Verification of Employment With an AI Agent: The Written VOE, the Day-of-Closing Verbal, and Why The Work Number Is a Credit Report
The Verification That Decides Whether the Income Is Real
Income calculation gets the attention because the math is visible, but the number an underwriter computes is only as good as the employment behind it. A borrower can hand over pay stubs that reconcile to the penny and W-2s that tie to the tax transcript, and the file still fails if the employer on those documents is not a real employer, or if the borrower stopped working there three weeks before the note date. Verification of employment is the control that stands behind the income, and it is the control most often run as an afterthought, a phone call someone makes the day before closing because the checklist says to.
We build the agent that runs across mortgage origination on lender platforms, and VOE is a good example of a task that looks like one step and is actually three, each with its own timing, its own documentary standard, and its own way of going wrong. The written verification early in the file, the verbal verification close to the note, and the reverification when the file changes are different controls, and an agent that collapses them into "confirm the job" is an agent that will confirm a job that no longer exists.
Three Verifications, Not One
The first is the written VOE, or the documentation that substitutes for it. Fannie Mae's Selling Guide accepts standard income documentation, pay stubs and W-2s, or a completed written Request for Verification of Employment (Form 1005), and it accepts data from a verified third-party database as an alternative. This is where the file establishes that the employment exists and that the income has the history the loan type requires. The agent's job at this stage is to assemble the documentation the profile calls for and to check it against the rest of the file, so the employer, the dates, and the income on the VOE reconcile to the pay stubs, the W-2s, and the 4506-C transcript rather than sitting in the file unread.
The second is the verbal VOE, and it is a timing control, not a documentation control. Under Fannie Mae's Selling Guide B3-3.1-07 on verbal verification of employment, the lender must verify employment within ten business days prior to the note date for a borrower with employment income, and within a longer window for self-employed borrowers, where the requirement is confirming the business exists through a third-party source within 120 calendar days of the note. The point of the verbal is to catch the thing the written VOE cannot: the borrower who was employed when the file opened and is not employed when the loan closes. A job loss, a resignation, a leave of absence between application and closing changes the ability to repay, and the verbal is the last look before the money moves.
The third is reverification, which the file triggers rather than the calendar. When something in the file changes after verification, a new pay stub that does not match, an income figure that moved, a discrepancy a QC review surfaced, the employment has to be looked at again against the new fact. This is the control that pre-funding QC leans on, and it is the one most likely to be skipped, because it does not appear on the original checklist.
What the Agent Runs at Each Stage
At the written stage, the agent reads the employment section of the URLA, determines what documentation the income type requires, and reconciles every employment fact across the documents in the file. Employer name, address, position, start date, and income have to agree across the pay stub, the W-2, the application, and any VOE form or database report. A mismatch is not a rejection, it is a flag with the specific fields that disagree, so the processor sends the borrower a precise question rather than a request to re-explain their whole employment history.
At the verbal stage, the agent tracks the note date and works backward to the ten-business-day window, because the verbal is only valid inside it. It schedules the verification to land inside that window rather than whenever someone remembers, and it verifies through a source it can stand behind: a direct call to the employer's HR line or published number, a listing in a verified third-party database, or the documentation the guideline allows for self-employment. The record it writes is the record a repurchase reviewer will read, the name and title of the person contacted or the database queried, the date, the phone number and how it was sourced, and the employment status confirmed. A verbal VOE that says "confirmed" with no trail is a verbal VOE that did not happen as far as an investor is concerned.
We added the day-of-note recheck after watching a file where the written VOE was clean, the income tied out, and a verbal placed inside the window reached an HR line that said the borrower's last day had been the prior week. The borrower had resigned to start a new job and had not mentioned it, because from their side the income was still coming. The file was days from funding. That is the entire reason the verbal exists, and it is why the agent treats the window as a hard gate rather than a formality: the loan does not advance to closing without a verbal inside the window, logged, with the status the source actually gave.
