As AI in various forms moves into the treasury space, this is a good time to take stock of the benefits and pitfalls of this increasingly popular set of technologies. A recent McKinsey survey finds that 50% of businesses have adopted AI in some form and part of their business. Less than a quarter of respondents report seeing significant bottomline impact, so it would seem AI is no silver bullet.

What is AI?

Despite the sometimes-creepy prescience of our phones and search engines, we are still far from the kind of AI popularised by the media – like Stanley Kubrick’s HAL. (HAL would be a ‘general AI’, whereas businesses today are using narrow AI; general AI remains a dream for the industry, and some say it may never arrive.) Although some AI technologies are making decisions in retail finance and insurance (as well as narrow domains such as trading), most treasury applications comprise AI assistance rather than control.

McKinsey breaks AI into two branches:

• Machine learning (ML) in various forms finds statistical relationships between different data points; treasurers can think of it as technical analysis on steroids (because machines can pile through more data and relationships than human analysts).
• Deep learning comprises neural networks and similar technologies that seek to mimic how the human brain operates by reinforcing data pathways that seem successful.

McKinsey further sub-divides machine learning into:

• Supervised ML requires human trainers to provide data with correct answers to train the algorithm.
• Unsupervised ML is more exploratory for example searching through data for interesting correlations.
• Reinforcement learning means that the algorithm tries different approaches and weights the most successful ones so that those are used more.

Deep learning sub-divides into:

• Convolutional neural networks.
• Recurrent neural networks.
• Transformers.
• Generative adversarial networks.

ML seems to be the most commonly used AI technique in treasury. It is suitable for forecasting and error detection. In some cases, neural networks are being used for fraud detection, but often this problem is similar to error detection.

Statistics

To some degree, treasurers can benefit from their experience with statistics to avoid some of the pitfalls with AI. ML especially equates quite closely with familiar treasury techniques like Value at Risk (VaR). ML will often develop more complex relationships than the correlations and volatilities used for VaR.

A common problem with VaR is that it relies on a limited historical dataset – often as short as one year. When markets change direction, as they periodically do, the last year is not a good guide to potential losses. Treasurers have developed mitigants such as stress testing to complement VaR’s weaknesses in this regard.

Bias is often discussed as a problem with AI.

With ML, treasurers need to be vigilant in an equivalent way. This is especially true because ML may develop more complex data relationships that can be difficult for treasurers to understand – the so called ‘black box’ or ‘explainability’ problem. So it is all the more important to be vigilant about the potential limitations of the dataset being used. For example does it cover a full economic cycle?

The year 2020 has provided a good example of this problem. For instance, forecasts built from two or three year old datasets will not have captured the turmoil caused by COVID-19. This is analogous to how the Asian currency crisis at the end of the last century ‘broke’ many VaR metrics.

Data

Most treasurers will have read that ‘data is the new oil’. This allusion is based on the importance of data to ‘feed’ AI in order to get the desired results. This is as true of cash flows as it is of clicks and visits that social media use to maximise their manipulation of our attention.

The answers, suggestions, and warnings that AI generates will at best be as good as the data that it is being fed with. Treasurers with effective data solutions, often data warehouses or lakes or oceans, have a big advantage in availing the benefits of AI. Although there are some AI technologies that handle unstructured data, and natural language processing may seem to offer an easy way to cut through data clutter, in practice treasury AI benefits from clean and well-structured raw data.

So treasury AI projects often start with collecting and cleaning underlying data. This can be a large project since, to obtain the most useful results, treasurers will often need to feed their AI with non-treasury data – for example, combining cash flows with accounts receivable and accounts payable data and even sales and production planning.

Explainability

As alluded to above, there is an issue with explaining AI’s forecasts and suggestions and, where relevant, decisions. AI will comb through datasets and find relationships in the data that are invisible to human observers and, even when exposed, do not make sense to humans. AI is playing with 1s and 0s – it has no common sense. On the other hand it has no preconceptions. When DeepMind’s AlphaGo beat Lee Sedol in March 2016, human experts had great difficulty fathoming that AlphaGo had not simply played with discipline but had in fact developed completely new strategies never considered by thousands of generations of go masters.

Managers will quite naturally feel reluctant to follow tactics or strategies that they do not understand. To some extent top management uses their understanding of subordinate experts to become comfortable with decisions they do not personally understand. But with AI there can often be no explanation, and there is no silver-tongued expert to get management on board.

This issue is such a big problem for the AI industry that it has spawned a new sub industry –eXplainable Artificial Intelligence (XAI) refers to a set of techniques that help show how a ML algorithm comes up with predictions and suggestions. XAI potentially enables humans to better understand (and trust) outputs from complex, ‘black box’ models and helps decision makers know the relevant data characteristics driving the output.

Perhaps most salient for treasurers is that this issue is another reason why the question is not how to replace analysts with AI but rather to find practical ways to augment analysts with AI. Treasurers will need all their domain knowledge more than ever, and they will have to leverage this with understanding what the AI is doing. AI is a new (albeit tireless) team member, and must be understood and managed like the human team members.

Bias

Bias is often discussed as a problem with AI. Training data may be skewed, especially when dealing with human issues. The conceptual equivalent for treasury would revolve around changes in historical patterns like the COVID-19 impact mentioned above, and to the potential for business model changes to trip up AIs honed as-is or business as usual (BAU) data.

Treasurers need to be vigilant when new product and or market segments are introduced because AI is definitely not clairvoyant. The AI will proceed on a BAU assumption and will often need to be re-trained. This is a challenge for new products and markets because there is no historical data with which to train the AI. Layering human assumptions will at best reduce AI’s ability to bring new light to issues. Even mixing up data from previous changes may not give the correct results – after all humans are also heading into uncharted territory.

There is no easy fix for this kind of situation. In the end it comes back to the collaborative AI concept – treasurers need to understand the limitations of their AI team members and manage them accordingly.

Conclusion

AI is tireless and fast, and also devoid of common sense. Treasurers need to get a feel for how to use AI productively, just as they need to understand the strengths and weaknesses of their human team members. AI is probably best thought of as a tireless ‘gopher’ assisting human treasury staff. And any attempt to enhance treasury with AI must start with clean and complete data.