The Truth About Treasury and Artificial Intelligence

It’s been a hot topic for almost every treasurer over the past year, but – hype aside – what is the true value of artificial intelligence (AI) within treasury? Nikolai Diekert, Director Product Management at leading TMS provider BELLIN, explores the concrete use cases for AI in treasury, providing a candid view on where the technology can add value and where it still has room for improvement.

Eleanor Hill, Editor, TMI (EH): Before we talk about AI in treasury, it would be helpful to clarify what AI is – and what it isn’t. Would you be so kind?

Nikolai Diekert (ND): Of course. You’re right – there are many contradicting and confusing definitions of AI. A good first step is to split the definition into ‘artificial’ and ‘intelligence’. What we mean by ‘artificial’ is that it is non-human, but created by people. While we tend to think of modern computer devices in this instance, analogue machines also fall into the ‘artificial’ bracket.

As for ‘intelligence’, there are even more definitions of this than there are for AI! But there is an interesting and useful definition on Wikipedia, which states that it is the ability to ‘perceive or infer information, and to retain it as knowledge to be applied towards adaptive behaviours within an environment or context’. Taking these elements together, AI can be viewed as any device that perceives its environment and takes actions that maximise its chance of successfully achieving its goals.

You also asked what AI is not, though, and this is very important. We often hear the words ‘machine learning’ (ML) uttered in the same breath as AI – but these are not one and the same. ML is a subset in the field of AI whereby algorithms build a model based on sample data and perform tasks or make decisions on real data without being explicitly programmed. The model is ‘trained’ on the sample data in various ways, supervised, unsupervised, reinforcement, self-learning and so on.

EH: Isn’t ML the scary part of AI – the bit that makes treasurers wonder if they will be replaced by machines?

ND: Some of the results that Google’s DeepMind has shown in recent years do make us wonder if machines will take over soon. In October 2019, for example, Google announced that its AlphaGo AI had beaten a world-class player at the ancient Asian board game Go – in other words, it played better than a human. For some people, this is a frightening thought, for others, it is something they are looking forward to.

But when asking whether the treasurer will be replaced by a machine, we have to be realistic. At this point in time, the machine still needs people to programme it. And there are some parts of human nature – like gut instinct and experience – that a machine cannot replicate entirely. Of course, we can’t be sure how fast the world will change!

EH: What are the main applications for AI within treasury?

ND: As I see it, there are four main areas of application for AI in the field of treasury. First is the automation of tasks. Any repetitive task that requires only minor decisions to be made could have AI applied to it, such as reconciliation of forecast with transaction data. Even if automation is already very advanced within a particular treasury department, there could still be advantages to using AI, such as making explicit instructions ‘redundant’ by using machines that observe human behaviour.

The second area where treasury could benefit from AI is forecasting. For many treasurers, producing an accurate and real-time cash flow forecast remains an elusive task. A handful of treasury teams are already using AI to improve their liquidity planning and risk forecast, having developed their own AIs alongside in-house data scientists with significant technology investment. There are also some treasurers who have purchased forecasting systems from vendors which leverage AI, but there is some debate as to the extent of the ‘intelligence’ here.

Next, we have support in complex decision-making. This is not the same as actually making the decision. It is providing analysis and different scenarios in order to assist the final decision. A good example would be determining which instrument to use for a specific hedge of a hard-to-understand risk. AI could also help treasurers to pinpoint the optimal time to issue a bond, for instance.

Last, but by no means least, is fraud prevention, especially in the area of payments. Here, software can predict and prevent electronic payment losses before they occur. Machine learning can automatically respond to variances in data, behaviours and trends. It learns patterns of fraudulent and legitimate transactions, to simultaneously minimise fraud and false positives.

EH: What are the limitations of AI and ML? What do treasurers need to know beyond the hype?

ND: Each of the practical applications I have outlined has its advantages, but also its challenges. For the repetitive tasks, one of the main considerations is that it remains unclear how big the advantage of using a ‘black box’ is compared with scripted automation with well-defined exceptions that need intervention from a human expert. Is this a case of AI for AI’s sake? Sometimes, I suspect this is true.

In the forecasting sphere, data quality, availability and classification are significant hurdles. Access to accurate, timely and holistic data, whereby silos are broken down, is a prerequisite for good ML. This is often a challenge for treasurers to achieve given confidentiality of data, legacy systems, manual errors and so on. Furthermore, anyone who says that applying AI to forecasting is a ‘plug-and-play’ project either a) uses a very simple statistic methodology that probably does not deserve the name of machine learning or b) does not truly understand the complexity of such a project. It’s also worth noting most AI projects in the forecasting space do not cover more than a short-term horizon of a few weeks because long-term factors are still hard to predict, even for more advanced AI.

As for support in complex decision-making, the main challenge is again the ‘black box’ and the question of how to ‘train’ the AI – because the number of high-impact decisions in treasury is relatively low and modelling them is extremely difficult. Moreover, the knowledge required to make such decisions requires years of experience, factoring in macroeconomic statistics, corporate strategy, risk aversion, and business relationships with banking partners. So, in this particular area, the use of AI is limited by the need for human expertise, which is tough to replicate.

When it comes to fighting fraud, this is arguably the field where AI has been applied for the longest – but more within banks and credit card companies than corporate treasury functions. To help minimise fraud, many corporate treasuries already have tools at their disposal, like central payment hubs, two-factor authentication, and straight-through processing. While AI can certainly have its benefits here, it remains to be seen whether it is the ‘ultimate’ answer without applying the measures mentioned above.

EH: Could you elaborate on the ‘other tools’ you mentioned there – give us an example?

ND: Without turning this into a commercial discussion, treasurers might want to look at tools such as our Vendor Verification solution. In a nutshell, this is a payment validation check, which is fully integrated in the tm5 treasury management system payments’ platform. It enables companies to configure whitelists and blacklists and every time a payment enters tm5, Vendor Verification compares the account information of the beneficiary with those lists. If the information does not match, the system either issues a warning or blocks the payment, depending on settings. Both single and bulk payments can be validated. Payments that could not be validated can then be manually approved, overriding the validation result, or confirmed as blocked.

EH: Will BELLIN be using AI in any of its future solutions, then? Or do you prefer other methods?

ND: AI for fraud is already up and running for some customers. We plan on rolling out sanction screening integration within Vendor Verification and are working on ML and outlier detection functionality, too. Based on historical data and potential risk, the system will provide recommendations upon which treasurers can decide to approve or block payments as they see fit. We are also exploring ways to leverage data sharing.

EH: So, what’s the bottom line on AI in treasury in your view?

ND: As we have discussed, AI certainly has potential use cases in treasury. There can be significant benefits, if it is deployed in the right way and for the right reasons. But it’s important to remember that AI isn’t a silver bullet and it isn’t always the optimal solution.

Instead, I prefer to see AI as an opportunity to ask ourselves where we can improve processes and underlying data quality, as well as enhancing the understanding of data by decision-makers. As it stands today, AI often creates benefits that stem from its prerequisites, rather than the results it delivers.