Is the current obsession with artificial intelligence in all things financial just hype, or is the technology truly at a tipping point? How will it affect treasury and what do treasurers have to know about it? By Leslie Holstrom.
Consultants, tech vendors and futurologists are in no doubt at all. Artificial intelligence (AI) is going to transform everything – including the financial services sector. The combination of limitless data storage, continued increases in processing power and the rise of distributed computing, and rapid advances in our ability to train machines in various ways, will hand large chunks of human activity to computers within the next five years.
Whether the most aggressive predictions come true or not, it is already true that AI is being used in sectors as diverse as finance, healthcare, transportation, manufacturing, the law and customer service. For corporate treasurers, this spread has two immediate implications: first, given the alleged benefits of AI in terms of productivity and data-driven insight-generation, how can they benefit internally from this technology? Second, how will the drive to adopt AI by the financial services industry affect the relationship between companies and their providers of financial products and services?
AI in the treasury and finance function
Current AI solutions, in general, rely upon a number of basic ‘skills’: some apply fixed algorithms to large datasets and recognise patterns which humans can then interpret as useful insights; some can be ‘trained’ to replicate processes to which they are repeatedly exposed, including those that require the ability to interpret natural language; and some can adapt their responses by ‘learning’ when they are exposed to new data.
Combining one or more of these abilities allows software to mimic customer service agents, to read and review complex documents, to match buyers to appropriate products and services and to provide sophisticated analysis of Big Data in close to real-time. In so doing, it is possible to imagine AI restructuring the entire operating model and the core processes of the finance function.
But before those more glamorous functions, AI will probably first be used to automate the mundane and repetitive.
Solving regulatory overload
Financial reporting and compliance, from KYC in trade finance and supply chain, to simply dealing with banks on a day-to-day basis, is increasingly complex, burdensome and expensive. It requires increasing numbers of staff to process huge volumes of data to generate the multitude of internal and external reports that are demanded by committees, the board, auditors, shareholders and regulators. Reporting and compliance adds little positive value to the business but a failure can create significant cost.
The first step in solving this problem is automated reporting for the most basic, standardised and formulaic reports. These may be mundane but they still involve large numbers of people performing low-value-added tasks. Some companies have already started to use robotic process automation (RPA) – essentially software ‘bots’ – to replace humans in these processes, but they have faced the problem that while the bots can be programmed to deal with some exceptions, they still stumble when tasks require even a small amount of ‘wisdom’. Adding machine learning allows RPA to deal with much more complex tasks and gives it the ability to go beyond volume-driven aggregation to functions such as intercompany reconciliations, the quarterly ‘close’ and earnings reporting.
In the UK, Arria has developed natural language generation software (NLG) that is being used across a wide range of industries to humanise and simplify the analysis of data heavy reports. KPMG has been using innovations from McLaren Applied Technologies (MAT) in its audit processes where predictive analytics automates evidence gathering and the production of complex data reports, saving time and improving client services. Deloitte recently announced a partnership with Kira Systems to aid in contract and document reviewing, and has already rolled out a customised version for audit processes with further applications being explored for tax and advisory practices.
And J.P. Morgan’s Contract Intelligence (COIN) system has replaced the 360,000 hours spent each year by lawyers and loan officers interpreting commercial-loan agreements and other contracts. COIN runs on a cloud-based machine learning system and as well as being many times faster than its human counterparts it has also managed to help J.P. Morgan decrease the number of loan-servicing mistakes that are, in part, driven by the need to interpret 12,000 new wholesale contracts every year. Again, it is the sheer scale of the data crunching required that will force companies to seek out this kind of solution.
The next level
More complex compliance requires more sophisticated solutions. Even a medium-sized firm may have to evaluate hundreds of tax and legal updates a week across an international network and techniques such as natural language processing and machine learning are being used to ‘understand’ the law, map compliance needs and even analyse the costs of compliance. By treating regulations as data, software will dynamically bring compliance into the enterprise risk environment, enabling treasurers to take a genuinely risk-based view of regulatory compliance.
These drivers have spawned an entire RegTech industry that aims to replace the present combination of scattered humans and fragmented legacy technology to ensure that trades, customers and the company are compliant.
The pace of change in global regulation and the complexity of AML/KYC and OFAC sanctions compliance are too much for existing instructional algorithms and rules-based systems. The latter flag cash transactions over a certain currency amount, block transactions to certain countries, use customer data to select accounts for additional monitoring, and categorise merchant accounts based on prior transactions. But they generate large numbers of false positives which need human intervention and they cannot cope adequately with deliberate attempts at fraud or evasion.
