Can Off-the-Shelf AI Review a Mortgage Loan Purchase Agreement, and What Is It Missing?

Off-the-shelf AI cannot competently review a mortgage loan purchase agreement (MLPA), because what it is missing is the documents the agreement incorporates and the market practice that gives it meaning. By off-the-shelf AI we mean any model, general-purpose or built for lawyers, that has not been trained on closed transactions like yours and supervised by the lawyers who negotiated them. Such a model learns from what its builders could collect, and no such model has read your deals. In our experience the provisions that decide a secondary-market trade’s legal economics are rarely analyzed in public, so, in our view, the model is unlikely to have seen them analyzed closely.

Why is this different from ordinary contract review?

Reviewing these agreements is different from ordinary contract review because, in our experience, the record a model can learn from holds little on how these provisions work. A search of the open internet for how a seller guide amendment right operates, or how an early payoff (EPO) window is measured, returns, in our experience, a familiar mix. It is mostly aggregators’ own guides, vendor marketing, and agency job aids. We find seller-side, provision-level analysis scarce, which is why we publish it.

In our view, that leaves an off-the-shelf model to learn the instruments at issue mostly from the counterparty’s paper and mortgage technology vendors’ marketing. The instruments are the MLPA, the mortgage servicing rights (MSR) purchase and sale agreement, the seller guide, and the warehouse facility.

What does an off-the-shelf model not know about a seller guide?

An off-the-shelf model does not know what the seller guide in your deal says, in the version in force, and a correspondent purchase agreement can hand that guide nearly every operative term. One older, public example: a correspondent purchase and sale agreement between a federal savings association, as purchaser, and a correspondent mortgage lender, filed with the Securities and Exchange Commission in 2003, runs five pages. Article I incorporates the seller guide “by reference in its entirety.”

Article V leaves the seller’s repurchase obligation “as more fully set forth in the Seller Guide,” and Articles VI and VII do the same with suspension, termination, and duration. Section 8.13 puts the guide above the agreement: “In case of any inconsistency between this Agreement and the Seller Guide, the terms and provisions of the Seller Guide shall control.” Only a commitment letter ranks higher than both.

Section 8.2 then splits the two documents. The agreement “may not be amended except by an instrument in writing signed on behalf of each of the parties hereto.” The seller guide “may be amended or supplemented by Purchaser from time to time in Purchaser’s sole and absolute discretion.” And a seller who “registers, locks or delivers a Mortgage Loan after receiving notice of a proposed amendment . . . shall be deemed to have agreed to such amendment.”

Five pages defer nearly every operative term to a document the model has not been given and rank that document above themselves. They also let the counterparty amend it unilaterally, by a writing that binds from a stated effective date. A model reading the five pages alone can see that the guide controls, but not what the guide says. What an aggregator can change without consent takes that right apart.

Which provisions does it read in isolation and misstate?

The provisions we would expect an off-the-shelf model to misstate when it reads them in isolation are the ones that work only together, starting with the early payoff window.

One large correspondent aggregator’s published delegated seller guide provides that an early payoff “exists when a Mortgage Loan is paid in full or when a curtailment in excess of 30% of the original principal balance occurs within 180 days after the Funding Date of such Mortgage Loan.” The same guide separately reaches loans eligible for sale to Fannie Mae or Freddie Mac and “paid in full within 120 days after the Funding Date of such Mortgage Loan,” cross-referencing both agencies’ guides.

The Fannie Mae provision, Selling Guide C1-1-01 (February 5, 2025), measures from somewhere else: “any loan that pays off within 120 days from the whole loan purchase date or the MBS issue date, as applicable.” The day counts match, but on their faces the starting events differ: the aggregator’s “Funding Date” on one side, Fannie Mae’s purchase or issue date on the other.

Neither the aggregator’s guide nor Fannie Mae’s provision, as read here, announces the divergence. EPO premium recapture in correspondent lending compares the published windows.

Then the knowledge qualifier. Under Fannie Mae Selling Guide A2-3.2-01 (August 29, 2017), Fannie Mae “may also require repurchase or a make whole payment if any warranty the selling lender made is untrue and, if the remedies framework applies, qualifies as a significant defect, whether or not the lender had actual knowledge of the untruth.” The next sentence supplies the exception: no such request “will be made if the warranty specifically states that a violation does not exist unless the lender had actual knowledge of the untruth and the lender has no such knowledge.” Quoting either sentence alone misstates the exposure.

In our view, the provisions that allocate a seller’s exposure operate on one another; summarizing any one in isolation risks a confident wrong answer.

What has already gone wrong when lawyers relied on it?

What has already gone wrong when lawyers relied on an off-the-shelf model is a sanctions order; the one we discuss here involved a general-purpose chatbot. In Mata v. Avianca, Inc., No. 22-cv-1461 (PKC), plaintiff’s counsel had filed an affirmation in opposition citing decisions that did not exist.

The U.S. District Court for the Southern District of New York ruled on June 22, 2023. It found that the two lawyers involved and their firm “abandoned their responsibilities when they submitted non-existent judicial opinions with fake quotes and citations created by the artificial intelligence tool ChatGPT, then continued to stand by the fake opinions after judicial orders called their existence into question.” It imposed a $5,000 penalty, jointly and severally.

The detail that carries outside litigation is in the findings of fact. The lawyer who wrote the filing asked the model whether the cases were real (“Is Varghese a real case”). “ChatGPT responded that it had supplied ‘real’ authorities that could be found through Westlaw, LexisNexis and the Federal Reporter.” Asked to check itself, the model confirmed itself.

The court was careful about the general point: “Technological advances are commonplace and there is nothing inherently improper about using a reliable artificial intelligence tool for assistance. But existing rules impose a gatekeeping role on attorneys to ensure the accuracy of their filings.”

A court filing has an adversary and a judge reading it. A misread survival period has neither, and the error surfaces years later, in a demand.

Where is AI genuinely useful here?

AI is genuinely useful here for administrative acceleration. For substantive drafting, and above all for the terms that define a trade’s legal economics, the AI we would use is not off-the-shelf. It is one trained on closed transactions like yours and supervised by the lawyers who negotiated them. That is the premise behind DealGPS.

Within substantive drafting, the risk of off-the-shelf AI is sharpest where the output defines a trade’s legal economics. In our view, that is true in the purchase agreement and the MSR purchase and sale agreement alike.

The facts the answer turns on

In our view, the thing to worry about is the document the model was never given. That is the seller guide the agreement incorporates, in the version in force on the day that matters. Where the agreement ranks the guide above itself, as the 2003 agreement does, a model reading the agreement alone can see that the guide controls, but not what it says. And the provisions that decide the trade’s legal economics work only together.

The answer turns on two facts: which guide version was in force, and what the agreement itself says about the guide and about the remedies as a whole. Those are questions of the documents, not of the model, and that is the work.