Wall Street’s Quiet Bet: LLMs Now Decide Who Gets a Loan

Lead

The next loan you apply for might never meet a human. Increasingly, a large language model trained on customer data, transaction histories and alternative signals will do the first, and sometimes final, judging. For consumers this promises speed and convenience; for investors and regulators it raises a stack of practical and legal questions.

Why this is happening now

• Falling inference costs. Cloud and GPU capacity from vendors like Nvidia have pushed real-time model checks into the realm of practicality.
• The underwriting lesson learned. Fintechs started handing off decisions to algorithms years ago; Upstart proved algorithmic underwriting can unlock approvals at scale. That capability is now cheaper and more portable.
• Demand for immediacy. Customers expect conversational interfaces and instant answers. LLMs can serve underwriting and customer-facing roles from the same codebase.

A short history and a bit of context

Think of credit scoring shifting from the slide rule era to something you carry in your pocket. FICO built a mid-20th century system around a simple, interpretable number. Machine learning widened the inputs. LLMs add contextual reasoning — parsing payment narratives, application text and public signals. That nuance can increase approvals, but it also makes the why behind a decision harder to reconstruct. What’s interesting is the trade-off: more subtlety, less transparency.

How it looks on the ground

Many banks and startups are using hybrid flows: models handle bulk cases and push odd or risky files to humans. Examples you’ll see in the wild:

• Upstart turned a model-first approach into a product. Peers and some incumbents are copying the architecture.
• Large banks are experimenting with internal LLMs to automate call-center triage, fraud detection and other bit-and-piece tasks, while renting heavy compute from public clouds.

Risks that matter — beyond the headlines

• Explainability. LLMs don’t hand over neat rationale. Regulators and courts will demand reasons for adverse actions. If a denial rests on an obscure token association, that’s a legal problem.
• Concentrated operational risk. Relying on a small group of cloud/GPU providers centralizes failure modes and bargaining power.
• Bias and drift. Macro shocks can shift model signals; what validated fine six months ago may stumble after a downturn.
• Adversarial gaming. Application text and public profiles are mutable. People — or bad actors — can engineer inputs to nudge outputs.

In practice, though, these risks aren’t theoretical. Teams underestimate how quickly a model’s edge cases multiply once it’s live.

Why investors should care

• Infrastructure pays. Chipmakers and cloud providers stand to gain from rising inference demand.
• Data moats matter. Firms that control payment flows or unique behavioral signals can train models others struggle to reproduce.
• Regulatory moves are binary and consequential. One enforcement action or legal precedent could either wreck a business model or raise the bar so high that compliance becomes a competitive advantage.

Signals to watch this quarter

• Mentions of AI-derived revenue or new cloud partnerships on earnings calls.
• Filings or guidance to CFPB, FDIC or state regulators about automated decisioning.
• Bank partnerships with niche vendors focused on model governance and auditability.
• Changes in approval rates, take-rates, charge-off guidance, or vintage performance after AI rollouts.

A contrarian note

LLMs won’t sweep away rule-based credit overnight. For high-volume, low-risk products, simplicity and interpretability still win. The real fight will be in the middle: small-dollar consumer credit and small-business loans where a bit of contextual understanding can move margins and approval rates.

Where this leaves us

Underwriting architecture is being rewritten. That promises efficiency and more personalized decisions, but it also forces a reckoning: balance exposure to infrastructure upside against model-specific and regulatory risks. For consumers, faster decisions mean greater access — and often less clarity about why they were accepted or rejected. The firms that win will be those that pair sophisticated models with rigorous governance and a tolerance for the messy work of operationalizing them.