Guide for lenders
An AI Governance Framework Lenders Can Actually Run.
Most organizations have more AI than they realize. Far fewer can say how it got there, what problem it solves, what it changed, or who owns it. This guide is the framework I build with banks, credit unions, and independent mortgage banks: how to inventory what you have, tier it by risk, write policy people will follow, oversee your vendors, and give the board a clear view of decisions, controls, and open risks.

Why Lenders Need AI Governance Now
AI can enter a lending organization through an approved project, an added vendor feature, or a tool an employee starts using independently. I begin by making that activity visible across origination, underwriting, servicing, and internal operations: what is being used, which data it touches, and who is accountable for the outcome.
A governance framework is not a brake on innovation. It is what lets you say yes with more confidence, because the rules of the road are written down. A documented approval path gives business, technology, compliance, and risk teams a shared basis for deciding what to pilot, what to limit, and what not to deploy.
The framework
Six Pillars of a Working Framework.
Inventory and Ownership
You cannot govern what you have not counted. Start with a register of every place AI or automated decisioning touches the business, including the AI hiding inside vendor tools you already own, and give each one a named human owner.
Use-Case Risk Tiering
Not every use case carries the same risk. Tier them: low-risk internal productivity tools, customer-facing assistance, and anything that touches credit, pricing, or borrower treatment. Each tier gets its own approval path and review cadence.
Policy and Standards
A short, readable policy that states what is allowed, what needs approval, what is prohibited, and who decides. Add standards for data use, human review thresholds, documentation, and disclosure. If staff cannot read it in ten minutes, they will not follow it.
Vendor Oversight
Buying an AI-enabled service does not remove your responsibility to oversee its use. Extend third-party risk management to cover model updates, data use, fair lending testing, decision records, and available audit rights. Ask the hard questions before signature, not after go-live.
Board and Executive Reporting
Boards do not need model math. They need a one-page view: what AI is in use, what changed this quarter, what incidents occurred, what is coming, and where the open risks sit. A standing agenda item, not an annual surprise.
Monitoring and Evidence
Governance is a loop, not a binder. Track accuracy drift, exception volumes, overrides, and complaints. Keep decision logs, version history, and review records so you can answer an examiner with evidence instead of memory.
Risk tiering
Tier the Risk Before You Write the Rules.
Tier 1 · Low Risk
Internal drafting, summarization, meeting notes, knowledge search
Approved tool list and a simple acceptable-use policy. Annual review.
Tier 2 · Moderate Risk
Customer-facing chat, marketing content, borrower communications, process automation
Use-case approval, disclosure review, human review path, quarterly monitoring.
Tier 3 · High Risk
Credit, pricing, underwriting support, quality control, anything touching borrower treatment
Executive sign-off, fair lending review, documented validation, continuous monitoring, and a rollback plan.
Board reporting
What the Board Actually Needs to See.
A standing AI agenda item, one page, every quarter. These six elements give the board real oversight without burying them in technical detail.
- AI inventory: what is in use, by whom, and what changed this quarter
- Risk posture: tier counts, open issues, and anything past its review date
- Incidents: errors, complaints, overrides, and how each was resolved
- Vendor movement: model updates, new features enabled, contract changes
- Pipeline: proposed use cases awaiting approval, with expected benefit
- Actions taken: policies updated, training delivered, reviews completed
Before approval
Create a Use-Case Intake Record.
I use the intake record to connect a proposed tool to a real lending workflow. A product name alone is not enough: the same tool can carry different risks when drafting an internal note versus preparing a borrower communication. Capture these details before assigning a tier or starting a pilot.
- Purpose and Owner
- Name the business problem, the workflow affected, the accountable business owner, and the people who will use the output. Define how you will measure benefit against the current process.
- Data and Access
- Identify borrower information, loan documents, and other sensitive data the tool receives. Record where data goes, who can access it, how long it is retained, and whether the provider can use it for training.
- Output and Borrower Impact
- Describe what the tool produces and whether that output could influence credit, pricing, underwriting, servicing, or borrower treatment. Record which decisions remain with a human.
- Review and Escalation
- Name the reviewer, the checks required before an output is used, and the escalation path for an error or uncertain answer. Confirm that reviewers have time, access, and authority to intervene.
- Pilot and Evidence
- Define a limited pilot, representative test cases, acceptance criteria, and the records you will keep. Include incorrect, incomplete, and unexpected inputs, not just successful demonstrations.
- Approval and Exit
- Record the decision, restrictions, unresolved issues, next review, and who can pause the tool. Set out the manual fallback before the workflow depends on it.
