Industry

The AI in Your Close

November 13, 2026

Fannie Mae's AI governance requirements took effect August 6. The letter names origination and servicing. Your loan data doesn't stop there. It's the second-to-last night before the closing. Something in the servicing subledger won't tie, and the person chasing it has been at it since dinner. They open the AI reporting tool they use every day, types the question in plain English, and gets a clean answer back — the variance is isolated, the entries behind it listed underneath.

Nothing went wrong here. Good tool, used well, by someone who knows her job. The close lands on time.

But the company sells loans to Fannie Mae. And as of August 6, if Fannie asks what artificial intelligence the company uses, what for, and what safeguards sit around it, the company must answer promptly. So: plain English in, a correct answer about loan-level data out — is that on the list?

The advice didn't change. The consequences did.

FHLMC came out with Guide Bulletin 2025-16 in December 2025, which the impacted Freddie Mac Guide Section 1302.2 and 1302.8. Essentially, it’s a comprehensive framework for deploying AI systems, including senior-level approval of AI policies, ongoing monitoring, training, and both internal and external audits. Fannie Mae followed with Lender Letter LL-2026-04 in April 2026, effective in August,- requiring documented policies reviewed annually, disclosure on request, and, governance of vendor and subcontractor AI that is "no less protective of these requirements.”  

None of them displaced what was already there. TILA, RESPA, HOEPA, and fair lending obligations never carved out an exception for software that's difficult to explain.

The inventory follows the loan

Both letters use the same framing. They apply to AI used in connection with the origination or servicing of loans. And neither one defines artificial intelligence or machine learning.  

That combination is easy to misread. "Origination and servicing" sounds like a map, so inventory work has gone where the map points: the LOS, the underwriting models, the borrower-facing chatbot, the servicing platform. All of that belongs on the list.

Loan data doesn't stop at the servicing platform; it keeps moving into the general ledger, the bank reconciliation, investor reporting, escrow analysis, and the loan-level P&L your CFO takes to the board. Those systems hold the same borrower information the letters care about. They just sit under a department that gets asked about SOC reports and segregation of duties, not about models.

Whether every back-office tool lands inside the GSEs' scope is a question for your counsel, and the language leaves room to argue. At the least, though, creating a policy that includes a full inventory of the systems, and defining what is in and out of scope, is a good place to start.  

Procurement catches vendors. It doesn't catch features.

Here's why the back office is the hardest part of the inventory, and it has nothing to do with anyone's diligence.

Your vendor management program is built to catch new vendors. Someone requests a tool, it gets reviewed, a questionnaire goes out, a contract gets signed, the vendor lands on a list. That process works.

It just doesn't fire when an existing vendor ships a feature.

The capabilities now appearing across the mortgage accounting stack arrive that way — document capture that reads and codes invoices, reconciliation that suggests matches, reporting layers that answer questions in plain English, assistants built into the platform. They don't come through procurement. They come through a release note. No purchase order, no new contract, no questionnaire, and no reason for anyone to think a review was owed.

Which is how a lender ends up with a vendor list that is genuinely complete and an AI list that is genuinely empty, both at the same time.

Walk the close calendar, not the org chart

The standard advice to inventory your AI is correct. But how do you do it?  

Take the month-end close checklist and the list of systems that touch loan-level or borrower data. For each one, five questions — the same inputs-ownership-usage frame, made operational:

  1. Has this vendor shipped an AI or "assistant" capability in the last 18 months? Release notes and product blogs, not the sales rep.
  2. Is it on by default, and who can turn it off?
  3. What does it actually see? The vendor's own documentation? Your configuration? Your loan tape? Those are very different answers.
  4. Where does that data go, and does anything leave your tenant?
  5. Does what you put in get used to train or improve the vendor's model, and who owns what comes out?

You'll finish faster than you expect, and you'll have produced the thing the GSE requirements letters actually ask for: a written record with dates, owners, and answers.

You'll also learn which vendors are worth trusting here. One who can tell you precisely what its AI was trained on and what it never touches has handed you something you can put in a file. One whose answer is that they take AI security seriously has handed you a sentence. Both go in the record, but only one is a control.

Then tier it, and be willing to accept a risk on paper

Not everything you find deserves the same scrutiny, and pretending otherwise is how governance projects die. Sort by two things: how much borrower data it sees, and how much decision-making authority it has. A tool that drafts a memo is not a tool that codes a transaction, and neither is a tool that shapes what a borrower is offered. Where the authority is real, keep a person in the loop. That judgment is the control.

One back-office wrinkle is worth naming. A model that returns a confidently wrong answer about a borrower tends to get caught because fair lending review exists and looks for exactly that. A model that returns a confidently wrong answer about a variance, at nine o'clock on the second-to-last night of the close, has nobody watching for it. Speed is the point of these tools. Speed is also what makes a plausible wrong answer expensive.

Fix what's at the top. For what's further down, a documented decision to accept a risk, with a named owner and a review date, is a legitimate answer, and a far better one than an undocumented gap. Perfect isn't the standard here. Written down is.

The GSEs are not asking you to be finished. They're asking you to show your work.

The part worth being optimistic about

None of this argues for slowing down. The AI arriving in mortgage accounting is useful; it puts answers in front of people who used to open a ticket and wait. The lenders who get ahead won't be the ones who locked it down. They'll be the ones who can adopt the next capability in a week, because they already know where their AI lives, what it touches, and who owns it.

That inventory is a Tuesday afternoon. The alternative is finding out what's in your stack during an investor review, in front of someone taking notes.

Two years ago, the honest answer to "should we be governing this?" was probably, eventually. It's a shorter answer now.

Getting a true handle on this requires expertise in both mortgage banking and cybersecurity. If you want help mapping where AI already lives in your environment, or building the record the GSEs now expect to see, contact our experts at info@richeymay.com.

‍

Richey May

About the Author

Richey May is a national accounting, advisory, and technology firm serving industries where compliance, strategy, and innovation intersect. Founded in 1985, the firm delivers audit, assurance, tax, business advisory, cybersecurity, and technology services with deep specialization in mortgage banking. Richey May is known for “digging deeper,” connecting strategy, technology, and risk management to help clients navigate complexity and capture opportunity. With decades of expertise and a commitment to anticipating industry shifts, the firm empowers organizations to transform financial and operational challenges into drivers of growth. Richey May was recently named a Top 50 Firm by Accounting Today.

No items found.

Ready to Grow and Scale Your Operations?

Gain unparalleled efficiency, compliance, and insights with Loan Vision. Begin your journey with us today.