Published: March 13, 2024
Within treasury departments, AI can be deployed in various ways to deliver benefits such as enhancing cash flow forecasting, mitigating fraud, and managing FX risk.
But what about the significant governance and bias concerns that are part and parcel of AI usage? Here, four industry experts examine the use of the continually-evolving technology – including generative AI – within the function and weigh up its pros and cons.
While AI may not be new, the amount of data that is now available for it to mine – and the rise of generative AI (GenAI) apps such as Google’s Gemini (formerly Bard) that interface with large language model (LLM) neural networks – mean the technology is now more accessible than ever.
The advent of the internet, cloud-based services and storage, and APIs as an easy means of plugging an end-use AI app into pre-existing systems, mean that AI can no longer be ignored. It has moved on from RPA or ML applications of the technology, which are still relevant, to GenAI applications.
As AI becomes more powerful and accessible, with GenAI, LLM-powered applications proliferating and democratising access, its time has truly come. AI in all its forms is already disrupting corporate treasury procedures, training, and the strategic aims of the function.
“AI is a transformative force that will change our roles in treasury,” says Dominic Lynch, Head of Treasury at GoStudent, an Austrian-based edtech company. He is also a member of the Austrian Corporate Treasury Association’s (ACTA’s) working group which is examining AI applications and, particularly, the pros and cons ofthe technology.
“AI will enable treasury to become a broader, more strategic and interesting role,” he continues. “More time will be spent on analytics, risk and execution topics in future, rather than back office functions.”
Cash management reconciliation is undoubtedly one of those back-office tasks that has already become automated.
But we can expect RPA and ML-driven end uses and other AI tools to proliferate still further in future and drive automation and transaction management activities across physical and financial supply chains.
Predictive AI and generative AI (GenAI) are two different approaches within the artificial intelligence sphere, each with its own set of characteristics and applications.
As the name suggests, predictive AI involves using algorithms to predict future outcomes based on historical data. It relies heavily on statistical techniques and machine learning to analyse patterns in data and make forecasts. It often uses supervised learning techniques where the algorithm is trained on labelled data to learn the relationships between input variables and the target variable.
GenAI, meanwhile, involves creating new data samples or content that resembles a given dataset.
Instead of predicting outcomes, GenAI models generate new, original content based on the patterns and structures they’ve learned from training data. Applications include data augmentation for training datasets, creating synthetic data for training AI models, and generating creative content – perhaps even writing a draft treasury policy.
Read on for more real-life treasury examples.
According to Enrico Camerinelli, Strategic Adviser, Datos Insights, (formerly Aite-Novarica), citing his market research and observations, “treasuries are particularly interested in the possibility of applying predictive AI to the following use cases”:
He adds: “AI should also be able to better integrate sustainability into investment and financial decision-making in future, which helps treasury alignment with strengthening ESG goals that are coming in for all corporates.
“By automating manual operations and reducing the time needed for information gathering, AI helps treasury departments to cope with any staff shortages and become more data-led and efficient. Treasury can apply more precise risk mitigants and take proactive decision-making to enhance liquidity management, optimise capital allocations, and so on. This data-driven approach also facilitates better scenario analysis, aiding in more strategic planning and risk mitigation. Overall, it boosts efficiency and financial stability.”
Thomas Mehlkopf, Head of Working Capital Management, SAP & Taulia, echoes this, adding: “The introduction of AI technologies goes beyond mere automation and has the potential to revolutionise how we approach financial strategies and risk management.”
Specifically, he says, it enables a more data-centric approach to risk and cash management functions within treasury. “AI facilitates sophisticated analysis, predictive modelling, and real-time decision-making based on vast datasets. This shift towards a data-led risk function not only enhances the precision of liquidity forecasts and risk assessments but also enables proactive responses to market dynamics. As a result, the role of treasury becomes more strategic, adaptive, and capable of navigating complex financial landscapes.”
Patrick Kunz, founder of Dutch consultancy Pecunia Treasury & Finance, elaborates on these benefits with examples taken from his varied career as an interim treasurer. Kunz is also Head of Treasury a.i. (ad interim) at the Social Hub, a Dutch hotel and co-working hub. In addition, he has fulfilled a number of similar roles at Booking.com, CarNext and Arcadis, a design and engineering firm.
“AI ML applications can help in the automation of cash flow forecasting,” says Pecunia’s Kunz, agreeing with his peers that this is a primary use case. “Both on the side of data aggregation, validation and, probably most importantly, on scenario analysis and learning from the data, AI can help.”
“For example, if an AI tool can see that a client has 30-day payment terms but the client structurally pays on day 40, it might make more sense to put the cash flow in the forecast on day 40, rather than day 30. The same goes for the inflows. Often, a hard part in forecasting is not only predicting the future but also predicting what will happen with the overdues, which are, according to standard data, already in the historicals but not yet in the actuals.
