How mortgage and lending teams can use AI on loan files safely

What borrower data is at risk when loan teams use AI?

Nearly everything in a loan file is at risk, so AI in mortgage lending starts as a data security question. A typical file holds Social Security numbers, credit reports, tax returns, W-2s, pay stubs, bank and brokerage statements, and the loan application itself. Under the FTC Safeguards Rule, that is customer information, and the rule protects it whether the borrower closes or not.

The risk usually starts with a reasonable request. A processor wants a summary of three months of bank statements. A loan officer wants help explaining a debt-to-income problem to a borrower. An underwriter wants a quick read on a self-employed borrower's tax returns. Each task is a good fit for AI, and each one sends a borrower's financial life to a vendor the moment someone uploads the file.

This article covers protecting that data. Using AI to make or inform credit decisions raises separate fair lending questions under the Equal Credit Opportunity Act and Regulation B, which are outside its scope.

Does the FTC Safeguards Rule cover AI tools that touch loan files?

Yes, though the rule never mentions AI. 16 CFR 314.2 names mortgage brokers as financial institutions, and nonbank lenders fall under the same rule. Once borrower data enters an AI tool, that tool is one of the systems holding customer information, so the rule's existing requirements apply to it.

Several provisions of 16 CFR 314.4 do most of the work:

• The risk assessment in 314.4(b) must identify reasonably foreseeable risks to customer information, which now include staff pasting files into AI tools.
• 314.4(c)(1) requires access controls that limit each user to the customer information needed for their duties.
• 314.4(c)(3) requires encryption of customer information in transit over external networks and at rest.
• 314.4(c)(8) requires monitoring and logging of authorized users' activity, and detection of unauthorized use of customer information.
• 314.4(e) requires security awareness training kept current with the risks the firm identified.
• 314.4(f) requires the firm to select service providers capable of appropriate safeguards, require those safeguards by contract, and periodically assess the providers.

An AI vendor that receives borrower data to summarize or draft for the firm is a service provider. It belongs in the vendor inventory with the loan origination system and the document vendors.

Does sharing a loan file with an AI vendor require borrower consent?

Usually not under federal law, if the vendor works only for the lender and the contract says so. Regulation P generally requires an opt-out notice before sharing nonpublic personal information with nonaffiliated third parties, but 12 CFR 1016.13 exempts service providers when the firm gives its initial privacy notice and a contract prohibits the provider from disclosing or using the information other than to carry out the purposes for which it was disclosed.

That contract term is where many AI tools fail. A consumer AI account whose terms allow the vendor to keep prompts or use them to train its models is hard to square with a contract that limits use to the lender's purpose. FTC staff have also warned that AI companies quietly changing their terms to allow training on customer data could be unfair or deceptive, which is a reason to recheck terms at renewal.

California adds a stricter layer. The California Financial Information Privacy Act generally requires a consumer's written consent before a financial institution shares nonpublic personal information with a nonaffiliated third party. The DFPI describes an exception for third parties performing services such as data processing, with conditions: the firm could lawfully perform the service itself, a written contract limits the provider's use of the data, only the information needed for the service is shared, and the firm receives no payment for the release. An AI vendor that meets all four conditions fits the exception; one whose terms let it keep prompts for its own model training is hard to fit within it.

Which loan file tasks can go to AI without borrower data?

More than most teams expect. Many useful tasks need no borrower information at all:

• Summarizing investor or agency guidelines, overlays and product matrices.
• Drafting template language for conditions letters, borrower checklists and status emails, with names and numbers filled in later.
• Explaining a general concept, such as how rental income is calculated, using made-up numbers.
• Turning internal procedures into training material for new processors.

Tasks that need a real file, such as summarizing a borrower's statements or reconciling income documents, belong only in a tool the firm has approved, under a contract that limits use, with access restricted to the people working that loan.

What controls make AI in mortgage lending safe?

Policy alone will not hold. A rule that says "never paste borrower data into public AI" breaks down the first busy month, so the controls have to enforce it:

• One approved tool, signed in with firm accounts. Pick an enterprise AI tool whose contract limits use to your purposes and excludes training. Require single sign-on so access ends when employment does.
• Block the alternatives. Use web filtering or tenant restrictions to block personal AI accounts and unapproved AI sites on firm devices.
• Data loss prevention on prompts and uploads. Set DLP rules that detect Social Security numbers, account numbers and loan numbers, and block or warn before they reach an unapproved destination. Our guide to DLP rules for account numbers and SSNs in AI prompts covers the setup.
• Logging you actually review. Keep the approved tool's audit logs and review them on a schedule, which supports 314.4(c)(8).
• Retention settings that match your disposal policy. 314.4(c)(6) requires disposal of customer information no later than two years after it was last used for the customer, with limited exceptions. Chat histories and uploaded files in an AI tool count, so set retention to match.
• Vendor file for each AI tool. Record the contract terms on use, retention and training, and review them at least annually. Our article on vetting an AI vendor's retention and training terms lists what to ask.

What happens if borrower data goes into an unapproved AI tool?

Treat it as a security incident and work through your incident response plan. Under 314.2(m), a notification event is the acquisition of unencrypted customer information without the consumer's authorization, and unauthorized access is presumed to be acquisition unless the firm has reliable evidence otherwise. If an event involves at least 500 consumers, 314.4(j) requires notice to the FTC no later than 30 days after discovery. Our article on reporting a breach to the FTC walks through the form.

A single upload rarely reaches 500 borrowers, but a bulk export of a pipeline report can. Whether an upload is a notification event at all depends on the facts, such as the tool's terms and whether the data was kept. California's own breach notification law may also apply, depending on the data elements involved, so bring in counsel early.

What should a lending team do this quarter?

Start with what staff already use, then close the gaps:

• Pull sign-in and web traffic data to find the AI tools staff already use with firm email or on firm devices.
• Choose one approved AI tool, confirm its contract limits use to your purposes and excludes training, and add it to your service provider inventory.
• Update the risk assessment and written information security program to name AI tools as systems that may hold customer information.
• Turn on DLP for Social Security numbers, account numbers and loan numbers in browsers and the approved tool.
• Block personal AI accounts on firm devices.
• Add a short AI section to security awareness training, with examples of tasks that need no borrower data.
• Set the approved tool's retention to match your disposal schedule, and confirm audit logs are kept.

For the broader Safeguards Rule picture for brokers, see the FTC Safeguards Rule for mortgage brokers.

How does s90 help?

s90 sets up the controls behind an AI policy for mortgage brokers and lenders: an approved AI tool tied to firm sign-in, DLP that recognizes borrower data in prompts and uploads, blocking for personal AI accounts, and logs your Qualified Individual can review. We also document each AI vendor for your Safeguards Rule program. If you want to see where loan file data goes today, we are happy to start there.