Payments, fintech and the Age of AI. Five predictions for the future.

AI is going to eat the payments industry and bring significant wealth to investors, bank partners and merchants who embrace it. The biggest winner will be consumers, who can benefit from great deals and personalised offers.

The problem

Current payment networks, built with 1970s technology, are ill-equipped for our digital world, let alone the coming impact of Deep Learning (AI). This is largely because old-fashioned payment systems technically require the use of a convoluted intermediary model, with a chain of participants to execute transactions.

The current system is based on a “pull” mechanism. When you pay with your card, you authorise the person you are paying to withdraw money from your account. The card numbers give them the information required to access your account. Essentially, payment chain participants all “pull” money from the customer’s bank account to the merchant’s bank account via the payment network.

You can see how it works in the diagram below.

Don’t laugh, you pay for this.

If aliens arrived today and compared our payments system to how sophisticated the rest of our digital world is, they would find it insane.

The legacy payment network has a number of fundamental problems.

Technical inefficiency. The payment network is a patchwork quilt of numerous players cobbled together as the system evolved over time. From a modern architecture perspective it lacks efficacy, to be polite.

High risk. The old-fashioned “pull” model means that account and identity information need to be shared throughout the payment chain, exposing user data to multiple attack vectors. Card fraud was $11.91 billion in 2021[1]  and has increased every year for over a decade[2]; in addition, 65% of card users will be victims of fraud at some point in their lives.[3]  There is no built-in way to associate a card number with an account holder.[4]  As a result of its architecture, the current payment model is intrinsically unsafe.

High cost. The technical inefficiency and risk characteristics of the card payment model naturally result in high costs. In the United States in 2021, merchants paid over $160 billion just in one part of payment fees.[5] The amount is similar, although lower, in Europe. These costs and inefficiencies disproportionately affect smaller merchants and online transactions.

Bad for the environment. Six billion payment cards are produced annually, adding 5.7 million tons of plastic to landfills each year.[6]

In summary, the current payment system is outdated, inefficient and insecure. Consumers pay for this through higher prices.

And, the system cannot be fixed technically. It is like devices before smartphones. You cannot upgrade an old Blackberry and stream Netflix on it. You have to throw it out and get a smartphone if you want modern technology.

It gets worse

The structure of the current payment system is not well suited for AI, as set out below.

Siloed data. The proprietary and segregated datasets held by intermediaries in the payment chain limit or do not allow AI to use input parameters. Even where some input parameters are available, since AI cannot access comprehensive data due to it being siloed, output accuracy is reduced. Payment networks like Visa, Mastercard and others operate independently and have their own data repositories. Even when attempting techniques like federated learning, to the extent that data is surfaced to third parties it is arbitrary, time-limited and not conducive to continuous training of large language models (LLMs).

Heterogeneous data. The heterogeneity of input data makes the use of LLM methods difficult. An example of payment data field diversity is set out below.

·       Numbers (e.g. formatting)

·       Categories (e.g. how transactions are classified between different companies)

·       Text (e.g. non-standardized and arbitrary event descriptions)

These heterogeneous fields increase perplexity for LLMs, reducing outcome accuracy. Heterogeneous and multimodal inputs make working with payment data even more complex than with text, music or images.[7] 

In addition, this information is often in different tabs or data structures. All of this means that even if these inputs can be used in an LLM they require considerable pre-processing for data cleaning.

Not comprehensive. Card payments, steps removed from bank accounts, provide inferior input parameters than the accounts themselves. In addition, the average user has 3.23 credit cards[8] so inputs from a single card issuer do not give a whole picture of user spending. Even if the issues of heterogenous data can be solved, the output would not give a complete insight into user behaviour.

For AI models intended to analyse transaction data, predict fraud, assess credit risk and generate insights into users, the lack of comprehensive and uniform data is a serious limitation. LLMs are only as good as the data input parameters used and the training they get. The intermediary structure of the current payments system makes this challenging.

The AI solution

Like many great technology inventions (think, the internet), AI cuts out inefficient middlemen and replaces them with direct connections. Where Kodak used to have a vast photo processing network people can now simply use their phone camera and send images between themselves.

AI will likely transform the relationship between customers and merchants into a direct one, cutting out card networks. This will replace the card “pull” model with a “push” model, where the user is in control, significantly reducing technical complexity and risk. The result will be much lower costs.

Direct relationships will also solve the AI data issues, allowing LLMs to be trained on better datasets. This would result in more accurate outputs. It could also open up benefits for merchants and their customers such as the ability to personalize the shopping experience.

Drivers of change

AI will not do the work all alone. There are other factors making this possible, set out below.

API technology. Application Programming Interfaces are a way for computer networks to communicate with each other. APIs are like pipes that allow data to flow, for example, from a bank to a technology company. This allows a bank to offer banking services via a fintech. Or open banking providers to enable customers of the same fintech to connect their existing bank accounts and seamlessly move money or share data.  

Open banking regulation. Open banking gives users control over their information and standardises data sharing. It was first introduced in the European Union via PSD2 in 2018 and has been (or is being) implemented in other jurisdictions like the UK and US. A key requirement for AI is data input homogeneity. Open banking is likely to result in unbundling, weakening the relationship between possessing customer information and sale of that bank’s products. Together with technology, this data access will put smaller banks on a more even footing compared to large money centre banks.

Five predictions about payments and fintech in the Age of AI

AI will radically disrupt low-tech, data-siloed, legacy businesses like payments. The likely impact on payments and financial technology are set out below.

1.     A direct relationship between merchants and their customers will emerge, cutting out old payment networks.

2.    Payments will become cheaper, faster and less risky.

3.    The richer datasets will allow merchants to offer personalised rewards.

4.  The user insights generated (including customer shopping preferences and merchant reward incentives) will eventually be a competitive alternative to search advertising.

5.    AI companies and their partner banks are likely to capture most of the payments market value over the next 5 years. They will also use the data to offer personalised financial products to users, transforming some fintechs into the largest players in the country.

[1] Card fraud losses, United States, Nilson Report, 2022. Global card fraud losses were $32.34 billion. US was 36.8% of global fraud and US makes up 25.2% of global GDP (World Bank, 2022). The total value of transactions using cards issued in SEPA amounted to €5.40 trillion in 2021, of which €1.53 billion was fraudulent (European Central Bank).

[2] Fraud has increased even with the introduction of new technologies like digital wallets. The underlying card “pull” model is intrinsically high risk.

[3] 2023 Credit Card Fraud Report, Security.org.

[4] Meaning: whoever has the card details can spend money from the card account; there is no biometric link to the bank account holder.

[5] Merchants pay over $160 billion just in interchange fees, costing the average American family over $1,000 a year, Nilson Report, 2022. Interchange, assessment and processing fees make up what is known as the Merchant Discount Rate, which costs merchants about 2% of sales. The MDR is usually higher for smaller merchants and online (card not present) transactions.

[6] Mastercard, Euronews, 2021.

[7] Deep Learning and Large Scale Models for Bank Transactions, Garuti et al., Prometeia Associazione, 2023.

[8] US adults. PaymentsJournal, 2022. The amount is similar in Europe.