should a specific rate or balance be unknown, the program will generate rates or balances based on what it can see in the bank statement, including calculation mechanisms driven by static data.

“We are also using out-of-the-box AI tools from Power BI, including the narrative and Q&A functions,” adds Del Natale. “Sometimes it doesn’t provide you with the right information, but you can train it to make that happen. We use many charts, which makes it even easier to understand the data.”

Charts can also highlight the discrepancies by using the static data in the model. It can also explain why a certain position might suddenly change, based on the static data. “We can also forecast certain items based on historical data, still using charts on which we can add some parameters and then see the charts moving around based on the data,” enthuses Del Natale.

This setup is used globally by the company’s businesses, up to the CFO, and Del Natale checks every month for new releases from Power BI to see what he can add to the current system.

“We will sometimes use a new functionality to improve our AI approach and dashboard,” he outlines. “Some treasurers still like Excel because they think they have mastered the data because they are manipulating it. But I would encourage them to try any system with AI tools, such as Power BI and Tableau.”

Welcoming aboard the co-pilot

While some rather bleak prophecies may suggest that AI will take the jobs of humans, treasurers have the opportunity to view AI tools, including generative AI, as a colleague who helps get the job done more efficiently.

Attevelt elaborates: “An AI assistant is essentially a co-pilot, a colleague that treasurers can put questions to. This functionality will be integrated into a lot of different applications over the coming years. Corporate treasurers will particularly notice the release of co-pilots in Excel and Power BI, which allows for several use cases.”

In this scenario, if a treasury’s data foundation is of good quality and is available in Power BI or Excel, then the treasurer can ask questions regarding their data. The co-pilot can format the data and extract essential insights. The treasurer can ask what requires their urgent attention – perhaps there’s an FX exposure somewhere that needs to be hedged – and the co-pilot can flag this and suggest potential actions.

“This could save treasurers a lot of time because instead of having to sift through their data manually, they can just ask the co-pilots, and they’re probably not going to miss anything anymore,” adds Attevelt. “Crucially, that data flow has to be refreshed automatically.”

Generative AI may also play a vital role in unstructured data, of which treasury handles large quantities, through files such as loan documentation, FX deal confirmations, and bank guarantees. The treasurer may want to extract some data from these automatically, which is where AI automation can help.

“There could be one page of a loan document where the treasurer needs a particular question answered,” suggests Attevelt. “Generative AI can be used to interpret the unstructured data and to automate certain processes, but also let the treasurer interact with the data. It’s one of the best use cases for the technology.”

The understanding of what AI can do has changed so much and so quickly, particularly over the past 12 months. Generative AI, in particular, has effectively enabled anyone in treasury to start utilising AI, whether using the co-pilot approach or embedded within the treasury technology. As Stark explains, these different approaches can also work together to bring a whole new dimension to cash management.

“Think of an organisation with a global cash structure supporting regional treasury centres worldwide,” he posits. “Then, imagine building what is effectively a heat map of that world. This would show where the company has buckets of cash or excess liquidity. Add to that how the company constructs the treasury rules and make that visible so it can be drilled down into, and also the different processes that treasury is setting up.”

For a treasurer thinking about a country where rules about notional and physical pooling have an extra layer of complexity, this approach is about creating a set of processes available on the surface that can be shown and displayed. Treasurers are not simply completing a process, they are also visualising the information and seeing the ‘before’ and ‘after’ effects.

“This can be plugged into tools such as Bloomberg or whatever the source of data is for treasury, using APIs to help bring in information to expand the reach of the process,” acknowledges Stark.

Generative AI can be used across a range of treasury processes. It can help bring another layer of automation to pooling or IHB operations. Applying it to a payment factory could mean that the treasury is not just centralising its payments but also making a payment factory with some intelligence.

“Generative AI can handle the rules and the exceptions, it can effectively provide that next level of automation that currently is either manual or semi-automated for most treasurers,” Stark enthuses. “It can effectively provide a set of outcomes that, as a human, the treasurer can then approve.”

There are some challenges to consider when looking at AI, which are around data privacy and how third-party providers train their AI algorithms.

Attevelt cautions: “One concern I have with generative AI models such as ChatGPT is about using it in an enterprise setting. That’s being addressed with the introduction of ChatGPT Enterprise, where [research and development firm] OpenAI has explicitly said it does not use private companies’ data to train their model. But corporates still need to trust OpenAI as a counterparty because they are still sending company-specific data to it.”

That is particularly the case when it comes to fine-tuning the model. One way around that is for companies to train their own models. For example, Meta – the parent company of Facebook – has open-sourced its LLM.

“The move by Meta allows anyone, even commercial companies, to take that LLM and fine-tune its training according to their own needs,” Attevelt adds. “Firms can deploy it in their own cloud environments, so there is no need to interact with any external company.”

Treasury teams are naturally concerned about protecting their data, and their legal and IT teams are obsessed with data security. The critical question for data privacy is not so much what corporates can do and not do, it’s more about understanding exactly where their data is.

“It’s not much different to when organisations first went to the cloud,” muses Stark. “Are you comfortable with the data controls and the level of security being provided? Do you trust that the vendor is certified in a way that enables your corporate compliance to be satisfied?”

If the answers to those questions are positive, that is good news, but treasurers should also ensure the answers are based on current information.

“Sometimes, the perception of AI and data security is based on something that’s six months old and, in this space, six months old means obsolescence,” Stark adds.

This challenge also points to the increasing need for treasurers to have technological expertise that some may not have been expecting to master when embarking on a career in finance.

Del Natale makes the case that treasury teams must now include hybrid people who understand the finance and treasury aspect but also an application data mindset and some programming skills. “Treasurers don’t all have to be developers, but some understanding is important,” he notes.

This point resonates because, while internal IT support can be beneficial, IT staff do not have the knowledge of the finance side, particularly with the often-complex nature of today’s treasury. This means treasurers must be able to discuss technology issues with IT, their bankers, TMS and ERP vendors, and fintechs.

“Treasurers need to be as agnostic as possible from IT, bankers and other third parties when it comes to technology because they have to understand enough to challenge them,” states Del Natale.

A rapidly evolving future

Before treasurers think about AI, they must have the right environment for proceeding with a potential AI treasury solution. This all comes down to the data, as so often in treasury.

“It’s a ‘garbage-in-garbage-out’ scenario, so start by understanding and mastering the data,” implores Del Natale. “Treasurers need to think about what they want to get from their data and what they want from AI.”

Stark agrees and urges treasurers to think of AI as a complement to their data strategy and the processes and tech stack that they already use.

“Treasurers need to have a data strategy to understand what data to pull where,” he notes. “For example, do they want their AI process, coupled with APIs and other tools, to call the bank in real-time? Treasurers must consider how they want to use this, and then how they want to build that into a larger process around something as simple as decisions around investing or borrowing.”

The potential of AI makes this an exciting time to be in treasury, with the opportunity to experiment with various smart technology tools to see what impact they might have.

Attevelt advises: “Start a proof of concept. If a treasurer has a specific use case in mind, whether they’re sure AI will be the right solution or not, getting some hands-on experience with AI and seeing how it works is important.”

If AI becomes a success in treasury, it could enable the function to lead and encourage the rest of the organisation to implement similar technology. “In that way, being a treasurer can be inspiring,” concludes Attevelt. “Treasurers have myriad data at hand and can potentially use these state-of-the-art technologies to do great things.”