Loan Boarding QC: The Data-Mapping Defects an AI Agent Catches When Servicing Transfers

The Map Is Where the Harm Gets Built In

A mortgage does not change when its servicing transfers. The borrower, the note, the balance, the escrow, and the loss-mitigation plan they are three payments into are all the same the day after the transfer as the day before. What changes is the system of record, and the loan gets there through a boarding process that maps the prior servicer's data onto the new servicer's platform. That map is where the trouble is, because a field that boards wrong does not stay an abstract data error. It becomes a wrong monthly statement, a payment applied to the wrong bucket, an escrow shortage that was never real, or a collection call to a borrower who is current on a plan the new servicer cannot see.

The CFPB has treated servicing transfers as a data-integrity risk for a decade, and the reason is that the harm is systematic rather than random. When a boarding map has a defect, it does not hit one loan, it hits every loan the defect touches, and it hits them all in the same billing cycle. Bulletin 2020-02 on mortgage servicing transfers is explicit that transferees are expected to have policies and procedures to ensure information is transferred accurately and to test and validate the data they receive, and Regulation X at 1024.38 puts the same duty in the rule: servicers must maintain policies reasonably designed to support the accurate transfer of information, with particular attention to loans in loss mitigation and with pending borrower requests.

We build the agent that runs across mortgage servicing on lender and servicer platforms, and loan boarding is a place where full-coverage reconciliation changes the outcome, because the defect that funds is always the one the sample missed. What follows is what the agent reconciles, how it ranks what it finds, and where the correction stops being the agent's to make.

Boarding QC on a Sample Misses the Loan That Fails

The structural problem with manual boarding QC is the same coverage problem that limits manual quality control everywhere. A human team validates a sample of the boarded population, because reconciling every loan's every field by hand against the source records does not scale against a transfer of tens of thousands of loans that has to be validated before the first billing cycle on the new platform. So the review sees a fraction, and a boarding defect that affects a subset of loans, the adjustable-rate loans, the loans with a particular escrow item, the loans mid-plan in loss mitigation, can pass a sample cleanly and still produce a wrong statement for every borrower in that subset.

An agent changes the coverage math the same way it does in pre-funding QC. It can reconcile 100 percent of the boarded loans against the data the transfer delivered and, where available, against the source documents behind that data, which turns boarding QC from a spot check into a gate the population passes through before it bills. That is the point of putting an agent here: not to review boarding faster, but to review all of it, because "all of it" is the only coverage that catches the subset defect before the borrower does.

What the Agent Reconciles on Every Loan

The agent recomputes and cross-checks the fields that drive what the borrower sees and pays. It verifies the unpaid principal balance, the interest rate, and the next payment due against the note and the prior servicer's records, because a balance or rate that boarded wrong produces a wrong statement immediately. For an adjustable-rate loan it checks the index, the margin, the caps, the current rate, and the next adjustment date against the note, since an ARM that boarded with the wrong index or margin will calculate the next change wrong and generate an adjustment notice that is wrong on its face. It validates the escrow balance and the escrow items against the most recent escrow analysis, because an escrow balance that transferred incorrectly is one of the most common and most harmful boarding defects, producing either a shortage the borrower does not owe or a cushion that hides one.

The agent checks the status flags that manual boarding drops most often, and these are the ones that cause the sharpest harm. A loan in an active loss-mitigation plan has to board with that plan intact, because a borrower three payments into a trial modification who boards onto the new platform as delinquent will get collection activity that Regulation X does not permit against a performing plan, and continuity of loss mitigation across a transfer is exactly the case 1024.38 tells servicers to protect. A loan in bankruptcy has to board with its bankruptcy status and any proof-of-claim posture intact, because collection against a loan in bankruptcy is a stay violation. A loan with a pending notice of error or information request has to board with that request alive, because the transfer does not reset the borrower's clock. The agent confirms each of those flags carried, and it flags the loan where the status did not survive the map.

It also reconciles the transfer against the notices the rule requires, because Regulation X at 1024.33 governs the transfer-of-servicing disclosures the borrower is owed, and a boarding process that gets the data right but the notices wrong has still created a violation.

Ranking Mismatches by Who Gets Hurt and When

Not every boarding mismatch is equal, so the agent does not treat them equally. It sorts what it finds by borrower-harm risk and by proximity to the next billing cycle, because a servicer validating a transfer against a hard go-live date needs to clear the harm-causing defects first, not work an undifferentiated exception list.

The top tier is the mismatches that hit the next statement and the borrower's money: a wrong balance, a wrong rate, a wrong payment, a wrong escrow figure. These produce an incorrect bill or an incorrect draft in the first cycle, so they have to clear before the cycle runs. The next tier is the status flags that govern whether the servicer may act at all: loss-mitigation continuity, bankruptcy, pending disputes, active forbearance. A defect here does not always change the number on the statement, but it changes whether a collection or foreclosure action is lawful, so it is high-harm even when it is not high-dollar. The last tier is the documentation and completeness gaps that need resolving for the record but do not, by themselves, produce a wrong borrower-facing action in the first cycle. Ranking this way means the servicer spends the tight pre-boarding window on the loans where a miss becomes harm, rather than distributing attention evenly across defects that are not evenly dangerous.

In our deployments the escrow and loss-mitigation-status mismatches are the two categories we watch hardest, because they are the ones that turn a data error into a wrong payment or an improper collection, and they are the ones a sample is most likely to skip when they affect a subset of the book.

Where the Correction Stops Being the Agent's to Make

The agent reconciles and ranks, and it does not silently rewrite the servicing system of record. That boundary matters because the servicing record is the operative truth for what the borrower owes and what the servicer may do, and a correction to it is a governed act with its own consequences. When the agent finds that an escrow balance boarded wrong, it flags the discrepancy with the source figure and the boarded figure and the analysis behind the source, and a servicing operator confirms and applies the correction through the servicer's controlled process. When it finds that a loss-mitigation plan did not carry, it surfaces the plan and the payments made against it, and the servicer restores the plan and honors it, because honoring an in-flight plan is a legal obligation the servicer owns, not a data fix the agent should make on its own authority.

The reason for the boundary is the same as everywhere the agent touches a regulated record: an agent that could overwrite the servicing system at machine speed and machine volume would be a faster way to propagate a mistake, not a safer way to correct one. So the agent's authority is to detect, quantify, and route, with the evidence attached, and the servicer's authority is to correct the record and take the borrower-facing action. The agent makes sure no harm-causing mismatch is invisible going into the first cycle. The servicer makes sure every correction is made through a process that is itself auditable.

What This Prevents

The boarding defects that generate complaints, examinations, and enforcement are the ones that reach the borrower before anyone catches them: the wrong first statement, the escrow shortage that was never owed, the collection call to a borrower on a plan, the ARM notice calculated off a bad index. Every one of those is preventable at boarding, when the source data is fresh and the loan has not yet billed on the new platform, and every one of them is expensive after boarding, when the fix means correcting a statement already sent, refunding a draft already taken, or answering a regulator about a collection that should never have happened.

Running the reconciliation at full coverage and ranking it by harm is what moves those defects from post-transfer discovery to pre-transfer correction. At Sei we build the boarding agent to reconcile every loan against its source records before the population bills, to put the escrow and status defects at the top of the list because those are where harm concentrates, and to flag rather than overwrite, because the servicing record is a truth a human has to correct through a controlled process. The transfer does not change the loan. The map should not either, and the job of the agent is to make sure it did not, on every loan, before the first statement goes out.