# The Credit Pull Is Not One Event: Undisclosed-Debt Monitoring, the Pre-Closing Refresh, and the Permissible-Purpose Line an AI Agent Cannot Cross

*September 26, 2026 · 7 min read · Ramkumar Venkataraman*

> A mortgage credit file is pulled at application and again before the note signs, and the gap between those two pulls is where a borrower opens a car loan the underwriter never saw. An AI agent that watches that gap touches FCRA permissible purpose, the Fannie Mae age-of-documents rule, and the re-underwrite decision. Where the agent monitors, what it is allowed to pull, and why it never gets to re-decide the loan on its own.

## Two Pulls and the Gap Between Them

A mortgage credit report is not pulled once. It is pulled at application to qualify the borrower, and it is checked again right before closing to confirm the borrower who qualified is still the borrower signing the note. Between those two moments, which can be forty or sixty days apart on a purchase, the borrower lives a life. They finance a car for the new driveway, they open a store card for the appliances, they co-sign for a relative. None of that was in the file the underwriter approved, and all of it changes the debt-to-income ratio the approval was built on.

That gap is a real underwriting risk, and it is also where investors lose money, so Fannie Mae and Freddie Mac both require lenders to have a process for catching debt the borrower took on after application. The common answer is undisclosed-debt monitoring: a service that watches the borrower's credit activity during the loan's open window and alerts the lender when new inquiries or tradelines appear. It is a natural thing to hand an AI agent, because it is high-volume, pattern-driven, and mostly produces nothing until it suddenly matters. It is also a place where an agent that is built carelessly walks straight into the Fair Credit Reporting Act.

We build the agent that runs across mortgage origination, and the monitoring window is one of the sharper examples of a surface where the compliance constraint and the system design are the same conversation. The agent earns its place by watching the gap. What it is allowed to do inside that gap is narrower than most teams assume.

## Permissible Purpose Does Not Renew Itself

Every pull of a consumer report needs a permissible purpose under the FCRA. For a mortgage, the purpose is the credit transaction the consumer initiated, set out in [15 U.S.C. 1681b(a)(3)(A)](https://www.law.cornell.edu/uscode/text/15/1681b), which allows a report to be furnished to someone who intends to use it in connection with a credit transaction involving the consumer. That purpose attaches to the application the borrower made. It does not float free, and it does not automatically cover every later access the lender might find convenient.

Here is the failure mode we engineer against. An agent monitoring the borrower's credit during the loan window is accessing a consumer report, or a product derived from one, on a cadence. If the agent's access is not tied to the live, still-open credit transaction the borrower initiated, it is accessing credit data without a clean permissible purpose, and the FCRA does not have a "we were being careful" exception. The statute's whole architecture is that a consumer report moves only for an enumerated reason, and monitoring that outlives the transaction it was authorized under is a pull without a purpose.

So the control we built is a permissible-purpose token bound to the application state, not to the borrower. The monitoring subscription opens when the loan file is active and the borrower's credit authorization is on record, and it closes the instant the loan funds, is denied, or is withdrawn. The agent cannot start monitoring a borrower it is not actively originating for, and it cannot keep monitoring one whose transaction has ended. In the deployments we have run, that token is checked on every monitoring event, not once at setup, because the thing that goes wrong is not the first pull, it is the ninetieth pull on a file that closed three weeks ago and nobody turned the subscription off.

## Soft Pull, Not Hard, and the Reason It Is Not Cosmetic

There is a design decision at the monitoring step that has consequences for the borrower. Monitoring can run on a soft inquiry, which does not affect the borrower's score and is not visible to other lenders, or it can be built on repeated hard inquiries, which do both. We run monitoring as a soft pull, and the reason is not politeness.

A borrower in the middle of a mortgage is rate-sensitive and score-sensitive, and a string of hard inquiries from their own lender during the loan window can lower the score the loan was priced on, which is a self-inflicted wound that can push a borrower out of a pricing tier or a program. It can also look, to the next lender who reads the file, like the borrower is shopping for credit they did not actually seek. A monitoring design that degrades the borrower's credit standing to protect the investor has the risk balance backwards. The soft-pull architecture watches the activity without touching the score, and it keeps the agent's monitoring invisible to the rest of the borrower's credit life, which is where it belongs.

## What the Agent Does With an Alert, and What It Does Not

An undisclosed-debt alert is a signal, not a verdict. A new auto-loan tradeline during the loan window might be a genuine new obligation that changes the DTI, or it might be a reported balance on a loan the borrower already disclosed, or a credit-monitoring artifact, or an identity issue. The agent's job is to turn the raw alert into something an underwriter can act on, and to stop well short of acting on it.

The agent reconciles the alert against the file. It checks whether the new tradeline is already represented in the disclosed liabilities, whether the payment it implies would move the DTI past the threshold the approval depends on, and whether the activity is consistent with the borrower's stated situation. When the alert is noise, the agent documents why and closes it, so the underwriter is not buried in false positives, and so there is a record that the signal was seen and dispositioned. When the alert is real and material, the agent packages it: the new tradeline, the recomputed [debt-to-income](/blog/ai-underwriting-aus-findings-du-lpa-conditions-sr-11-7-model-governance) impact, and the specific question the underwriter now has to answer, which is whether the loan still qualifies as structured.

