In short: LLMs used in lending, whether chatbots, document extractors, agents or anything that feeds a credit decision, fall under the same expectations as other models: inventory, validation, monitoring, explainability and fair-lending testing. The practical rule is to keep LLMs out of the final credit decision, constrain what they can say to consumers, test them against fixed evaluation sets, and log everything. This article is an overview, not legal advice.
Lenders and lead businesses are adopting LLMs quickly: chatbots for borrower questions, agents for operations, extraction from bank statements and pay stubs, generated marketing copy. Each of these touches rules that were written before LLMs existed but still apply. Regulators have been clear that new technology does not create exemptions.
This article summarizes the main obligations and a practical control framework. Your compliance and legal teams should adapt it to your products and jurisdictions.
The Rules That Already Apply
Model risk management. Banking supervisors' model risk guidance (the Federal Reserve's SR 11-7, adopted with the OCC) expects every model to be inventoried, validated by people independent of its developers, monitored in production and governed. Many lenders and their partners apply the same standards even when not directly supervised. LLM-based systems that produce estimates or classifications are models under these expectations.
Adverse action and explainability. Under the Equal Credit Opportunity Act and Regulation B, lenders must give specific reasons when they deny credit or offer less favorable terms. The CFPB stated in Circular 2022-03 that this applies even when decisions rely on complex algorithms, and later guidance said generic checklist reasons are not enough when they don't reflect the actual drivers. The CFPB withdrew both circulars on May 12, 2025, but the requirement they interpreted, specific principal reasons under ECOA and Regulation B, is unchanged; see Explainable ML and Adverse Action Notices. An LLM's free-form judgment is very hard to explain this way.
Fair lending. Models must not discriminate on prohibited bases, directly or through proxies. Text and conversational data are rich in proxies (language, dialect, names, neighborhoods), so LLM-derived features need careful testing.
UDAAP and consumer communications. Anything a chatbot tells a borrower is a communication from you. Misstated rates, invented promises or an inability to reach a human can be unfair or deceptive. The CFPB has examined the risks of chatbots in consumer finance.
Calling and messaging rules. The FCC has ruled that AI-generated voices are "artificial" under the TCPA, so consent requirements apply to AI voice agents.
Privacy and security. Financial privacy and data-security rules apply to everything you send to an LLM provider, including transcripts and documents.
State AI laws and frameworks. Several states have enacted or proposed rules for "high-risk" AI, including lending decisions, and the NIST AI Risk Management Framework is a widely used voluntary reference. Expect this landscape to keep changing.
Where LLMs Fit, by Risk Tier
| Use | Example | Risk | Key controls |
|---|---|---|---|
| Internal analysis | Analyst copilot summarizing yield | Low | Access controls, accuracy checks |
| Data extraction | Income from bank statements | Medium to high | Field-level validation, human-labeled test sets, fallback to manual review |
| Consumer communication | Chatbot answering product questions | High | Approved content only, disclosures, human handoff, transcript review |
| Credit-adjacent features | Intent or hardship signals in a score | High | Proxy and disparate-impact testing, documentation, monitoring |
| Final credit decision | LLM approves or declines | Very high | Avoid; use explainable models |
A Practical Control Framework
1. Inventory everything. List every LLM use: vendor, model version, prompts, data flows, owner and risk tier. Include "shadow" uses by teams.
2. Separate deciding from talking. Keep credit decisions and pricing in validated, explainable models, as we recommend for default prediction and lead scoring. Let LLMs gather, extract and explain, but not decide.
3. Pin and version. Fix model versions and prompts in production. Treat a model upgrade or prompt change as a model change that triggers re-testing.
4. Build evaluation sets. For each use, maintain a test set with expected outputs: extraction accuracy, policy-compliant answers, refusal of out-of-scope requests, adversarial prompts. Run it before every change and on a schedule.
5. Constrain outputs. Ground consumer-facing answers in approved content, block unapproved numbers and claims, and require structured outputs where the result feeds a system.
6. Test for fairness. Check LLM-derived features and outcomes for disparate impact across protected groups, using proxy methods where you don't collect demographics.
7. Keep humans in the loop. Require human approval for consequential actions by agents, and give consumers an easy path to a person.
8. Log and monitor. Record inputs, outputs, model versions and actions. Monitor accuracy, drift, complaint themes and escalation rates.
9. Manage vendors. Review LLM providers' data use, retention, security and change-notification terms, as you would any critical third party.
Compliance as an Advantage
Teams that build these controls once can adopt new models quickly and safely, while competitors without them either move recklessly or not at all. For the architecture that makes this easier, with LLMs around the decision engine rather than inside it, see Agentic AI in Lead Gen.
Frequently Asked Questions
Can lenders use LLMs in credit decisions?
It is risky. Lenders must give specific adverse-action reasons and manage fair-lending and model risk, which is hard with free-form LLM judgments. The common safer approach is using LLMs for data extraction and communication while explainable models make the decision.
Does model risk management apply to LLMs?
Yes. LLM-based systems that produce estimates or classifications used in business decisions are models under model risk expectations: they should be inventoried, validated, monitored and governed.
What are the compliance risks of lending chatbots?
Misstating terms or making unapproved promises, failing to provide required disclosures, trapping customers without a human option, mishandling sensitive data, and, for voice agents, violating TCPA consent rules.
Want to use LLMs without creating compliance debt? Contact Plato AI to design LLM and agent systems with model risk controls built in from day one.