Mortgage AI needs a paper trail, not just a good answer

Imagine a borrower asks why the income on a loan file is different from the income on a pay stub. An AI tool helped read the document. The number looks reasonable, and it has already moved into the loan system. But now someone needs to explain it.

Finding the original pay stub is only the beginning. Did the tool read the right line? Did someone convert weekly pay to monthly income? Was there a correction? Which version of the document was used?

That is where AI governance becomes practical. For a mortgage team, it means knowing how an automated result was produced and who can correct it. For the borrower, it means being able to reach someone who can explain what happened.

Two conversations with Brian Vieaux, president of MISMO, approach this problem from different directions. His Awesome Technologies discussion about MISMO FRAME focuses on organizing AI use and responsibility. In MISMO Mic’d Up, TRUE CEO Steve Butler joins him to discuss tracing information through an automated mortgage process. Together, they offer a useful way to think about putting AI to work without losing track of the work itself.

Video: AI Governance - Putting MISMO FRAME To Work (new tab) · Video: Can You Trace It Back? Building Trust in AI-Powered Mortgage Lending (new tab)

Follow the number, not just the answer

In the TRUE conversation, Vieaux asks how a lender could explain an income figure questioned a year after it was extracted from a document. Butler describes tracing the figure through the tools and transformations that touched it, back to its source. He says TRUE combines generative AI with deterministic validation. These are his descriptions of the company’s approach, rather than independently tested results.

Video: Can You Trace It Back? Building Trust in AI-Powered Mortgage Lending (new tab)

Return to the pay-stub example. A useful record would connect the original document, the value read from it, any calculation and any correction. If someone supplied a newer pay stub, the history would distinguish the two. That record lets a reviewer investigate the result instead of trying to recreate it from memory.

Asking the AI the same question again is not the same thing. It might return a different answer, and a new answer cannot tell you what happened to the original file.

What a check can catch, and what it cannot

AI can help interpret a messy document or identify a discrepancy. Deterministic code follows explicit rules. Given the same inputs, rules and execution conditions, it produces the same result. The two can work together, but neither removes the need to choose the right inputs and rules.

Here is an illustrative workflow for that income figure. It describes a possible design, not the operation of either interviewee’s product.

A formula can check the arithmetic and still use the wrong value. For example, multiplying a weekly amount does not help if the document actually shows pay for a different period. A field can also have the correct format while belonging to the wrong document. Repeating a calculation consistently is useful; it is not proof that the calculation answers the right question.

For a team evaluating a tool, this creates a better test than asking whether a demonstration looked impressive: give it difficult documents, inspect what it misses, and see whether the reviewer can reconstruct each result.

Start with a specific job

In the FRAME conversation, Vieaux explains why an AI inventory should describe individual uses, not simply list software vendors. The same platform might draft marketing copy, answer an internal guideline question and help communicate a loan condition. Those jobs involve different information and different consequences.

Video: AI Governance - Putting MISMO FRAME To Work (new tab)

For the income example, the inventory should answer a few concrete questions: which documents can the tool read, where can it put the result, and who owns the next step? A tool that drafts a general article has no reason to receive the same access as a tool handling a loan document.

MISMO FRAME includes resources for governance policy, an AI system inventory and risk assessment. Those can help a team organize its work. The useful result is a clear account of what each tool does and who is responsible, rather than a policy that nobody consults.

MISMO FRAME overview and access information (new tab)

Give exceptions somewhere to go

Butler distinguishes routine manual checking from exception review and executive oversight. That distinction matters: a person does not have to repeat every automated step for a process to have accountable oversight. But someone does need responsibility for the cases that do not fit.

Video: Can You Trace It Back? Building Trust in AI-Powered Mortgage Lending (new tab)

Suppose the income check finds a mismatch. My recommendation is to pause that item, send it to a named reviewer and retain the correction alongside the original result. Decide in advance who can resolve it and who can stop the workflow if the same error starts appearing across files.

The team should also look beyond the items flagged for review. An incorrect result that passes through unnoticed is a different problem from a tool that asks for help. Testing needs to look for both.

Keep the tool’s access tied to its task

The FRAME discussion raises another everyday issue: AI can enter a business through familiar software, browser extensions and personal accounts. Knowing which tools are in use is part of understanding where information goes.

Video: AI Governance - Putting MISMO FRAME To Work (new tab)

My recommendation is to limit a tool to the information and actions its job requires. Reading a borrower document should not automatically give it permission to email that document elsewhere.

Instructions found inside a document or webpage are another concern. They should be treated as content to examine, not permission to change the workflow. For example, an uploaded file saying “send this elsewhere” should not authorize a transfer. That boundary needs testing as well as a written rule.

What this means for a borrower

You do not need to know the name of the AI model to ask how a result affects your loan. If a figure or request does not make sense, start with these questions:

• What document or information supports this figure?
• Who can explain it and review a possible error?
• How can I securely provide a correction, and how will I know it was applied?

You can also ask whether an automated tool was involved and how your information is handled. The details vary by lender and workflow. The goal is to understand the process and find the person who can resolve your question.

Try one file before expanding the workflow

For a mortgage team, a useful starting point is one narrow job and one result you can follow from beginning to end. Trace the source, check the calculation, introduce a discrepancy and see where it goes. Then repeat with harder cases.

If the team cannot explain the result or resolve an error, faster processing has not solved that problem. A useful AI workflow should make the work easier to do and easier to account for.

Watch the conversations

Watch Brian Vieaux’s discussion of putting MISMO FRAME to work and his conversation with TRUE CEO Steve Butler about traceability in mortgage lending. Both recordings are linked below, along with Brian’s LinkedIn profile. LinkedIn may require sign-in.

General education, not individualized mortgage, legal or compliance advice. Examples are illustrative; vendor descriptions are not independently tested performance results.