The tiers in this guide are a starting framework, not a regulatory classification. I recommend reviewing the actual use, data, and borrower impact with your compliance and risk teams. An internal tool is not automatically low risk if it handles sensitive information or influences a lending decision.
Due diligence
Ask Vendors for Evidence, Not Just Assurances.
I connect vendor due diligence to the intended use case. Ask for written answers and supporting records your technology, security, compliance, and risk teams can evaluate together. A polished demonstration does not answer these questions.
- Where Does Our Data Go? Ask about hosting, subprocessors, retention, deletion, access controls, and whether prompts, documents, or outputs are used to train any model.
- What Has Been Tested? Request validation relevant to the proposed lending workflow, known limitations, error examples, and any applicable fair lending evaluation. Ask what the evidence does not cover.
- What Changes After We Buy? Establish how model versions, new AI features, and material changes are communicated. Ask whether features can be disabled and when your team can retest before use.
- Can We Reconstruct an Outcome? Confirm what inputs, outputs, versions, human reviews, and overrides can be recorded or exported, with appropriate protection for borrower information.
- What Happens When It Fails? Review incident notification, support escalation, rollback options, and data return or deletion at exit. Document gaps and resolve them before expanding the pilot.
Illustrative workflow
Work Through a Borrower Communication Example.
Consider an AI tool that drafts a request for missing loan documents. This is an illustrative example, not a client case study. I would start with a controlled, human-reviewed workflow rather than allowing the tool to send messages automatically.
Define the Boundary. The tool drafts from an approved checklist and authorized loan information. It does not determine eligibility, change pricing, or invent documentation requirements. The business owner and compliance team assess the actual risk before approving the pilot.
Test the Review. Use approved test data to check incorrect document requests, missing facts, confusing language, and unintended disclosure. A trained reviewer compares each draft with the loan record and approved requirements before anything reaches a borrower.
Measure and Decide. Compare review time, correction rates, and escalation patterns with the existing process. Keep the approval conditions and review evidence. If inaccurate requests or unexpected data use appear, pause the tool and return to the established manual workflow while the issue is investigated.
After launch
Keep Monitoring Connected to Action.
I recommend defining the monitoring plan before launch, with a named owner for each measure and a clear response when a threshold is crossed. Match review frequency to the risk and your institution's requirements rather than treating the sample tier cadences above as universal rules.
- Quality: Review errors, corrections, overrides, and exceptions against the pilot baseline. Check whether output quality changes across relevant document types and workflows.
- Borrower Experience: Track complaints, unclear communications, and escalations. Involve compliance and risk teams in evaluating borrower-treatment concerns.
- Change: Reassess when the vendor changes a model, the data changes, or the business expands the use case. Record the retesting and approval decision.
- Response: Specify who investigates, who can restrict or stop use, how affected work is reviewed, and what evidence is needed before restarting.
Bring the findings back to the board report: what changed, what the evidence shows, which issues remain open, and which decision leadership needs to make. This guide supports planning and discussion; it is not legal advice or a guarantee of regulatory compliance. Review institution-specific obligations with qualified counsel and your compliance team.
When leadership needs to make a specific approval or escalation decision, use the companion AI governance for lenders board decision brief to summarize the evidence, borrower impact, unresolved risks, and conditions for proceeding.
Order of operations
How to Put It in Place.
- 01
Take the Inventory
List every AI and automated tool in use, including AI features available in your LOS, CRM, and servicing platforms. Confirm which features are enabled, who uses them, and whether your current approvals cover them.
- 02
Assign Owners
Every use case gets a business owner who can answer for it: what it does, what data it touches, and who reviews its output. No owner means no deployment.
- 03
Tier the Risk
Sort the inventory into the three tiers. This is where governance gets practical: low-risk tools move fast, high-risk tools get the scrutiny regulators expect.
- 04
Write the Policy People Will Read
One to two pages: what is allowed, what needs approval, what is off-limits, and who decides. Attach the standards and procedures as living documents.
- 05
Stand Up the Approval Path
A short intake form, a cross-functional review for Tier 2 and 3, and a documented decision. Fast enough that people use it instead of routing around it.
- 06
Report, Review, Repeat
Put AI on the board agenda, review the inventory quarterly, and re-tier as tools and regulations change. Governance that does not update itself becomes shelfware.
Next step
Want Help Standing This Up?
I help lenders build this framework as a focused governance engagement: inventory, risk tiers, policy, vendor oversight, and board reporting, sized to your organization and your exam calendar. Bring your current state to a discovery call, or read how the consulting engagements are structured.
