Technical Guide | Financial Services

    Shadow AI in Bank Lending Operations: Fair Lending Exposure From Unapproved AI Tools

    Shadow AI in bank lending creates fair lending exposure because unapproved tools can receive applicant data, shape credit decisions or adverse action wording, and leave no record that model risk management or compliance can review. Written policy and after-the-fact review show intent. Policy enforced at runtime, in the execution path, can show what happened in individual interactions and tool calls.

    1. Inventory lending workflows and the data each handles
    2. Tier uses by risk and set rules per tier
    3. Include vendor and embedded AI in scope
    4. Tie AI-influenced steps to existing fair lending testing
    5. Test whether per-interaction evidence can be produced on request

    Questions to ask when evaluating runtime controls for lending

    • Does enforcement operate in the execution path, or does it discover and report after the fact?
    • Which surfaces are covered: AI sites, coding assistants, agents, MCP servers, models and tool calls?
    • Which actions apply per interaction (allow, redact, block, escalate for human approval), and how are approvals recorded?
    • What evidence is produced for each interaction, and can it be exported for examiners and model risk management?
    • Where is evidence stored, how long is it retained, and can retention and residency be configured?
    • Can the control be self-managed in the bank's environment, and which data leaves it?

    What shadow AI means in a lending context

    Shadow AI in lending is the use of generative AI tools, browser-based AI sites, coding assistants and embedded AI features by lending staff and adjacent teams without model risk management or compliance review. The users are not only underwriters. Credit analysts, collections staff, marketing teams, vendor managers and developers who maintain decisioning code all have reasons to reach for these tools.

    The core problem is that the bank cannot say how the tool behaves, what it received, or what it produced. That matters because fair lending obligations do not depend on which tool generated an output. A federal document from 2024 states that AI use in financial services may complicate compliance with fair lending and anti-discrimination laws, and it names the Fair Housing Act, the Equal Credit Opportunity Act (ECOA) and the Fair Credit Reporting Act. As of November 2026, nothing in that framing suggests an unapproved tool changes a bank's responsibility for the outcome.

    This is therefore an evidence and control problem, not only a policy-writing problem. A policy that prohibits unapproved tools does not tell a risk leader whether anyone used one on Tuesday afternoon.

    Mapping the exposure points to fair lending and model risk concerns

    The 2021 interagency request for information on AI in financial institutions highlighted three areas of risk management challenge: explainability, data usage and dynamic updating. They give a workable structure for seeing where an unvalidated tool leaves a documentation gap.

    Data input. Applicant data and credit criteria pasted into an external tool leave the bank's controlled environment. The request for information notes that AI may use traditional or alternative data, and commenters have flagged alternative data as raising heightened fair lending concern. A bank cannot assess whether a variable has a close nexus to creditworthiness if it does not know the variable was used.

    Output influence. A tool that summarizes a file, suggests a risk view or drafts a credit memo can steer the human reviewer. If the output reflects proxy variables or patterns learned from historically biased data, the effect may be disparate treatment or disparate impact. Outcome testing of the kind banks describe for approved models cannot cover a tool that is not in the inventory.

    Adverse action explanations. CFPB Circular 2022-03, issued May 26, 2022, addresses adverse action notification requirements for credit decisions based on complex algorithms. When an unapproved tool drafts or rewords the reasons given to an applicant, the explanation may not match the actual basis for the decision, and no record shows how it was produced.

    Third-party and embedded AI. Vendor features and AI embedded in everyday software can change how work is done without a procurement or validation event. Industry commenters have argued that banks should hold detailed knowledge of vendor models, meaning the categories of variables, the techniques and the key outputs. That is difficult for features the bank does not know it has.

    Missing records. Model risk management frameworks typically cover development, implementation and use, validation, and governance, policies and controls. A tool used outside the framework has none of these, and the bank has nothing to show for it.

    What policy and after-the-fact review can and cannot show

    Policy, approvals, training attestations and periodic review establish what the bank intended and what it formally authorized. They are necessary. They do not establish what happened in a specific interaction.

    Discovery-based approaches have a similar limit. Scanning for AI domains or applications can show that a tool was reachable or visited. It generally does not show which applicant data was submitted, what came back, or whether the output reached a credit decision or a customer communication. Reviewing activity after the fact also means any exposure has already occurred.

    This gap matters most when a risk leader is asked to show which AI tools touched lending workflows, what each was permitted to do, and what actually occurred. Documented approvals answer the second question only for sanctioned tools. For the other two, the bank needs records produced at the moment of use.

    Tiering helps decide where the gap is least tolerable. Industry commenters have described banks using dynamically updating models mainly in lower-risk uses such as marketing or customer interaction, and avoiding them for credit origination, where explainability is required. A sensible policy applies the same logic to unapproved tools: stricter handling for origination, underwriting support and adverse action drafting, more latitude for low-risk drafting that involves no applicant data.

    What runtime enforcement in the execution path adds

    Runtime enforcement places policy where the interaction happens, between the user or agent and the model, tool, site or coding assistant. Each request can be evaluated against rules and then allowed, redacted, blocked or escalated for human approval. Because it acts on the individual interaction or tool call, it produces a record of that interaction rather than an inference drawn later.

    In lending, the practical uses are specific. Redaction can remove applicant identifiers before a prompt leaves the environment. Blocking can stop credit criteria or underwriting rationale from reaching an unapproved site. Escalation can route a request that touches adverse action language to a person who approves or rejects it, with the decision recorded.

    There are tradeoffs. Rules that are too broad interrupt legitimate work and push staff toward unmonitored personal devices. Rules that are too narrow miss the surfaces where use actually occurs. Enforcement also does not validate a model, test outcomes for disparate impact, or replace fair lending review. It narrows the gap between written policy and runtime behavior and supplies evidence that those other functions can use.

    Trussed enforces policy in the execution path across models, agents, tools, MCP servers, AI sites and coding assistants, with actions that include allow, redact, block, rate limit and escalate for human approval, and generates continuous audit-ready evidence from that enforcement. It supports self-managed deployment in the customer cloud with configurable retention and data residency. Its scope is individual interactions and tool calls. It does not provide model provenance or end-to-end tracing of agentic workflows. Trussed is deepest in insurance among regulated financial sectors, and its approved production proof includes 18+ months of continuous production in a regulated enterprise, 99.95% uptime on live production traffic, and 100+ models under one policy plane.

    Shadow AI exposure points in lending

    Data input

    Applicant data, credit criteria and underwriting rationale entered into unapproved tools.

    Output influence

    AI-generated analysis or summaries that shape a credit decision without review.

    Adverse action reasoning

    Drafted explanations that may be inconsistent or undocumented.

    Third-party and embedded AI

    Vendor features and browser tools that sit outside model inventory.

    Missing records

    No evidence of which tool was used, what it was permitted to do, or what occurred.

    Assess whether your controls can produce evidence, not only policy

    Walk through your lending workflows with our team and see how enforcement in the execution path and audit-ready evidence apply to your model risk and fair lending questions.

    Talk to an Expert
    Next step

    How does your AI governance program compare?

    See where your program has gaps in less than 2 minutes.