# Pre-Funding QC With an AI Agent: The Fannie D1-2 Review, the Defects You Catch Before the Wire, and the Reverifications the Rule Still Wants a Human to Judge

*August 5, 2026 · 7 min read · Ramkumar Venkataraman*

> Post-closing QC tells you how many defective loans you already sold. Pre-funding QC is the only review that changes the outcome, because it happens before the money moves. An AI agent can run the full-file pre-funding review at 100 percent of the pipeline instead of a sample, catch the income-calculation error and the data-integrity break before closing, and drive the reverifications the rule requires. What the agent computes, what stays human, and how the defect taxonomy has to be built so the review is defensible.

## The Only QC Review That Changes the Outcome

Post-closing quality control tells a lender how many defective loans it has already sold to the secondary market. It is a required, useful, backward-looking measurement, and it does nothing for the loan under review, because the loan has closed and funded and, often, been delivered. The [Fannie Mae Selling Guide at D1-3-01](https://selling-guide.fanniemae.com/sel/d1-3-01/lender-post-closing-quality-control-review-process) requires a post-closing random sample, either a minimum 10 percent of the prior period's acquisitions or a statistical sample sized for a 95 percent confidence level with a 2 percent precision, with the review cycle and defect-rate reporting running on a defined monthly cadence. That process produces a defect rate. It does not prevent a defect.

Pre-funding QC is the review that can. The Fannie Selling Guide at [D1-2-01](https://selling-guide.fanniemae.com/sel/d1-2-01/lender-prefunding-quality-control-review-process) requires a pre-funding review process precisely because catching a defect before the loan closes lets the lender fix it while fixing is still possible, reverify the income, correct the data, cure the disclosure, rather than repurchase the loan or eat the loss after delivery. The problem lenders have with pre-funding QC is coverage. A human pre-funding team reviews a sample, because reviewing every file by hand before it funds does not scale against a pipeline. So the review that could prevent defects sees a fraction of the loans, and the defect that funds is usually the one the sample missed.

An AI agent changes the coverage math. It can run the pre-funding review on 100 percent of the pipeline instead of a sample, which turns pre-funding QC from a spot check into a gate. We build that agent. What follows is what it actually reviews, where its judgment stops, and how the defect taxonomy has to be structured so the review holds up to an examiner and to Fannie's own reviews.

## What the Agent Reviews on Every File

The pre-funding review has a determinate structure, because the defects it looks for are defects against known standards: the loan's data, the loan's documents, and the consistency between them. That is exactly the kind of review an agent runs well, at volume, without fatigue.

The agent recomputes the numbers the file asserts. It recalculates qualifying income from the source documents rather than trusting the income figure keyed into the loan origination system, because a transposed number or a misapplied [self-employed cash-flow analysis](/blog/ai-income-calculation-self-employed-1084-4506c-reasonableness) is one of the most common and most consequential pre-funding defects. It recomputes the debt-to-income ratio from the recalculated income and the liabilities on the credit report and the application. It checks the loan-to-value against the appraised value and the sales price. It verifies that the assets documented support the funds the loan structure requires.

The agent checks data integrity across systems. The figures on the [HMDA loan application register](/blog/hmda-data-integrity-ai-lar-field-population), the closing disclosure, the note, and the LOS record all describe the same loan, and they have to agree. A rate on the note that differs from the rate on the CD, an occupancy on the 1003 that differs from the occupancy the appraisal describes, an APR that does not reconcile to the finance charge, each is a defect the agent surfaces before the file funds rather than after an examiner finds it.

The agent verifies document completeness and validity against the loan's requirements. The documents the profile required are present, the tax returns are complete rather than partial, the transcripts match the returns, the appraisal is the current form, the flood determination is in the file, and the dates are inside the windows the loan program allows. This is the same document discipline the [point-of-sale intake agent](/blog/ai-mortgage-point-of-sale-urla-intake-trid-application-condition-prevention) runs at the front door, applied again as a gate before the money moves.

## The Reverifications the Rule Requires

Fannie's pre-funding requirements at D1-2-01 call for reverification of specific loan components on the pre-funding sample, including verbal or electronic reverification of employment, reverification of assets and Social Security numbers, and validation of the appraisal or other property valuation. Reverification is not a recalculation of what the file already contains. It is going back to the source independently to confirm the file is true, and it is where the agent's role and the human's role separate.

The agent drives the mechanical part of reverification. It initiates the verbal verification of employment through the employer or the authorized third-party payroll source, submits the [IRS transcript request through the 4506-C process](https://www.irs.gov/pub/irs-pdf/f4506c.pdf) to confirm the income documents against what was filed, and runs the SSN validation against the issuing authority. It collects the results, compares them to the file, and flags every discrepancy, an employer that reports a different title or a termination the file does not show, a transcript that does not match the return in the file, an SSN that does not validate.

