Accurately predicting cash-flows is incredibly difficult, hampered by treasurers only having access to rudimentary software tools and spreadsheets. And while the prediction can provide guidance, it is not wholly reliable. Patrick Kunz, managing director of consultancy firm Pecunia Treasury & Finance, says having even 70–80% accuracy in cash flow forecasting (CFF) would be considered a success.
The innovation of artificial intelligence (AI) in CFF seems to provide the solution. While existing CFF software can utilise historic data, financial algorithms and predictive analysis, AI can dig into invoicing patterns, error recognition and discrepancies. It can even go as far as to flag potential fraud and model possible scenarios. AI can also be integrated into the accounts payable and accounts receivable processes.
Most significantly, these AI-enhanced processes are already a reality.
Henry Benamram, senior vice-president of corporate and business development, North America, at business banking software firm OneUp, says: “What we describe is not happening tomorrow: it is already here. We have worked with BNP Paribas and Lloyds Banking Group to deploy the OneUp Business Assistant to customers. Through banks, we have 800 customers benefiting from these AI services.
“This technology is moving very fast; things that were in concept a year ago are already in the market.”
However, the idea that CFF automation requires simply inputting data would be an oversimplification. “AI is not the Holy Grail which will solve all problems, but it will help with making forecasting more efficient,” Mr Kunz cautions.
Garbage in, garbage out
To reach the point of efficiency means getting a handle on large sets of cleaned data, with full understanding of where the data is from. “If your data quality is low, it’s garbage in and garbage out. This will mean the CFF will be terrible as well,” Mr Kunz says.
With that in mind, getting to the point of feeding data into the AI model will require significant work from the user end. Building a data set internally can mean drawing across multiple sources and different departments working together. Data must be fed in slowly, so any issues in one data set can be identified and remedied before the next set is introduced.
“There are AI tools that extract data out of existing tools, such as enterprise resource planning (ERP) systems, procurement systems and vendor systems,” Mr Kunz says. “The ERP system will be the main source, but the data within it is very short term. It is the accounts receivable, the accounts payable, information on taxes and salary payments, but that’s often lacking beyond two to six months of data.”
Minimal data can be bolstered by looking to outside sources. Sourcing the data externally can create a better picture, but also increases the associated risks due to unknown elements within the data sets. This is particularly true if the data is being used in more complex scenarios.
It’s through continued usage [of AI forecasts] that trust is built in the model
Bob Stark, global head of market strategy at treasury solutions provider Kyriba, says: “When an organisation is expanding into a new market and does not have the historical data, it can use proxy or synthetic data, taken either from another region or predictions to make new data. This can also be used with mergers and acquisitions to understand a cash flow situation where there is no existing data available.”
When data has been accumulated, it needs to be prepared. This can necessitate a deep-dive into the existing data sets to understand a range of metrics from the nature of a transaction, the supplier, the end customer and tax implications.
Mr Benamram says: “Once there is clean data, you can start identifying patterns to predict what will happen and create a forecast. It takes single-user data to find repeating patterns and cross-user data. All data is anonymised.”
An unexpected consequence of data cleaning can be finding existing issues. For example, it may find that certain clients are habitually late in making payments. In this case, Mr Kunz says it is up to the corporate to take action with the customer directly, or to contact their credit department. “If they’re not meeting payment terms, no tool will fix this,” he adds.
In some cases, the problems can be an internal issue. Mr Kunz has seen this arise with previous customers: “In one case we found a client had a spike in activity on the 15th of each month. The chief financial officer (CFO) did not know why this happened. After digging, we found the admin person was inputting data in the middle of the month, as they were originally told that as long as the monthly totals were balanced the date did not matter. This is a good example of having the data, but not the accuracy.”
Risk factors
The nature of AI systems means they are highly complex and cannot be used as a plug-and-play software tool that treasurers may have become accustomed to for other aspects of their business.
“If the engine is trained in a certain industry, it cannot be plugged into another industry,” explains Mr Benamram. “You have to be sure the engine takes all industries into account, evolves as the market evolves, and does not assess criteria that are not ethical.”
The risks around data are apparent when considering the position of companies looking to adopt AI for CFF, and are faced with using the highly abnormal data that had been collected over the past three years.
“If a company is going to teach its algorithm based on data from 2021 and 2022 to predict 2023, the business conditions in those years were obviously very different from normal,” Mr Stark says. “The value of the US dollar would have an impact if you were exposed to the currency. There are many factors which are not down to bias, but about what is truly different. Understanding what you’re training your algorithm to include, and what to ignore, is incredibly important. Users have to be actively conscious of this, rather than assuming the AI tool will figure it out.”
Being cognisant of AI’s limitations, even in the smartest systems, means such systems can require human intervention to function correctly. “Some AI tools will convert the data and predict the future. The risk is while there is logic, they need to be challenged. For example, some tools needed to be shut off during the Covid-19 pandemic because the tool couldn’t handle such a sudden change in behaviour,” Mr Kunz says.