The Fraud Surface VOE Sits On
Employment and income misrepresentation is one of the more common forms of mortgage fraud, and the VOE is where a fabricated employer either holds up or falls apart. A fraudulent file will often have internally consistent documents, because the same party produced all of them, so consistency across the pay stub and the W-2 is necessary but not sufficient. The verification has to reach outside the documents to a source the borrower did not control.
The agent's contribution here is to treat the employer as a fact to be verified independently, not as a field to be copied. A business whose only evidence of existence is the documents in the file is a business the agent flags for third-party confirmation. A phone number that traces back to a cell phone rather than a business line, an employer address that resolves to a residential property or a mail drop, an HR contact whose details match the borrower's own contact information, these are the patterns that separate a real employer from a manufactured one, and they are patterns worth surfacing before the verbal call rather than after. The agent does not adjudicate fraud, it assembles the signals and routes them, because a misrepresentation determination and any resulting referral to the lender's BSA function is the lender's to make, not the agent's.
The Part Most Programs Get Wrong: The Database VOE Is a Credit Report
Instant, database-driven VOE has made verification faster, and it has quietly moved a compliance obligation into a place many lenders do not look. When a lender pulls employment and income data from a payroll database such as The Work Number to make a credit decision, that data is a consumer report and the database operator is a consumer reporting agency under the Fair Credit Reporting Act, 15 U.S.C. § 1681. The obligations that attach are the credit-use obligations, not the employment-screening ones, and the two that matter most for VOE are the permissible-purpose requirement and the adverse-action duty a user of a consumer report owes under 15 U.S.C. § 1681m. If the lender denies or worsens the loan terms in whole or in part because of information in a database VOE, section 1681m requires an adverse-action notice that identifies the reporting agency, the same as with a credit score.
The trap is treating the database result as a neutral data lookup rather than as a consumer report. A lender that declines a file because the database showed a different income than the borrower stated, and never sends the FCRA adverse-action notice naming the database, has an FCRA violation sitting inside a step it thought was purely operational. Our agent tags every employment data source with what it is, so a database VOE that contributes to an adverse decision carries the FCRA flag into the adverse-action workflow with the reporting agency identified, and the notice names the source the way the statute requires. The borrower's dispute rights under FCRA run against that data too, so when a borrower says the database is wrong about their employment, the agent routes it as an FCRA dispute rather than as a customer-service question.
The Reverification Loop That Catches the Late Change
The verifications above are point-in-time, and loans change after each point. A borrower whose income was verified in week two can have a new pay stub in week five that shows a lower base and a bonus that is not guaranteed, which changes the qualifying income the file was approved on. The agent watches for the fact that should trigger a fresh look: a new income document that does not match the verified figure, a QC finding on the employment, a change in the borrower's stated employment, or the simple passage of time that pushes the verbal outside its window and requires a new one. When the trigger fires, it does not silently update the income, it opens a reverification with the specific reason attached, because the underwriter owns the decision about what the new fact means for the loan.
This is the same discipline we apply everywhere the agent touches a credit-relevant fact: the agent produces and verifies the inputs, and surfaces the change, and the human owns the determination. VOE is a place that discipline earns its keep, because the cost of a missed employment change is not a condition, it is a loan sold to an investor with a rep and warranty that the borrower was employed, made against a borrower who was not.
The Honest Read
Verification of employment is three controls wearing one name. The written verification establishes the employment, the verbal inside the ten-business-day window catches the borrower who left the job after application, and the reverification catches the change the file surfaces later. An AI agent earns its place by running each one on its own timing, reconciling the employment facts across the whole file, and treating the employer as something to verify independently rather than a field to copy, because employment misrepresentation is exactly the kind of fraud that survives internally consistent documents.
The part that separates a compliant program from a fast one is the database VOE. When employment data pulled from a payroll database drives a credit decision, it is a consumer report and the FCRA obligations attach, adverse-action notice and dispute rights included. At Sei, we build the VOE agent to run the verifications on the guideline's clock, to carry the FCRA tag on every database source into the adverse-action path, and to route the reverification when the file changes rather than closing it clean and moving on. The phone call is the easy part. Knowing which of the three verifications you are running, and what obligations the source you used carries, is the job.
Ramkumar Venkataraman
CTO & Co-Founder