Consumer demand for instant access to banking and other financial services is redefining IT strategy for banks and financial service companies.

Banks are now looking to increase innovation and agility to accelerate software development while simultaneously combating the increasing risk of fraud and keeping an eye on cost-efficiency.

Many of these companies are looking towards artificial intelligence as a one-stop solution to combating fraud. It’s already being used to detect fraud in financial transactions based on real-time identification of discrepancies against a predefined set of rules. The number of AI deployments for this use case is increasing rapidly, to the extent that the business value of AI in banking is projected to reach $300 billion by 2030.

However, the full integration and scaling of AI solutions is complex, and the architecture must be streamlined before AI’s promise can be fulfilled. Jumping on the AI bandwagon for the sake of it is misguided; organisations should instead carefully consider the advantages and limits of the solutions they are looking to incorporate.

Integrating AI into fraud detection

AI right now exists as either a separate part of banks’ technology ‘stacks’ or as a grafted-on part of a larger database management system.

In order to deliver dependable real-time fraud detection however, all financial services firms must truly integrate AI to assist with tasks like transaction scoring – the process of classifying client transactions to identify customer behaviours. This isn’t a straightforward process, though. AI models delivered using software or platforms designed for processing huge volumes of data in batches, don't perform or have the appropriate operational characteristics for real-time or at scale uses.

To support fraud detection, AI should be deployed in a series of machine-generated, data derived algorithms. These algorithms should be streamlined into application form to allow them to ‘infer’ things about new data – essentially, identify potentially anomalous activities and flag behaviour that should be labelled as potentially fraudulent – based on their previous training in how to identify illegal transactions or behaviours. This process is commonly known as running ‘inference’.

The traditional home of AI has been in a computational or application layer; however, architecturally, inference operates more similarly to the data layer. This shift has massive implications. This is because putting AI architecturally closer to the source data and allows for inference to be performed alongside other data operations for optimisation purposes. Doing this is essential to future success in the context of ever-expanding data intensiveness and requirements, and the real-time data velocity requirements that modern applications demand.

Above all, AI must have all the markers of a high-performing database to perform at its optimum level. That means it needs to be highly available, stable, connected by a common interface, and able to be scaled in a cluster. And, of course, it must be able to respond within a real-time performance envelope.

Key considerations when incorporating AI

Large firms are currently investing significant resources in custom solutions to use AI in fraud detection, according to research by Deloitte, with a quarter of market frontrunners investing over $10bn a year in AI/cognitive technology-based projects. However, current AI solutions take time to deliver ROI within the sector, as banks have to consider the costs of developing and integrating AI models and product rollout. There’s no good shortcut for developing these models, so any efforts to economise typically come from running inference or developing a simple set of rules.

Most inference is fundamentally a simple process – you take a subject you want to run your rules against, feed it to the AI model and use the output from that model as your result. But this process often isn’t straightforward. It requires different programming languages, various tools and a series of unfamiliar concepts to the average enterprise developer. The real key to making it successful is to make the production process – the process of setting up inference – as easy as possible through the fast and persistent availability of memory.

The limits of current AI

Deploying AI for bespoke services demands the writing of tight, effective production-ready code, especially for the use of AI in fraud detection, which must happen in real-time and have a low occurrence of false positives. AI is still developing in this regard – the code and tools used by data scientists often require extensive customisation to become useful to enterprise developers and must be specifically modified to run at scale and in real-time.

AI works best when it has access to a large amount of compute power and high data bandwidth.

The squeeze to develop these low false-positive models means they’re often developed by data scientists, many of whom rely on retrieving data from disk, rather than from main memory. This disrupts developers’ attempts to orchestrate actual inference in real-time, as the seek time when searching for data is too long. Some tools are catching up, though, and inference is beginning to be treated as a real cog in the machine of enterprise software.

Overall, demand is growing for a more standardised approach that pulls and processes a variety of data sets simultaneously. Once the industry adopts this maturity in inference and opts for in-memory databases, AI’s use in fraud detection will become more widespread.