The agent does not re-decide the loan. A new material debt found during the window can require a re-underwrite, a changed-circumstance redisclosure, or in some cases a path toward adverse action on a loan that was conditionally approved, and every one of those is a decision the lender is accountable for. An AI agent that read a new car loan and quietly flipped the file to denied would be making a credit decision without the human accountability the law assigns to the creditor, and it would be doing it on data the borrower has a right to dispute. The agent surfaces and recomputes. The underwriter decides, and when the decision is adverse, the borrower gets the notice and the reasons the [ECOA and FCRA adverse-action rules](/blog/adverse-action-notices-ai-credit-decisions-ecoa-reg-b) require.

## The Fresh-Credit Refresh and the Age-of-Documents Clock

The second pull, the one right before closing, exists because credit documents go stale. Fannie Mae's Selling Guide sets a maximum age for credit documents, measured from the document date to the note date, and the lender has to be inside that window at closing or re-verify. The policy lives in the Selling Guide's allowable-age-of-credit-documents topic, reachable from the [Fannie Mae Selling Guide](https://selling-guide.fanniemae.com/) and operationalized in [Desktop Underwriter](https://selling-guide.fanniemae.com/sel/b3-2/desktop-underwriter-du). A loan whose credit documents have aged past the limit cannot simply close on the old pull. It needs a refresh, and the refresh can surface everything the monitoring was watching for, all at once, on the day everyone wants to sign.

The agent manages the age-of-documents clock as a system fact, the same way a point-of-sale agent manages the [TRID disclosure clock](/blog/trid-loan-estimate-closing-disclosure-ai-agents-timing). It knows the document date and the projected note date, it knows the maximum age, and it flags the files that will age out before they close early enough that the refresh is a scheduled step, not a closing-table fire drill. When the refresh runs and the credit is clean, the agent clears the condition with the new report referenced. When the refresh surfaces new debt, the agent runs the same reconciliation it runs on a monitoring alert, and the file goes back to the underwriter with the recomputed picture.

Running the clock as a system fact is also what keeps the refresh itself inside permissible purpose. The refresh is a pull on the live transaction, authorized, documented, and tied to the closing that is about to happen. It is not a speculative check. The agent will not refresh a file that is not heading to closing, because there is no transaction reason to, and a pull without a transaction reason is the thing we do not do.

## Why This Is a Model, and How We Govern It

The moment the agent reconciles an alert and recomputes a DTI that an underwriter relies on, it is doing the kind of input-to-estimate work that [SR 11-7](https://www.federalreserve.gov/supervisionreg/srletters/sr1107.htm) treats as a model, and we govern it as one. It has a documented intended use, it is validated against files that human underwriters dispositioned, and we measure the thing that actually matters, which is whether the agent's alert reconciliations agree with a careful re-review on the hard cases and route the genuinely ambiguous ones to a person rather than closing them.

The failure we most want to avoid is a confident false negative: an agent that closes a real undisclosed-debt alert as noise because the new tradeline looked like a duplicate of a disclosed one. An illustrative internal target we hold is that the agent's disposition of material alerts matches independent re-review at a high rate while it over-routes rather than over-closes, because on this surface the cheap mistake is sending a human an alert that turns out to be nothing, and the expensive mistake is funding a loan with a debt the file never accounted for. We tune for the cheap mistake on purpose. The agent's access to credit data also sits inside the lender's [GLBA Safeguards](/blog/glba-safeguards-rule-ai-vendors-banks-non-banks) program, scoped and logged, because a monitoring subscription is a standing flow of sensitive consumer data and it has to be governed like one.

## The Honest Read

The credit pull is two events with a risky gap between them, and watching that gap is real work an AI agent does well. The constraint is that monitoring is still a consumer-report access, so it lives or dies on permissible purpose, and permissible purpose is tied to the transaction the borrower initiated, not to the borrower and not to the lender's convenience. Bind the monitoring to the live file, run it as a soft pull so it does not degrade the borrower's own credit, close it the instant the transaction ends, and the agent is doing something both the investor and the borrower are better off for.

What the agent cannot do is turn an alert into a decision. A new debt found in the window is a signal that goes back to the underwriter with the math redone, and the credit decision, including an adverse one, stays with the human the law holds accountable for it. At Sei we build the monitoring and the pre-closing refresh as governed inputs on the live transaction, because the version of this that gets a lender in trouble is not the one that misses a car loan, it is the one that was pulling credit it had no current purpose to pull. Watch the gap. Keep the purpose clean. Leave the decision where it belongs.

---

_Source: [https://seiright.com/blog/fresh-credit-undisclosed-debt-monitoring-ai-mortgage-permissible-purpose-fcra](https://seiright.com/blog/fresh-credit-undisclosed-debt-monitoring-ai-mortgage-permissible-purpose-fcra) · Sei AI_