How LLM Assistants Create Fair Lending Risk in Pricing Exceptions

Why pricing exceptions are the exposure

Examiners already look closely at pricing discretion. FDIC fair lending guidance lists broad discretion in loan pricing as a risk factor and notes that discretion can exist even when rate sheets and fee schedules are in place, if loan officers can deviate from them without clear and objective criteria. The same guidance asks whether an institution that allows discretion tracks exceptions to pricing criteria, and it calls for special attention where financial incentives accompany that discretion.

CFPB ECOA baseline procedures ask similar questions: whether pricing exceptions are allowed, what guidelines govern them, how often they occur, and what standards apply to justifying and documenting them. As of October 2026, Regulation B, as published in the eCFR, prohibits creditor practices that discriminate on the basis of the factors it lists.

An LLM assistant that drafts, justifies or recommends exception requests sits inside this existing scrutiny. It does not create a new category of examination topic. It adds a new channel through which discretion is exercised, and one whose individual interactions are usually not recorded in a form an examiner can review. Federal Reserve and DOJ material from 2018 shows how discretion plus incentives has produced higher prices for minority borrowers even after legitimate factors were considered. That history is the baseline for assessing assistant-influenced pricing.

What interaction-level enforcement can and cannot address

Enforcement in the interaction acts on individual prompts, retrievals, outputs and tool calls. Examiners, by contrast, assess pricing through comparative file review, statistical analysis, or both, comparing borrowers by prohibited-basis group under the institution's own standards. Those are outcome methods, and interaction controls do not substitute for them. Redaction is also partial by nature: a rule can remove fields it knows about, but it cannot enumerate every proxy.

• Addressable in the interaction: Redacting protected-class fields and known proxies from prompts and retrieved data; blocking rationales that cite non-permitted reasons; restricting which pricing actions the assistant may take.
• Addressable in the interaction: Escalating exception recommendations or pricing-system writes for human approval, and recording what was permitted, attempted and approved.
• Outside the interaction: Whether the exception criteria themselves are objective, and whether incentives attached to discretion are appropriate. This is pricing governance.
• Outside the interaction: Whether the underlying model behaves consistently and without learned bias. This requires model validation, and explainability remains a recognized fair lending challenge for AI.
• Outside the interaction: Whether realized pricing differs by prohibited-basis group after controlling for legitimate factors. This requires statistical and comparative fair lending testing.

Policy rules to define before deployment

Rules need a documented standard to enforce. Start with the institution's list of permitted exception reasons, since examiner reviews turn on guidelines, justification and documentation. Then express the controls as allow, redact, block and escalate decisions.

• Input redaction: Remove protected-class attributes and approved proxy fields from prompts and retrieved context, and flag free-text references to them.
• Data access scope: Limit retrieval to the fields the exception process documents as relevant, by role and product.
• Rationale validation: Block or return for revision any generated rationale that does not map to a permitted exception reason code.
• Tool-call restriction: Default the assistant to read-only or draft-only against pricing and LOS systems. Explicitly list any permitted write actions.
• Human approval: Escalate any rate adjustment or system write to a named approver role. Consistent with OWASP guidance on confirming high-risk actions, treat pricing changes as that class.
• Change control: Version the rules and review them under the post-deployment practices NIST describes, including monitoring, incident response and override.

Placing this within model risk management

Treat interaction enforcement as one layer. Keep model risk management, compliance and fair lending testing as distinct roles. CFPB procedures ask whether models and processes are periodically tested for fair lending risk and how risk from service providers is managed, so the assistant's model provider belongs in that service-provider view. Interaction logs do not replace outcome testing, and controlling the assistant does not remove the discretion and incentive risks that exist without any AI.

A practical sequence: define permitted exception criteria and approver roles in pricing governance; validate the model behind the assistant; enforce rules in the interaction; and feed per-interaction evidence into the same monitoring that tracks exception frequency and pricing disparities by group.

Trussed enforces policy in the execution path for individual interactions and tool calls, with allow, redact, block, rate limit and escalate for human approval, and generates audit-ready evidence from that enforcement. Its depth is in insurance, a regulated financial sector, and it does not perform statistical fair lending testing or replace pricing governance.