An AI solution ‘learns’ to identify problem transactions by analysing every data point in the entire transaction database. It develops rules of its own based on, for example, customer location, transaction timing, social media activity and relationships with other customers.
AI in AR
The same logic is driving the application of the RPA/AI combination to accounts receivable. AI algorithms are ideally suited to resolving some of the key problems in receivables because they can learn from previous experience and behaviour patterns. This allows them to build up a picture of good and bad credits, early and late payers, fraudulent activity and even message repair and exception handling.
One key receivables problem has always been the level of manual intervention required to cope with the mismatch between the way customers deal with billing and the way a particular treasury system would like them to. The multiple exceptions, errors and idiosyncrasies of the process defeat both the simplest systems that used optical character recognition (OCR) with templates and also more intelligent rule-based solutions.
Now however, companies like US-based HighRadius, Receivable Savvy and Germany’s collectAI are using self-learning solutions to automate AR processes to make debt collection more efficient, reduce costs, improve cash flow and increase customer retention rates.
These types of software do everything from work out the best channel on which to contact debtors to figuring out which payments relate to which invoices despite customers’ habit of using one payment to fully or partly pay multiple invoices and optimising early payment incentives to a particular customer using algorithmic invoice discounting to build intelligent supply chain finance solutions.
Others, like YayPay look at a customer’s payment habits and behaviours and uses machine learning to predict the potential day of their payment. This forecast is then used by the treasury team (or other automated solutions) to target the most significant outstanding debt.
AI in AP
In the same way, the advent of intelligent mobile bill processing technology able to execute transactions to any schedule or set of rules, will transform accounts payable. One set of solutions uses supervised machine learning to teach a software tool the key data points on a set of scanned invoices. After the training phase, the AI-enabled tool can perform the data extraction completely on its own and only brings invoices to the attention of a human where it does not recognise them.
Here, the main benefit of artificial intelligence is the ability to speed read huge volumes of invoices and to apply approvals rules to them. Almost as a by-product of those processes, these systems aggregate large volumes of data that, properly analysed, can be used by treasury, procurement or business units to ensure the supply chain is as efficient as possible. Similar solutions are available for T&E processing and monitoring.
All these developments are recognisably a linear continuation of the long-time treasury drive for process efficiency and productivity: cost cutting to you and me. As well as forming the backbone of internal treasury, these AI-based solutions will transform next-generation shared service centres.
Where AI gets more interesting is in its promise to revolutionise strategic and value-added treasury functions.
Forecasting and ERP
What have treasurers put at or near the top of their list of challenges for the last decade? Forecasting – cash forecasting, forecasting for risk management, forecasting underlying business variables from customer orders, to supplier payments to procurement need and to inventory levels.
That treasurers struggle with forecasting is no surprise. Accurate forecasting requires the intelligent evaluation of a large number of internal and external variables, the weighting of those variables, comparison with historical patterns, the incorporation of real-time data from the business, procurement and elsewhere and a view on how good business units themselves are at understanding their situation. At even a small company, this process involves far more data points than a human, even one equipped with Excel or even a good ERP system, can accurately model.
Treasurers ought not to feel too bad – it turns out, across hundreds of studies across a wide range of sectors, that a small number of fairly statistical algorithms, applied to past data, almost always outperform even the most qualified humans.
The latest solution is AI-based ERP systems which promise to optimise operational models and transform business operations. So, for example, as announced in mid-February, SAP’s S/4/HANA Cloud ERP product now incorporates predictive analytics and some machine learning capabilities in its Analytics Cloud.
IBM Watson Analytics, Amazon QuickSight, Microsoft Azure and Google’s Cloud platform also incorporate AI and it seems clear that those treasuries able to centralise the required datasets will be able to choose from a wide range of providers able to intelligently crunch it and provide the forecasts they need. Most firms will simply rent these solutions from the cloud.
In addition, companies like Microsoft are making their AI capabilities, for example in language processing and face recognition, available through APIs and they are linking with the largest open source AI platforms, such as H2O.ai, allowing other developers to add AI to their products.
AI-enabled ERP solutions will combine intelligent data analytics, smart automation, smart data gathering and sensor technology with deep learning, natural language processing and the technology to respond appropriately to changing situations in real-time. We are only at the beginning of this process, but companies like SAP are already deploying these technologies and corporates will need to re-organise all IT- and data-reliant processes to incorporate the changes. The impact on staff and organisational structure is hard to overstate. Is the day approaching where an ERP (or TMS) system will refuse a treasurer’s instructions on the grounds that they are sub-optimal?