What the agent does not do is conclude what a discrepancy means. A verbal VOE that comes back with a slightly different job title can be a data-entry difference or a material misrepresentation, and the difference is a judgment about the borrower and the file that a QC analyst makes, not a rule the agent applies. The agent's contribution is that the discrepancy is found on every file rather than on a sampled few, is documented with the source and the timestamp, and is put in front of the analyst with the file context attached. The judgment stays human because Fannie's requirement and the fraud exposure both live in the interpretation, not in the detection.

## The Defect Taxonomy Has to Be Built Before the Agent Runs

A pre-funding review is only as defensible as its defect taxonomy, the classification that says what counts as a defect and how severe it is. Fannie's framework, described across the D1-1 and D1-3 sections, distinguishes findings by severity and requires a target defect rate for the highest severity level on the post-closing random sample. If the pre-funding agent flags findings against a taxonomy that does not map to the lender's post-closing severity definitions and to Fannie's expectations, the pre-funding results and the post-closing results speak different languages, and the QC program cannot show the pre-funding gate is actually reducing the post-closing defect rate.

We build the agent's classification to the lender's existing defect taxonomy, not to a taxonomy of our own. A finding the agent raises carries a defect category (income, assets, credit, collateral, compliance, data integrity), a severity that maps to the lender's and Fannie's severity levels, the specific document or data element the finding is about, and the recomputation or reverification that produced it. That structure lets the QC team do the two things the program depends on: route the file to the right cure before funding, and roll the pre-funding findings up against the post-closing defect rate to prove the gate works. A pre-funding process whose findings cannot be tied to the post-closing defect rate is a process that cannot demonstrate its own value, and the demonstration is what keeps the investment funded.

## The Model-Risk Question the Agent Raises

An agent that recomputes income and reclassifies defects is making determinations that feed the loan-sale decision, which puts it inside the [model-risk-management perimeter](/blog/model-risk-management-ai-agents-sr-11-7-nist-rmf) even though pre-funding QC is not usually thought of as a model. The recalculation logic has to be documented and validated, the classification has to be tested against a human-reviewed benchmark, and the agent's error modes, the defect it misses and the false defect it raises, have to be measured and monitored. A pre-funding agent that silently drifts, that starts passing files a human would have flagged, is worse than no agent, because the gate reports green while the pipeline degrades.

The control we run is a continuous benchmark. A sample of the agent's pre-funding decisions is re-reviewed by the QC team, the agreement rate is tracked, and disagreements feed back into the recalculation and classification logic. The agent's job is to review every file; the human's job is to keep reviewing the reviewer, because a QC gate that no one audits is a single point of failure sitting in front of the wire.

## The Failure Mode We Engineer Against

The pattern that produces the worst pre-funding outcomes is the program that runs pre-funding QC as a small human sample and post-closing QC as the real measurement. The sample is too small to be a gate, so defects fund. The post-closing review finds them after delivery, when the only remedies are repurchase, indemnification, or a loss. The defect rate is a number the QC team reports to senior management and to Fannie, and it stays stubbornly above target because the review that could have moved it, pre-funding, never had the coverage to matter.

The architecture we run makes pre-funding the gate and post-closing the audit of the gate. The agent reviews every file before it funds, recomputes the numbers, checks the data integrity, drives the reverifications, and routes findings to the QC team against the lender's own severity taxonomy. The human analysts spend their time on the judgment the agent surfaced, the discrepancy that might be misrepresentation, the finding whose severity is a call, rather than on the mechanical recomputation the agent already did. Post-closing QC then measures whether the gate is holding, and because the pre-funding and post-closing taxonomies are the same, the program can show the defect rate falling as the gate's coverage rises.

The economics follow the coverage. The [Freddie Mac 2024 Cost to Originate study](https://sf.freddiemac.com/docs/pdf/cost-to-originate-full-study-2024.pdf) traces how much of the origination cost sits in rework and quality remediation, and a defect caught before funding is cured at a fraction of the cost of a defect caught after delivery. Pre-funding QC at full coverage is the review that converts that remediation cost into prevention, and an agent is what makes full coverage possible.

## The Honest Read

Pre-funding QC is the only quality-control review that can change a loan's outcome, and lenders under-invest in it because a human team cannot review the whole pipeline before it funds. That constraint is the one an AI agent removes. The agent runs the full-file pre-funding review on every loan, recomputes the numbers, checks the data integrity across systems, and drives the reverifications the rule requires, which turns pre-funding from a sample into a gate.

The judgment does not move. Whether a reverification discrepancy is a data difference or a misrepresentation, whether a borderline finding clears or holds the loan, whether a defect is severe enough to stop the wire, these are the QC analyst's calls, made on the findings the agent produced. At Sei, we build the pre-funding agent to expand coverage to the whole pipeline and to feed the human the judgment calls with full context, and we keep a human auditing the agent's decisions so the gate itself stays honest. Post-closing QC will still tell you your defect rate. Pre-funding QC, run at full coverage, is how you make that number go down.

---

_Source: [https://www.seiright.com/blog/pre-funding-qc-ai-agent-fannie-d1-2-reverifications-defect-taxonomy](https://www.seiright.com/blog/pre-funding-qc-ai-agent-fannie-d1-2-reverifications-defect-taxonomy) · Sei AI_