Banks Are Quietly Handing Credit Decisions to LLMs — Investors, Pay Attention

A new wave is rolling through finance — and it already smells like underwriting

For the past 30 years credit scoring evolved by accretion: FICO, then ML classifiers on tabular data, then alternative signals like mobile and transaction histories. What’s different this time is both scale and texture. Large language models can read messy, unstructured inputs — emails, call transcripts, chats, scanned documents — and turn them into features underwriters can actually use. That changes the input set in a way that matters.

Why banks are trying this

• LLMs can shave hours off underwriting by extracting facts from messy documents, cutting down on manual review and easing straight-through processing.
• They can surface narrative-based fraud signals that traditional models miss, and adapt faster when new tricks appear.
• For lenders the arithmetic is simple: faster decisions, fewer reviewers, lower operating costs — which matters most for thin-margin credit products.

But it isn’t magic.

Concrete trade-offs (and real headaches)

There are three messy problems investors should care about.

• Explainability. Regulators and compliance teams expect a traceable reason for adverse actions. LLM outputs are often opaque, so banks are being pushed toward hybrid stacks that keep an explainable model alongside the LLM.
• Data and bias. Narrative text carries socio-economic cues that correlate with protected classes. Without strict testing and governance, lenders risk amplifying disparate impacts and inviting legal trouble.
• Concentration risk. A few cloud providers, model vendors and chipmakers are becoming the plumbing for modern underwriting. If everyone uses the same models or embeddings, systemwide fragility increases.

None of these are theoretical. They show up in audits, litigation risk assessments, and even in the cost of capital.

Where you’ll see this first

LLMs will augment front-line workflows, not rewrite credit policy overnight. Expect early deployments in places like:

• Ingesting and classifying documents for mortgages and small-business loans.
• Fraud triage for debit and credit originations, using call and chat transcripts.
• Automated red-flag detection during merchant onboarding for commercial lending.

The immediate market effect isn’t just lower costs. It’s margin pressure on smaller lenders and a possible edge for banks that hold their own models and data.

Winners and losers — an investor’s checklist

• Winners: AI infrastructure vendors and chipmakers that get both capex and recurring cloud spend.
• Cloud and platform leaders that can bundle compliance tooling and model hosting — sticky revenue if done well.
• Losers (or at risk): small banks and non-bank lenders who outsource models; they concentrate operational and legal risk.

Look for these signals in earnings calls and filings:

• Clear disclosures about LLMs in production underwriting.
• Capitalized software or strategic partnerships with model vendors.
• Compliance memos describing explainability frameworks or adverse-action workflows.

Practical questions for management

• How big are pilots and how dependent are they on third parties?
• Are they investing in in-house model governance and data lineage, or outsourcing the hard parts?
• Which suppliers of inference chips and cloud services are they tied to — infrastructure choices often precede revenue upside.

A broader view

This moment echoes earlier inflection points. When FICO and credit bureaus centralized decisioning, credit expanded — and sometimes too quickly. LLMs could widen the data underwriters use and lower frictions for credit to flow, which is a macro tailwind for consumer lending volumes. But if errors or bias scale, it becomes a policy headache.

For investors the right stance is nuanced: favor infrastructure plays and disciplined banks that build governance and auditability into their stacks, not firms chasing novelty. The next big winners will be those that make these models explainable and resilient — and find a way to charge for that reliability.