The impact of bank AI on treasury
It is clear that the incorporation of AI into existing and new treasury technology will have an increasingly profound effect on the systems, status and staffing of corporate treasuries. But it may be the use of AI by their providers of financial products and services that has the greatest impact.
Transaction, risk and asset management are, along with credit provision, the core products treasurers need from their banks. So how will banks’ adoption of artificial intelligence change the world for corporate treasury?
Artificial intelligence may well create interesting new money management tools, hedging systems and transaction management services, but by far the most important implication of AI for any bank customer is how it will transform their ability to derive insights from their data.
We have already seen in the retail market that AI-driven bots can act as financial advisers, matching customers to appropriate loan, credit card and investment products. AI is also being used to improve retail credit scoring by looking at core data more intelligently at firms like online lender Elevate, Equifax and ID Analytics. Experian too is researching machine learning while relying for now on traditional regression methods.
The same basic idea will be applied to corporate banking – and while the banks emphasise how AI will lead to better service, more accurate matching of products and services to client needs and better pricing, treasurers need to think of the possible downsides.
Banks have unimaginably large databases of customer behaviour which, until now, they have been unable to analyse in any meaningful way. This data is not simply basic data on loans, timely repayments and core financial variables, it includes details of every payment made and received, the timing and location of those payments, the behaviour of cashflows over time, data on FX flows and hedges, data on indebtedness over economic cycles – and, importantly, data on all the counterparties to these transactions and, of course, the amount of money the bank has made from all this activity.
The combination of Big Data techniques, distributed cloud technology and AI will increasingly mean that they will be able to analyse this data. So what will they do with it? On the upside, it should mean that banks are able to offer a set of products radically improved by the addition of AI-driven advice. So, for example, banks will be able to analyse corporate transaction data across the cash cycle to determine which bills should be paid when, which customers should be offered discounts and what those discounts should be. It could suggest how payment terms could be altered more broadly to improve the overall P&L and to model the impact of suggested packages of products and services tailored to the client.
The data could be used to look at balance sheet optimisation, the best sources and types of funding for particular projects or acquisitions, for asset management and hedging. And, because banks will use AI for their own AML/KYC compliance, they will be able to help corporates with their supply chain and supply chain finance optimisation.
In addition, because the banks will be able to look at all their data, across all sectors, geographies and company sizes, they should be able to provide the kind of benchmarking and advisory services that treasurers have sought for the last 20 years.
However, there is a flip side to all this. The banks’ investments in AI and advanced analytics will be driven first and foremost by a desire to improve their own profitability. Their new-found knowledge may well be used to examine which clients are profitable and which are not; how different products and services should be priced according to individual customer’s needs; which clients are actually far more of a credit risk than existing internal models and ratings suggest because of their exposures to other companies or risks made visible by these advanced analytics; which clients do not meet their AML/KYC requirements because of the nature of their third-party relationships or payment flows – and so on.
Already the banks have finally agreed to pool anonymised credit data to improve their own risk management and pricing methodologies without AI. London-based credit risk management startup Credit Benchmark was founded in 2012 by ex-Goldman Sachs and Lehman Brothers employee Mark Faulkner. It specialises in pooling, aggregating and anonymising credit risk data from leading global banks, so that financial institutions can make better risk management and capital allocation decisions.
In AML/KYC, there are companies like ComplyAdvantage – an AI and machine learning RegTech startup to provide businesses with a feed of proprietary anti-money laundering (AML) risk data as well as on-boarding screening solutions and a monitoring platform for know your customer (KYC) processes. The system collects data from sources such as Interpol’s watch list, international sanctions and media reports to automate due diligence on clients that pose a criminal risk. Off the back of this data it can provide solutions like enhanced due diligence reports, risk and compliance advice and HR services. There is no guarantee that this kind of development will benefit all corporate customers and AI will simply make the analysis more profound.
Not quite yet
The more ambitious of these advances are some way off. Banks have found the task of aggregating and centralising their global databases extremely difficult and they will continue to do so. Only when that problem is solved can they then begin to apply AI and other techniques to those datasets and start to modify their offerings meaningfully.
In the meantime, AI is being applied to discrete functions that do not rely upon that aggregation of customer data. So companies will be able to benefit from advances in, for example, asset management where companies like Aidyia and Kensho are using machine-learning to run portfolios.