While much of the activity in the AI markets are focused on the tech giants chasing ever-increasing model sizes and compute budgets, financial company FICO is going the other way with smaller, smarter, and purpose-built models specific to their client needs. With the launch of its Focused Foundation Model for Financial Services, consisting of Focused Language Models for Financial Services (FLMs) and a Focused Sequence Model for Financial Services (FSM), the data and analytics veteran is staking a claim that narrow beats broad when it comes to applying AI to real-world financial decision-making.
“Unlike general-purpose LLMs, FLMs are purpose-built for highly precise, deterministic decisioning and regulated agentic tasks,” says Dr. Scott Zoldi, Chief Analytics Officer at FICO. “They are not primarily intended for use in open-ended conversational chatbots.”
Built Small, Aiming Big
FICO claims that its FLMs can produce high quality results with a 1,000-fold reduction in computing requirements compared to general-purpose LLMs. That efficiency comes from deep domain curation. Rather than building their model on petabytes of text across the web, FICO trains its models on financial services data, filtered to retain only task-relevant patterns. The result is a compact, highly tuned model that produces higher quality results with faster inference.
These aren’t just theoretical gains. Using its focused models, FICO cites a 38% improvement in compliance adherence for a customer communications client in Asia Pacific, and a 35% ‘lift’ in transaction analytics based on U.S. credit card fraud detection data across U.S. banking clients. Both metrics were observed over a 24-month production period, using client baselines as reference points.
Zoldi is quick to point out that performance doesn’t degrade as complexity grows. “FLM performance scales predictably with data and task complexity,” he shared.
Compare that to what the big tech companies are doing. OpenAI’s GPT, Google’s Gemini, and Anthropic’s Claude are all pushing toward massive, generalist models meant to solve any problem thrown at them. But these models are expensive to fine-tune, unpredictable in regulated scenarios, and often opaque to audit. When hallucinations occur, they’re difficult to catch before damage is done.
In contrast, FICO’s models aren’t just smaller for efficiency’s sake. They’re meant to be less general-purpose and creative but more reliable. In financial services, that trade-off is necessary.
Trust Scores: The Anti-Hallucination Metric
Central to FICO’s GenAI framework is the concept of a Trust Score, which is a numeric score that indicates how closely a model’s output aligns with required task responses. These scores are derived using knowledge anchors, institution-defined examples of correct and incorrect behavior.
“Banks set thresholds based on their tolerance of potential hallucinations or damaging outcomes,” Zoldi explains. “Outputs with the lowest Trust Scores exhibit the highest levels of hallucinations, and where clients require consistent responses, these low trust scores reliably signal when the [model] has drifted.”
It’s a fundamentally different approach from confidence scoring in other LLMs. Trust Scores don’t reflect internal model certainty. Instead, they reflect distance from required levels of institutional truth. The logic is straight out of FICO’s legacy in credit scoring. Decision confidence isn’t just about probability, it’s about alignment with policy. Provide a score, and let organizations make a decision about what to do with that score.
Compare this to the healthcare sector, where startups like Hippocratic AI and Nabla are also attempting to rein in hallucinations with domain-specific LLMs. Most companies are still trying to figure out how to determine trustworthiness at inference time.
Financial AI With Guardrails
FLMs are not simply general-purpose models stripped down for efficiency. FICO says that they are structured from inception for regulated environments. That means co-designed hardware and software pipelines, explainable AI components, and built-in observability layers that can satisfy the most exacting compliance requirements, from GDPR to BCBS 239.
The financially-focused models are used in two phases. Phase 1 builds general financial knowledge, and Phase 2 localizes for region-specific laws and practices using synthetic and seed data from clients. This results in highly contextual models that understand not just what to say, but how to say it in ways that won’t land a bank in regulatory hot water.
Zoldi emphasizes that this isn’t just localization by translation. “All regions have different challenges, but the biggest challenge is that organizations struggle to define their seed of task data. Experts often disagree on interpretation and that takes time.”
Clients can layer domain-specific capabilities on top of the base FLM using post-training with their own data. Rather than self-service fine-tuning, FICO guides the process, incorporating red-teaming and rigorous validation.
“Today, we see clients most interested in defining the business problems to be solved with FLMs, optimizing KRIs/KPIs and obtaining value, not in directly training models themselves,” explains Zoldi.
Leaning Into Specialization
In other industries, similar specialization is emerging. In legal tech, Casetext (now part of Thomson Reuters) launched CoCounsel, trained exclusively on legal documents and optimized for legal reasoning. In biotech, companies like Recursion and Genesis Therapeutics are creating small foundation models trained on molecular data.
In the financial services industry, there’s similar activity happening. While JPMorgan has filed patents for its own finance-specific AI models, few banks have the infrastructure or domain depth to pull off what FICO is offering out of the box. For many, financially-focused models represent not just a tech play but an outsourcing of risk.
That vertical focus also allows FICO to sidestep many of the ethical issues plaguing more open-ended genAI systems. Hallucinations, data opacity, and explainability gaps are easier to manage in narrower scope data and task environments.
FICO’s ecosystem integration is another strategic lever. The models feed into its broader analytics suite, drawing on decades of credit risk modeling and compliance tooling. This creates a feedback loop where FLMs improve not just in language capabilities but in domain-specific decision accuracy.
“Our models don’t operate in isolation. They’re deeply integrated across FICO’s entire analytical ecosystem and continuously learn from real-world data points,” Zoldi says.
Looking Forward
Over the next 12–24 months, FICO expects to expand multilingual support and roll out updates to its base models optimized for evolving enterprise needs. Expect more agentic functionality too, as FLMs are prepped to serve as autonomous actors in AI-driven workflows.
And in keeping with its Responsible AI stance, FICO plans to extend its auditable AI blockchain to financial model development. Every training decision, every anchor update, every production deployment will be logged, traceable, and reviewable.
“It is critically important for language model providers to walk-the-walk, not just talk-the-talk,” Zoldi says.