Effective usage of AI requires a degree of trust to be built up between human and machine, says Nilesh Vaidya, global industry head, retail banking and wealth management at Capgemini. “AI models are extensions of quantitative predictive analytics, and the hallucinations experienced in these models is lower than in generative AI models,” he says.
“The risk departments on these cash forecasts need ongoing validation of the loss ratios. Many AI forecasts are initially used a lot for internal purposes, and it’s through continued usage that trust is built in the model. Of course, banks use a higher verification threshold when these models are shared outside the enterprise.”
Discrimination and bias
Even when the data has been input, AI still has multiple risk factors. In Deloitte’s July 2023 note ‘Banking on the bots: unintended bias in AI’, it is stated that bias can creep in across three areas: the input of data, the development stage and after training.
“Some methods of AI training may obscure how data is used in decisions, creating the potential for discrimination, for example if race or gender data was used in credit decisions or insurance premiums,” the Deloitte note states. Should this creep in, it could create an “ongoing cycle of bias”. It can also be impacted by a lack of diversity among development teams.
You need to be sure the data analysed is not discriminative, specifically along ethnicity and local geography
In the post-training stage, the AI can, over time, shift towards discriminatory decisions based on previous results. “As AI systems self-improve and learn, they may acquire new behaviours that have unintended consequences, for example if an online lending platform began rejecting loan applications from ethnic minorities or women more than other groups,” the Deloitte note states.
Marie-Vincente Beau, communications and marketing manager at OneUp, stresses that this is an important area of focus. “You need to be sure that the AI is not discriminating against people. You need to be sure the data analysed is not discriminative, specifically along ethnicity and local geography. You cannot allow AI to analyse the time it would take to pay a loan, for example, based on ethnicity and where they are located, and even less make a link between these two.”
She suggests that ethnic and geographic data analysis should be avoided to prevent discrimination or discriminatory statistics, cautioning that an AI can easily create a new data set if it is presented with information such as where an individual lives, what their financial situation is and how long it takes them to make payments.
Where banks fit in
While the development of AI systems is largely being driven by technology companies, banks are a vital partner in ensuring success. While they can provide hard data, they also have the soft skills built up through their relationship with a client.
“Banks are influential advisors to corporate CFOs and their finance teams,” says Mr Stark. “Any insight they can provide, especially around external data, is very useful. It is possible for the teams to, for example, draw a consensus across their top three banking partners in areas such as interest rates, inflation and foreign exchange (FX).”
Indeed, the process can be mutually beneficial for the banks. “Banks have a high interest in their customers having more accurate, longer-term cash and liquidity projections, for the simple reason that if CFOs and treasurers have greater confidence in their forecasts then they can have more strategic discussions with their banking partners,” Mr Stark says. “This can extend beyond simply borrowing or FX trading, and instead focus on holistic risk management programmes or new ways to unlock liquidity within the cash conversion cycle, for example.”
For smaller companies that do not have a large treasury team, having the data built into their banking portals can give them easy access to real-time financials. “It also improves customer service, as banks can tell them what they need to do based on their data by pushing them daily tips to better manage their businesses. It’s a win-–win for the small and medium-sized enterprises and banks,” says Ms Beau.
There is a high degree of involvement the corporate can have in the system implementation, explains Mr Stark. “A question we get asked is how does the corporate support this? What is the reliance on vendors or the bank, versus what the corporate can do itself? For example, does a CFO have the resources to build a model, clean the data, to understand the new results? Can they surround the ERP with the apps and application programming interfaces needed to make it work? We are asked this by IT teams and chief investment officers, on how to make this reachable by our organisation.”
Next-generation AI
With AI for CFF already in use, the question becomes how the tool can be expanded in the near future. One hurdle can be the cost. Mr Kunz says that he only sees his largest clients using the tools and hopes they will become more mainstream over time. The amount of time it takes to implement the software and the technical staff needed to execute this may also hamper widespread adoption in the near term.
For those looking across the scope of potential, there are possibilities abound. Mr Vaidya believes there are three frontiers emerging for the next generation of AI. These are the advent of enterprise-wide CFF for large corporates, which are largely limited to local forecasts, the ability to connect CFF with supply chain finance providers to drive product innovation, and improved cross-border CFF combined with real-time payment capabilities.
More on cash management
Innovations in other areas of AI will inevitably feed into CFF. “The most impactful future development will be how to leverage generative AI and large language models like ChatGPT,” Mr Stark says.
“The past year has demonstrated how AI is in reach for treasurers looking to advance their treasury data strategy. Banks are in a fantastic position to support treasurers by understanding their role in providing the tools and data to help improve AI-driven data predictions and unlock greater process automation through the use of embedded AI.”
