The term Artificial Intelligence (AI) is inescapable in Fintech today. It is often thrown around, with little knowledge of what it actually means.
One certainty is it isn’t a cure-all for many of the issues affecting the industry. AI opens a host of questions, beginning with how it should be regulated and continuing through how it is most effective.
Fraud detection is one area in which AI can contribute. How can AI be deployed to accelerate the delivery of personalized, multilingual fraud alerts? How can it be paired with data visualization to spot fraudulent patterns within complex data sets?
Rosaria Silipo has some of the answers. The head of data science evangelism at KNIME, Silipo has three decades of experience working with AI.
She is in a unique environment at KNIME, a leading open-source platform for low-code/no-code data analytics. KNIME supports more than 300,000 global users on its open-source KNIME Community Hub. It facilitates the creation of end-to-end data workflows and enables machine learning and advanced analytics.
AI isn’t new. Three decades ago, it was discussed in a digital network context before progressing into web data, data science and mining, and deep learning. Today, AI is more about consuming data from elsewhere.
Silipo said an early issue for AI was that the hardware didn’t keep up with expectations. That led to memory and calculation problems. Long-term and short-term memories weren’t improved, and the hardware didn’t have the power calculators required.
Then came the Netflix challenges in the mid-2000s, which had the goal of developing accurate recommendation engines, which built on gains in data mining.
However, hardware still wasn’t up to the challenge. That left AI with little business value.
The climate soon changed as deep learning and new hardware allowed old algorithms to be improved for business usage. Deep learning networks grew. Companies invested in architecture that could train deeper neural networks.
KNIME offers two products. The first is KNIME Analytics Platform, a free, open-source low-code/no-code software that allows anyone to make sense of data. KNIME Hub allows users to collaborate on solutions created with the Analytics Platform. The Community Hub is open to the global community while the Business Hub is installed into private infrastructure.
It produces an interactive dashboard. Models can also be trained to clean data as it is ingested and trained in neural networks.
“We like to keep (the Community Hub) free and open source so everybody can use it and contribute to expanding it,” Silipo said.
Open source is a necessity because the community is needed to synthesize the volume of data and algorithms. That volume of participation helps stakeholders stay in front of new technologies.
Silipo said there are many theoretical questions related to regulations. AI can quickly improve processes, but left unguarded, that comes with costs. Assurances must be included to ensure accuracy, privacy and ethical behavior.
“Whatever your AI produces, you need to analyze it and ensure what you’re giving to somebody else is safe,” Silipo said. “It’s a very early stage for these kinds of things.”
Developers are responsible for knowing the laws and regulations impacting your product. Silipo said some companies have AI auditing departments that check models.
Multilingual alerts are an important development in fraud detection and customer service. AI has a clear role in detecting fraud, but it can also be used to communicate with customers quickly and concisely.
Silipo said communication can be harder than detection. The process might begin with developing language templates for the most popular customer languages. The templates are developed, and the various languages are entered in the fields.
Then, those systems have to be maintained and updated as new alerts arise. Even if they don’t know a certain language, companies must ensure their communications are accurate. As the number of languages they need to communicate in grows, maintenance becomes more complicated.
“With AI, it’s easy,” Silipo said. You can write a warning that says particular things.” Provide a large language model the correct references for a new law, and it can generate appropriate communication in the needed languages.
Silipo worked at ING and Rabobank before joining KNIME. Those experiences taught her that in 2024, AI has limits. Changing the input, even slightly, changes the results. What was the logic behind the results? More progress is required.
“The part where I couldn’t control the output reliably… That was the part that stopped me from using AI for the outlier detection,” she explained.