# The Documents That Are Not Done When the Loan Funds: Trailing-Document and Collateral-File Reconciliation With an AI Agent, Without Touching What the Custodian Certifies

*September 22, 2026 · 4 min read · Ramkumar Venkataraman*

> A funded loan is not a delivered loan. The recorded security instrument comes back from the county weeks later, the final title policy trails, and until the collateral file is complete and certified the loan is a reps-and-warranties exposure. Fannie Mae requires a document custodian to take physical custody of the note and certify it, and for eNotes the authoritative copy is controlled on the MERS eRegistry. Here is where an AI agent reconciles the file and chases the trailing pieces, the match key that stops it from clearing the wrong exception, and the certifying act it is not allowed to perform.

## Funded Is Not Delivered

A loan that has funded feels finished to everyone except the people who have to sell it. The borrower has the keys, the wire has cleared, and the file moves to post-closing, which is where the quiet risk lives. The note exists, but the collateral file is not complete, because several of the documents that make the loan saleable do not exist yet on the day it closes. The security instrument has to be recorded by the county before the recorded copy comes back, and that can take weeks or, in a slow jurisdiction, months. The final title policy trails. Intervening assignments trail. Until those pieces arrive, are matched to the note, and the file is certified, the loan sits in an inventory that reads as an asset and behaves as a liability.

That liability is not abstract. When a loan is delivered to an investor with an outstanding collateral document, it is a representation the seller made and has not yet backed, and an unresolved trailing-document exception is one of the cleaner ways a loan comes back as a repurchase or indemnification demand. So the post-closing desk is a control surface, and it is exactly the kind of high-volume reconciliation work an AI agent is good at, right up to the legal act it is not allowed to touch.

## What the Custodian Actually Certifies

The rules here are specific, and they draw the line the agent has to respect. Fannie Mae requires the mortgage documents to be held by an approved [document custodian](https://selling-guide.fanniemae.com/sel/a3-3-04/document-custodians) that takes physical possession of the note and the collateral documents, and the custodian's [custody of the mortgage documents](https://selling-guide.fanniemae.com/sel/a3-3-05/custody-mortgage-documents) is the basis on which the loan is certified for delivery. Certification is a legal attestation by the custodian that the required documents are present and conform. It is not a status field. It is an act with liability attached, performed by the party the rule names, and it is not something an agent gets to assert on the custodian's behalf.

For an eNote the mechanics change but the boundary is sharper, not softer. An eNote has no paper original, so possession is replaced by control, and the authoritative copy is governed on the [MERS eRegistry](https://selling-guide.fanniemae.com/sel/b8-7-01/mortgage-electronic-registration-systems-mers-inc). Immediately on execution the eNote is registered with a unique 18-digit Mortgage Identification Number, and [delivering the eMortgage](https://selling-guide.fanniemae.com/sel/c1-2-04/delivering-emortgages-fannie-mae) means transmitting a Transfer of Control and Location request so the eRegistry shows the investor as Controller and Location of the authoritative copy. Transferring control of an eNote is a state change on a system of record, and it is precisely the sort of action that should never happen because a model inferred it was time. It happens because a human authorized it, and the agent's job is to have the file ready so that authorization is a confirmation and not an investigation.

## Where the Agent Earns Its Place

Inside those boundaries, most of the post-closing workload is reconciliation and chase, and both are work an agent does well. The agent reconciles the collateral file against the required document set for the loan's investor and product, identifies what is missing, and opens or ages an exception for each gap. It chases the parties who owe the trailing pieces, the county recorder for the recorded instrument and the title company for the final policy, on the cadence the exception requires rather than whenever someone remembers to check. When a recorded instrument arrives, the agent reads it and compares it to the note and the closing file, so a mismatch in the legal description, the borrower name, the recording data, or the MIN surfaces as a flagged exception instead of certifying its way into a pool.

The value is in the completeness, not the cleverness. A human post-closing team samples and prioritizes because it cannot look at everything, which means the exception that ages quietly is the one nobody was assigned to that week. An agent reconciles every file on every cycle, so the trailing-document report reflects the actual state of the inventory rather than the last time a person had capacity to update it. That is the same shift from sampling to full coverage that changes what [quality control catches](/blog/ai-mortgage-quality-control-pre-funding-post-close-100-percent-coverage), applied to the collateral file instead of the credit file.

## The Match That Almost Cleared the Wrong Exception

The failure I want to name came out of shadow mode, and it is the reason the agent's match logic looks the way it does. Early on, the agent matched an incoming recorded security instrument to an open exception primarily on borrower name and property, which is how a human skims the same task. On a refinance where the same borrower had two loans against the same property in a short window, it matched the recorded instrument to the wrong loan and would have cleared the wrong exception, which is worse than leaving both open, because it reports a file as complete that is not. Shadow mode caught it precisely because the agent's proposed clear was compared against the human's, and they disagreed.

The decision we made from that has held. A trailing-document match keys on the recording data and the MIN and the loan number together, not on the borrower, and any match that would clear an exception affecting the loan's saleability is presented to a human for confirmation rather than cleared by the agent. Reconciliation the agent owns. The clear that changes whether a loan can be delivered, a human signs. Everything the agent does writes back inside the [system-of-record boundary](/blog/ai-agents-inside-the-los-system-of-record-boundary-write-back-mortgage), with a dated trail of what it matched and on what evidence, so the certification the custodian performs is built on a record an examiner or an investor can follow.

## Why the Boundary Is the Product

It would be easy to build an agent that certifies files, transfers eNote control, and clears exceptions on its own, and it would be faster, and it would be exactly the design that turns a quiet post-closing backlog into a pool of loans certified on evidence nobody checked. The reason we do not is not caution for its own sake. It is that the two acts with real legal weight here, the custodian's certification and the transfer of control on the eRegistry, are acts the rules assign to specific parties and that a model has no business asserting. The agent's job is to make the file so complete and so well-reconciled that the human's act is a confirmation, which is where the time actually gets saved, because the slow part of post-closing was never the signature. It was assembling the truth the signature stands on. At Sei we put the agent on the assembly and leave the attestation where the rule puts it, and the loans deliver faster because the exceptions were worked, not because the controls were skipped.

---

_Source: [https://seiright.com/blog/post-closing-trailing-documents-collateral-file-document-custodian-ai-mortgage](https://seiright.com/blog/post-closing-trailing-documents-collateral-file-document-custodian-ai-mortgage) · Sei AI_