In part one of our series about artificial intelligence (AI) we considered the state of AI today. In this piece, we look at how the payments industry is already using AI to cut fraud, reduce transaction times and offer customers better service.
Every year, citizens of the European Union make 122 billion digital payments using payment cards, bank-transfer apps, e-wallets, mobile wallets and other payment methods. Without AI, there is no way the payments industry could process so many transactions so quickly and keep fraud and error rates down to an acceptable minimum.
Why AI is necessary
The massive volume of digital payments is both a help and a hinderance to spotting fraud. On the one hand, it strains existing fraud-detection systems to the limits. On the other hand, it provides developers of artificial intelligence with the data they need to train their algorithms.
A traditional rule-based fraud-detection system might consider a range of variables, for instance location, the type of merchant, the amount being spent and so on. Thus, if a user spends more than usual, with an unfamiliar merchant in a previously unvisited location, this will probably end up being flagged as a possibly fraudulent transaction. It’s why your card sometimes ends up frozen when you go a bit nuts on holiday and spend too much on skiing gear.
The problem with this model is that it’s too rigid to cope with increased volume and complexity. For one thing, it yields too many false positives. Only 1.49% of all transactions are fraudulent. But like our imaginary skiing splurchase, in today’s highly diverse and mobile world many purchases do not fit well into a rigid rules-based model of fraud detection.
A 2017 study of merchants in North America found that 79% used manual reviews to determine whether at least some anomalistic transactions were likely to be fraudulent. On average, merchants manually reviewed 25% of all transactions. According to another study, 52% of transactions flagged as fraudulent were false positives.
By 2020, global merchants are expected to be processing 726 billion digital payments every year. With those kind of volumes, no one can afford to rely heavily on a manual review process.
How AI is used in payments
Using AI, usually machine learning, can significantly improve fraud detection and reduce false positives. For example, a payments AI might look at a whole range of factors and assign a risk score to each. A merchant with a good track record might have a low risk score, say 15%, but an unfamiliar IP address, time zone or location might attract higher risk scores. This process can be repeated for hundreds of factors, with the final average score determining whether the transaction passes the merchant’s threshold for being flagged up as fraudulent.
It’s possible in this way to analyse vast quantities of data to build a much more sophisticated picture of what “normal” looks like. Because it’s not limited to working within set rules, an AI can analyse this larger body of data to look for unexpected commonalities between both fraudulent and non-fraudulent transactions.
A financial institution using AI to detect fraud benefits not only from being able to process transactions in real time, something it could not do if manual checks were required, it’s also constantly finding new ways to recognise the anomalies that allow it to identify a fraudulent transaction from a real one. This is in addition to the advantage of being able to process a vastly greater number of transactions in near to real time.
Beyond fraud detection
Fraud detection is the most common use for AI in finance but it’s not the only one. The same faculty for finding patterns and defining new variables can also be used to spot potentially useful or worrying connections and behaviour as part of the know-your-customer (KYC) process. Using AI in this way, institutions can process larger quantities of data, from a range of sources including all the customer’s accounts and other products with that institution and soon, with the advent of open banking, with other institutions as well.
Credit scoring is another area in which the industry is already using artificial intelligence. Again, the challenge here is to analyse data across millions of accounts to spot patterns which correlate strongly with a risk of default. Once AI has been used to derive these models it can then check individual applications and customers against them, going beyond simple models based on a narrow range of factors such as past spending and expected income.
But as well as being a positive for the industry, the use of AI for credit scoring also points to some of the risks both for institutions and consumers. Unless the algorithms are rigorously tested and weighted to avoid replicating bias that already exists in the data, or introducing new biases by making false correlations, then there exists the danger that credit-rating-AIs could unfairly deny some people access to financial services. We’ll look at this risk, and what the industry is doing about it, in a future instalment of this series.