“Another LLM-based use case involves an AI application reading extensive loan documentation and providing summaries of it to shorten the working timeframe for treasury control or accounting,” continues Kunz.
“Or AI can help find relevant information and clauses in complex loan documents more quickly, which is no small thing when you consider these can have 500-plus pages. Translating emails and internal conversations can be automated with smart LLM software as well. I know Microsoft’s Copilot LLM, powered by its Azure OpenAI cloud-based service, can do this.
“RPA applications of the technology, which have been around for a while, are becoming more mainstream, cheaper and useful,” adds Kunz. “With Power BI, for example, Excel automation of the simplest repetitive tasks can be achieved much more cheaply. Treasuries just have to start looking at their processes and identify which ones are repetitive and should be automated. And then just do it with AI-powered RPA. I have seen one company having a full-time RPA person moving from department to department helping with this automation. The guy is a money-maker for the company.”
AI tools – both predictive and generative – will also enable treasurers to take a deeper step into the future core of treasury, which is, in Lynch’s opinion, means risk management. “For instance, I could train an AI model on reconciled transactions, understanding the behaviour of each category to give more accurate liquidity forecasting, especially if I closely aligned it with a financial planning and analysis [FP&A] base plan, user groups, and other potential expected exposures,” he says.
“It can then identify deviations from historical and recent datasets to improve my forecasting capabilities and highlight nuances. If you have a rolling hedging strategy, it’ll be a few less clicks to analyse and execute risk strategies.
“You can dive deeper into the data with AI, present the findings with less time consumption, which will have a larger positive impact on the company,” continues Lynch. He has also already used the OpenAI-developed ChatGPT application, the enterprise version supported by Microsoft, to improve team collaboration, minute recording and data security at his firm. “It’s by far the most powerful treasury personal assistant.”
Other ChatGPT end uses already deployed at GoStudent include:
“GenAI, when speaking to your data lakes, will allow for better insights and faster results in future,” enthuses Lynch. “Once a potential model can understand the behaviours and trends of a business model, it could allow for insights to unseen financing opportunities, such as better working capital management with trade financing.”
In short, the potential use cases of AI are essentially limitless. Indeed, Lynch is already foreseeing another FX risk mitigation step. “If an AI model can see an ERP user in Brazil, for example, and know they are expecting a new last-minute vendor invoice of 100,000 BRL on the 25th day of the month, for the next three months [but not recurring thereafter], then you can build something around that. The data can be sent via API to the TMS, or data lake, and this could allow an AI automation job to be carried out to foresee an updated temporary exposure for BRLs. This updates the exposure reporting mechanism, and you can have the AI model recommend an update to an associated three-month-long FX hedging strip, if it breaches the target balance.
“In addition, you could get the model to understand this is only a three-month temporary vendor and therefore exclude it from any future forecasting models. This reduces noise in subsidiary and group liquidity planning.”
Despite the clear potential benefits of AI, however, “we still need humans to verify and control AI tools,” cautions Kunz. “You need a human checking the AI’s work, especially in GenAI apps that link to LLMs, as these make mistakes, even while writing down responses very convincingly,” warns Kunz. “Data security and sources are also an issue. AI doesn’t reference data or know if it’s correct.”
As ever, good procedures and a diversity of contributors when writing models and apps are essential to prevent bias, and good governance is essential to police systems when they are live. A human must always remain in the loop, especially during testing and experimental end uses during pilots and first deployments. Ongoing oversight is also essential.
The latest example of AI going horribly wrong occurred in January this year when the delivery firm DPD had to disable part of its online support chatbot in the UK. The chatbot swore at customer Ashley Beauchamp and criticised DPD when prompted by the Beauchamp, who shared his experience with millions on X, formerly known as Twitter[1].
A new update had caused the DPD chatbot to behave in this unexpected manner and has since been rectified. But this minor and relatively benign example is just one of many instances where AI has run amok because its programming, governance, and oversight limitations detailing the parameters it should operate in, were deficient.
Air Canada, meanwhile, has suffered a similar fate – with financial and reputational consequences to boot. After its automated chatbot provided a customer with incorrect information, leading him to purchase a ticket at full price under false pretences, the company has been ordered by a tribunal to pay compensation. Adding fuel to the fire, Air Canada faced a backlash for trying to disassociate itself from the mistake by asserting that the chatbot was accountable for its own actions[2].
Do treasurers really want to release such a system on their departments, accounts, forecasting, inventory, payments, cash management and other such activities? The answer should be ‘yes’, in order to gain the efficiency, automation, and data insights that humans cannot provide. But only if the AI is pre-programmed and controlled correctly. Governance pre- and post-deployment is all important.
