FICO, the eponymous credit score provider, announced three artificial intelligence language models for financial services firms on Tuesday: Focused Foundation Model, Focused Language Model and Focused Sequence Model.
In developing its own language models, FICO is offering an alternative to the large foundation models that have become popular in the industry, like OpenAI's ChatGPT, Anthropic's Claude, Google's Gemini.
According to FICO's Chief Analytics Officer Scott Zoldi, his team looked at the way banks were using generative AI, "where you use a pre-existing model that you have not trained yourself, and you don't actually understand what data is in there," he told American Banker. "And then you surround it with all kinds of tooling like retrieval augmented generation to try to get it to make a sensible decision."
This leads to lower accuracy decisions, bias and hallucinations "because you don't know exactly how that model was built," Zoldi said. "We said, for certain types of problems, particularly those in financial services, we need to build these models from scratch as small language models."
FICO conducted proofs of concept for these models in banks in the areas of underwriting, collections compliance and customer communication compliance.
"Models that are focused on a specific domain are likely to be more accurate and have fewer hallucinations," Mike Gualtieri, vice president and principal analyst at Forrester Research, told American Banker. "Financial institutions want to use AI, but have been cautious because of unpredictable results. FICO trains its model on a corpus of generally available financial information, but its customers will ultimately be able to create their own focused model using FICO's model training process."
FICO's focused language models focus on specific tasks, such as making lending or fraud decisions, or consumer compliance.
"The approach was on this concept of focus," Zoldi said. "Focus basically means we will curate the data for a specific purpose." A fraud detection model, for instance, will only be fed data related to fraud and scams. This leads to higher accuracy, fewer hallucinations and a higher trust score, he said.
This focus allows clients to have control over the data the models consume. "Many of our customers are concerned about what data is driving the decision, even if it sounds plausible, and now they have more control over it, and that plays out in higher and better accuracy and an ability to explain how this model got built," Zoldi said.
FICO's focused foundational model is essentially a small language model, according to Megha Kumar, research vice president at IDC.
"Small language models allow for improved accuracy because they are trained on domain specific information," Kumar told American Banker. "They also need lesser resources compute-wise and training levels do not have to as extensive as large language models, resulting in lower costs. Financial institutions are cautious with how they use AI and having a domain-specific solution that can easily be adapted will be attractive."
Every output of these models receives a trust score from 1 to 999. "The higher the score, the more trust you have," Zoldi said.
The trust score is based partly on coverage — how much data was available to support the answer. "If I have only three examples of a financial instrument in Kazakhstan, it'll produce an answer, but there's not enough data, there's not enough coverage and statistical relevance there," Zoldi said.
It's also based on "knowledge anchors," which are questions or prompts the model is supposed to answer. There might be 100 or 200 of these defined by experts at the financial institution.
"Think about the person who's responsible for the proper way to communicate with the customer," Zoldi said. "We take those knowledge anchors, and through a generative AI technique, independently validate whether or not the prompt and the response aligns with the knowledge anchors." If they don't, the trust score is likely to be low.
In proofs of concept, FICO has compared the work of its models against human underwriters, and it's done as well as some of the best underwriters, Zoldi said.
"Now your customer base is getting answers more quickly, more consistently and more accurately, and the human's job is to oversee it, review it and agree with it," Zoldi said. "But a lot of that heavy lifting has been done."
Gualtieri expects this kind of focused-model approach will become mainstream for every industry because it offers more accurate and specialized information compared to more generic AI models.
"This approach is a breakthrough in how all institutions will use AI models," he said. "FICO's innovation is not only showing the future of AI models for finance, but for every other industry. The potential weakness of this approach is that it is likely to lag the frontier models on new capabilities such as reasoning."
