A leap from FICO to foundation models

The consumer loan aisle is shifting faster than most customers — and many bankers — realize. What began as scorecards, then moved to gradient-boosted trees, is now flirting with LLM-powered underwriting. Regional and community banks, short on engineering talent but desperate to keep pace with fintechs, are piloting generative AI to speed underwriting, automate income verification and tailor offers.

Why banks are rushing in

• Speed and scale. LLMs can compress days of paperwork into minutes of automated parsing and income reconciliation. When margins are thin and attention is fleeting, that matters.
• Better signal extraction. Early pilots suggest models pull signals from nontraditional sources — paystubs, emails, rental ledgers — that legacy credit models often miss.
• Cost pressure from fintechs. Startups rewired consumer-lending economics with ML; traditional lenders feel the squeeze and are following suit.

Not just efficiency — survival, practically speaking

Big banks have proprietary data and in-house ML teams. Smaller institutions do not. For them, third-party LLM stacks — via cloud partners or specialist vendors — are less a tech vanity project and more a way to stay in the game: faster decisions, fewer manual reviews, smarter cross-sell. It’s a defensive play as much as an offensive one.

Real risks under the convenience

Speed is seductive. But there are real tensions here.

• Regulatory exposure. Explainability matters in fair-lending reviews. If a model declines an applicant because of a latent text pattern, how do you produce a defensible reason to a regulator? CFPB and OCC examiners are flagging opaque approaches more often.
• Bias and disparate impact. Models trained on web-scale data can inherit social biases and unintentionally amplify proxies for protected classes. Even well-meaning features can skew outcomes.
• Third-party concentration. Smaller banks outsourcing AI end up depending on a few cloud and model vendors. That concentrates systemic risk — an outage or a problematic audit at a vendor ripples through many firms.

A quick history: FICO to explainable ML to foundation models

Credit scoring has always been a trade-off between predictive power and interpretability. FICO brought standardization. Tree-based models boosted performance while keeping some interpretability. Now foundation models push the needle toward performance at the cost of transparency. Regulators and technologists are scrambling to respond — and sometimes they’re a step behind.

Guardrails being tested in practice

• Use LLMs to augment, not to finalize. Human-in-the-loop remains a common mitigation.
• Keep an auditable feature set before the LLM step so you can point to defensible inputs in reviews.
• Run adversarial fairness tests and counterfactual simulations across demographics.
• Prefer smaller, domain-tuned models when they meet the needs, rather than defaulting to opaque web-scale giants.

What’s interesting is that some of the most effective controls are simple: transparency about inputs, careful logging, and conservative use-cases.

What this means for consumers and investors

Customers may see faster approvals and more personalized pricing. But hidden denials or subtle price segmentation tied to biased signals are real risks. Investors should watch banks that loudly promise AI-driven margin gains but lack governance; those firms face reputational and regulatory costs down the line.

Who to watch

• Fintechs that sold the ML underwriting promise are expanding into adjacent credit products.
• Cloud providers and chipmakers remain the invisible rails behind these pilots.

At stake

Generative models are not a cure-all for underwriting. They deliver real gains, yes, but also open new risk channels: explainability gaps, bias amplification, and concentration around a few vendors. The next 6–18 months will be revealing. Firms that pair sensible governance with realistic product scope will have the advantage; those that rush without guardrails may face audits, fines, or worse.

Pedro Marini