Dominic Lynch was hired as the first treasurer at GoStudent in March 2022. He built out the treasury function from scratch and implemented a plan that included a TMS from FIS (Integrity Edition, installed in just five months) and a centralised banking relationship with Citi. Interestingly, GoStudent has also integrated a data lake with AI-powered data-crunching capabilities via APIs between these partners – winning a TMI Award for Innovation and Excellence in the process.
The data lake, linking to the TMS and its banking partner, and the AI tool installed to get the most out of it, supports analytical capabilities in the risk arena. An advanced liquidity model, with better working capital, is also possible thanks to AI’s ability to spot patterns and enhance accurate cash forecasting.
AI-power with a data lake, fed with information from the TMS and bank, has also enhanced the investigation and adjustment of payment files. For instance, transaction data fed into the AI can detect issues in payment failures and prevent them happening again by highlighting and implementing the changes. This unique aspect of the project has much to teach other, older established corporate treasuries, which perhaps haven’t yet fully realised the power of AI.
“We are presently pulling the data into the data lake, so we can carry out further analysis reporting on different transaction movements which we then feed into the AI forecast model,” continues Lynch, as he sets out one of the many end uses of its innovative new solution. “We are now able to see the transaction reports right in the TMS. As a result, we have been able to reduce FTE by 1.5 across other functional departments.”
Automation and efficiency have increased at GoStudent since the establishment of its new treasury and supporting systems. Moving forward, the use of AI analytical power will be useful in predicting interest rate risk, which is currently particularly volatile, currency movements, and other such risk mitigation duties, while also enabling the organisation to be more strategic and data-led in its future planning.
Read more about this project in TMI’s 2023 Awards Yearbook: tmi-awards.com
AI does have other drawbacks beyond reliability, too. “It requires substantial initial investment for set-up, implementation and maintenance,” admits Camerinelli.
“The accuracy of AI models depends on the quality and quantity of data as well, which can be challenging to obtain especially for treasury business data that is – necessarily – kept highly confidential. Hence privacy issues can arise.”
Other challenges include the current fragmentation of the technology stack across ERPs, TMS’ and bank proprietary systems, which is where APIs can help, although the difficulty of integrating in-house and third-party vendor applications shouldn’t be minimised. Additionally, there’s a steep learning curve for employees to understand and effectively use AI systems. Over-reliance on AI could also lead to complacency in manual oversight.
Skill sets therefore need to be updated in line with AI developments. SAP & Taulia’s Mehlkopf comments: “There is growing demand for human skills in data science, machine learning, and AI technologies. As the treasury function continues evolving to embrace AI, professionals need to be proficient in handling and interpreting large datasets, and qualifications related to data analytics and AI will become more attractive and relevant.”
There will also be a shift in working methodologies, with a greater reliance on AI-driven insights for decision-making. Mehlkopf continues: “Such a shift necessitates a collaborative approach, with treasury teams working alongside AI technologies. Understanding the power of AI but also its limitations – especially with regard to data access or potential blind spots – will be key to fully leverage the benefits.”
Therefore, the future treasury workforce will need a blend of traditional financial expertise and tech know-how, in order to leverage AI tools, he believes. “For this future workplace to thrive, it is essential to foster a culture of continuous learning to keep pace with advancements in technology and ensure a seamless integration of AI into day-to-day operations.”
But whether or not the workplace is thriving, AI should never be seen as a silver bullet. As Kunz concludes: “Don’t think AI will be your solution to everything. Don’t start looking into AI or LLM applications if you don’t have a strategy around it. And don’t deploy it yet if you don’t have the resources in people, skills and money, to implement it properly.”
Good deployment means a step-by-step approach. “First, look at your data sources, the number of ERP systems, how to connect them, the data quality, validation and so on,” advises Kunz, warning that otherwise it will be “garbage in, garbage out – even with the best tools”.
In many ways, Covid-19 has been the ultimate test for treasury teams and their technology. With the spotlight firmly on treasury management systems (TMSs), treasurers are asking whether they have the best tech for the job – in a remote-working environment where instant access to ‘decision-ready’ data is critical. Is there a better TMS out […]
According to this year’s responses to the EACT’s annual pan-European survey, the pandemic has not changed the top priorities of corporate treasurers. Instead, it has crystallised the need for increased digitisation. The European Association of Corporate Treasurers’ (EACT’s) annual survey reveals the top priorities for treasurers during the year ahead, identifies critical challenges facing the […]
Danone’s multi-phase technology-fuelled project tackling risk in Asia has created a more agile, resilient, and connected operation Over the last 12 to 18 months, accelerated by the onset of the pandemic, Danone’s treasury initiated an ambitious multi-phased project with a view to transforming its treasury and risk management in Asia. The team has enabled faster […]
