As banks ramp up investment in responsible AI, financial institutions of all sizes are turning to custom-built, focused language models as an alternative to large-scale generative AI tools.

Banks should develop smaller AI systems designed for specific business functions, allowing for better control, security and explainability, Scott Zoldi, chief analytics officer at FICO, told Bank Automation News.

Accessible models

Focused language models are gen AI-based and can be built by banks of any size without support from a large cloud provider, Zoldi told BAN.

“Just about any organization can build its own language models. And they should.”

— Scott Zoldi, chief analytics officer, FICO

FIs can gather their own data and develop custom language models tailored to their specific domains and tasks and apply advanced language and sequence modeling techniques while maintaining control over the data and model development process, supporting responsible AI initiatives, Zoldi said.

“That’s going to be incredibly powerful for each and every organization,” he said.

Smaller language models will be a game-changer for FIs that lack the resources of large banks, Amir Hajian, chief science officer at Arteria AI, a fintech that provides banks with AI-powered digital documentation services, told BAN in a recent episode of “The Buzz” podcast.

“It’s AI that you can use in real time, Hajian said. “It won’t break the bank, it’s accurate, it’s very flexible, and you can use it wherever you want, however you want.”

Banks that work with Arteria include the $1.7 trillion Citi and the $558.2 billion Goldman Sachs.

Data control

While “focused language models” is sometimes used interchangeably with “small language models,” FICO distinguishes the former by the focus on what data goes in and doesn’t go into the development of those models, Zoldi said.

FICO has developed its own focused language model and is exploring applications that will be powered by the model, such as pricing technologies, he said.

To further improve reliability in its own models, FICO has been refining a trust score framework which is used to assess how dependable a model’s output is on a scale of one to 999, Zoldi said.

“You don’t know if it’s hallucinating or making a bad outcome unless you know the [correct] answer,” he said, adding that trust scores will indicate when an outcome is likely incorrect.

With the addition of trust scores to help gauge the reliability of model outputs, focused language models will become standard across operations, with organizations maintaining 15 to 20 models used in different facets of their business, Zoldi said.

“It won’t be some sort of really interesting thing out in the sky somewhere … It’ll become common,” he said.