Tech Time: Core Concepts of Artificial Intelligence
Just to get it out there: Artificial intelligence as a topic deserves the hype. However, the surface of AI has barely been scratched. Throughout history, the original incarnations of new technologies have often been dim views of their true potential. The original movie cameras, for example, were used to simply film stage plays, since theater was the medium people knew at the time. Tomorrow’s problems, and the technology we’ll need to solve them, are barely definable today. What is certain is that the technology used to solve those problems will be driven by AI. To take full advantage of the opportunities afforded by AI, it is important for credit unions to gain a better understanding of the core concepts that define this emerging area.
The Opportunity for AI
As the complexity of credit unions’ digital offerings continues to balloon, making sense of all streams of data is a major task that could be undertaken by a learning system. AI could present ideas and patterns not based on spotty anecdotes, but on a real-time, thorough analysis of available data.
This future is not far off—a credit union’s offerings soon will be tailored to a segmentation of an individual member, based not on marketing and product decisions made by a manager, but by a learning system that has analyzed that specific member’s behavior to provide relevant, completely personalized results. Interactions with members may even be automated, with a chatbot or AI voice-based assistant passing along the advice and guidance, as well as taking questions and modifying its output based on those questions.
Breaking Down AI: Data Science and Machine Learning
As is often the case with nascent technologies, terminology becomes mainstream with little or no true understanding. Two terms that often require some distinction are “data science” and “machine learning.” Machine learning is a tool used within data science. Data science encompasses a great many disciplines, including statistics, computer science and mathematics. Its tools are generally statistical analysis tools. Data scientists create models to simulate and predict statistical outcomes, reviewing various input parameters. The models are then run over and over with the inputs tweaked to generate differences in the output, which is later analyzed.
Traditional data science meets artificial intelligence in two key areas: the automation of model generation and the exponential increase in scope of the input parameters. A machine learning system first starts with a generalized model of what it needs to analyze, created by a data scientist. In supervised learning situations, the scientist teaches the algorithm what the model’s end state should look like. The AI then takes vast amounts of input and tweaks the model based on the errors it finds in describing and comparing the data to the supplied end state. When the learning process is complete, a model has been produced that has adapted to changing parameters across a vast data landscape and is now prepared to analyze new streams of data.
This is essentially how sophisticated machine learning-based fraud analytics work. The models are taught what fraud looks like and how it looks different from normal activity using safe activity records and fraudulent activity records. Models are then tested against known outcomes, and their effective rate is measured against a known quantity. So, after teaching my fraud model what fraud looks like, I could test it by passing it 100 normal cases and three fraud cases, seeing if it detects the cases I know to be fraud. When an acceptable recognition rate is reached, based on the use case, the model is deemed ready for general use.
The Future of AI in Fintech
The ability to analyze a tremendous amount of data practically instantaneously makes this an exciting area of development. Models are able to learn to adapt to data changes much more quickly than an analyst who is spot checking data manually. Modern software is built with rules to detect fraudulent behavior. These rules have to constantly be updated by a programmer. In machine learning, the rules are updated on the fly by the software itself, no human intervention required.
Fraud analysis is only one example of how machine learning is used to automate traditional analysis methods. In the fintech world, applications include automating time-intensive functions, such as contract and legal documentation review and generation of compliance reports. Machine learning can also be used to automate highly repetitive manual processes, like back-office item review and simple call center functions. Predictive analytics like those used for fraud detection are being used for marketing segmentation.
Still, these services only scratch the surface of where AI and machine learning are going. It’s difficult to envision the future. Much like how our film predecessors never foresaw today’s modern motion picture, some general directions in AI look promising. As more and more knowledge-worker tasks are automated, the knowledge-worker’s role changes into more of a teacher, and less of an implementer.
Think for a moment about the engineers who spend countless hours writing software that has to constantly keep up with the myriad ways your members interact with your credit union. What if those engineers could be given sophisticated models that already encompass a basic understanding of your business. Those engineers would be freed from maintaining every changing data point and could focus on higher-level trends, teaching their AI tools what they know, so that the tools can help inform decisions going forward.
Some may argue that using AI and machine learning removes the human element. However, what I see is an opportunity for credit unions to better engage members by using technology to tailor an approach specifically designed for each individual that is unique and more meaningful.
Danny Piangerelli is CTO at Malauzai, Austin, Texas.