The Classification That Happens Before the Ratio
A debt-service-coverage loan is marketed as the simple one. No tax returns, no paystubs, no employment to verify. You take the property's rent, divide by the payment, and if the ratio clears the program's floor the loan qualifies. Teams building AI for non-QM point at exactly this and say it is the easiest workflow to automate, because the math is a division problem.
The math is the easy part, and it is not where these loans get lenders in trouble. The thing that governs whether a DSCR loan is a well-priced product or a compliance exposure is a classification the file has to make before anyone computes a ratio: is this extension of credit primarily for a business purpose. If it is, the loan sits outside Regulation Z's ability-to-repay rule, which is the whole premise of pricing it without income documentation. If it is not, the lender has just originated a consumer-purpose mortgage with no ability-to-repay determination behind it, and no amount of coverage ratio cures that.
An AI agent that computes DSCR and skips that determination is not automating underwriting. It is automating the production of a classification error at volume, which is worse than doing it by hand because now it is consistent.
What the Exemption Actually Says
The reason a DSCR loan can skip income documentation is 12 CFR 1026.3(a), which exempts from Regulation Z any extension of credit primarily for a business, commercial, or agricultural purpose. The ability-to-repay rule at 12 CFR 1026.43 applies, by its own terms, to a consumer credit transaction secured by a dwelling. A loan that is primarily for a business purpose is not a consumer credit transaction, so 1026.43 does not reach it. That is the chain the entire product rests on, and it is three links long.
The official commentary to 1026.3(a) does not give you a rent-coverage shortcut. It says the creditor must determine in each case whether the transaction is primarily for an exempt purpose. There is no safe harbor that says "investment property equals business purpose." A borrower buying a single rental with the intent to live in half of it, or a borrower whose real plan is to move in after a season, can produce a loan that looks like an investor deal on the application and is a consumer-purpose loan in fact. The commentary also warns that the primary-purpose test is decided on the facts, not on which disclosure box got checked, so a lender that hands out Regulation Z disclosures on a business loan has not thereby made it a consumer loan, and a lender that skips them has not made a consumer loan into a business one.
So the determination is a facts-and-circumstances judgment the creditor owns, and the file has to show the work. That is the part a coverage-ratio calculator cannot do and the part an examiner will ask about first.
How We Build the Agent to Make the Determination Defensible
We do not let the agent conclude "business purpose" from the loan program. The loan program is the lender's label, and the label is the thing under examination, so using it as the input is circular. Instead the agent assembles the business-purpose record from the facts in the file and scores how well those facts support the exemption, against the lender's own encoded policy for what a business-purpose investor loan has to show.
In practice that record is built from a short, specific set of signals, each tied to a document in the Loan Graph: the borrower's stated purpose on the application, occupancy declared as investment, an existing rental or business entity in the borrower's profile, a lease or market-rent analysis on the subject, the borrower's other financed properties, and whether title is being taken in an entity. Any one of these in isolation proves little. Taken together they either tell a consistent business-purpose story or they contradict each other, and the contradictions are the finding.
The agent's output on this step is not a yes. It is a business-purpose memo: the facts it relied on, the lender's policy rule each fact satisfies, and a confidence score. Anything above the lender's threshold clears to the ratio work. Anything below it routes to a human underwriter as a named exception, with the conflicting facts surfaced rather than buried, because an under-threshold business-purpose call is exactly the kind of judgment that belongs to a licensed person and not a model. This is the same triple-check path we run on every decision: the agent reasons against the encoded rule set, an independent pass re-checks it, and the result carries a score and a citation.
Then, and Only Then, the Coverage Ratio
Once the business-purpose record stands, the ratio work is real underwriting, not arithmetic, because the inputs are contestable. The agent computes DSCR as qualifying rent divided by the monthly housing expense, and every term in that fraction is a place where programs differ and where a defect hides.
Qualifying rent is the first. For a property with a signed lease, the lender's program may use the in-place rent, or the lesser of the lease and the market rent on a Form 1007 or 1025, and the vacancy treatment differs from the agency 75% rule because these loans are not agency loans. The agent applies the lender's program rule, not its memory of how Fannie treats rental income, which is a different product governed by a different guide. The denominator is the second: principal, interest, taxes, insurance, association dues, and for some programs a reserve or a stressed rate rather than the note rate. A ratio computed on the note rate when the program requires a qualifying rate above it is not a rounding error, it is the wrong number carried into the credit decision.
The agent shows its arithmetic the way an auditor reads it: the rent figure and the document it came from, the payment components each cited to their source, the program rule for vacancy and qualifying rate, and the resulting ratio against the program floor. A reviewer checks the citations rather than taking the number on trust, and the record that proves the ratio is the same record that answers the investor's later question about how this loan was qualified.
The Failure Mode We Engineer Against
The first version of a system like this will happily pass a loan where the business-purpose facts are thin but the coverage ratio is strong, because a 1.35 DSCR feels like a safe loan and the model learns that strong ratios correlate with approval. That is the failure we design against, because it inverts the actual risk. A strong ratio on a loan that is secretly consumer-purpose is not a safe loan. It is a Regulation Z violation with good collateral, and the collateral does not cure the violation.
So the business-purpose gate is hard. The agent cannot net a strong ratio against weak purpose facts, because those are answers to two different questions and only one of them decides whether the ATR rule applies. We keep them as separate determinations with separate confidence scores precisely so that the system cannot launder a classification problem through a good ratio, which is the quiet way these programs accumulate risk that surfaces in a repurchase demand or a state exam two years later.
Why This Is a Model Decision Under Governance
The business-purpose classifier and the ratio engine are both models, and we treat them as models under SR 11-7. That means the classifier's inputs, its policy rules, and its error rates are documented, and the threshold that decides what clears automatically is a governed control the lender's model risk owner sets, not a setting an engineer tunes quietly. When the lender changes its definition of an acceptable business-purpose investor loan, that is a change to an encoded rule with a version and an owner, and the loans underwritten before and after the change are distinguishable in the record.
The fair-lending edge matters here too. DSCR programs are priced and qualified with wide latitude, and latitude applied by a hundred underwriters is where disparate treatment lives. A consistent, documented determination run the same way on every file is testable in a way a pile of individual judgments is not, which is the argument we make to a lender's compliance team: the automation does not weaken the fair-lending posture, it makes it auditable, as long as the criteria are the lender's and written down.
Where the Underwriter Still Signs
The agent does the assembly and the arithmetic. It builds the business-purpose record, computes the ratio against the program, reconciles the rent and the payment to their sources, and clears the files where everything agrees above the lender's threshold. What it does not do is make the business-purpose call on a contested file or sign the credit decision, because the determination that keeps a DSCR loan outside ability-to-repay is a legal conclusion the creditor is accountable for, and that accountability does not transfer to a model no matter how good the ratio looks.
We built Sei for exactly this division, because the easy framing of DSCR as a division problem is the framing that produces the expensive mistake. The ratio is the part that looks like underwriting. The business-purpose determination is the part that is underwriting, and it is the part worth automating carefully, so the file explains itself the day an investor or an examiner asks why this loan never got an ability-to-repay determination and was right not to.
Ramkumar Venkataraman
CTO & Co-Founder