AI on the Collateral Desk: The AVM Quality-Control Rule Now in Effect, the Reconsideration-of-Value Process, and the Appraisal-Independence Line the Agent Cannot Touch

Two Rules That Redrew the Collateral Desk

For most of the last decade the collateral desk ran on a stable set of rules: order an appraisal, run it through the GSE collateral tools, review the exceptions, and clear the loan. Two regulatory changes in the last two years moved the ground under that process, and both bear directly on where an AI agent can operate.

The first is the interagency Quality Control Standards for Automated Valuation Models rule, published in the Federal Register on August 7, 2024 and effective October 1, 2025. It requires institutions that use an AVM in a covered credit decision or securitization determination to maintain policies and control systems designed to ensure a high level of confidence in the estimates, protect against data manipulation, avoid conflicts of interest, require random sample testing and reviews, and comply with applicable nondiscrimination law. The nondiscrimination factor, added to the four Dodd-Frank quality-control factors, is the one that changed the compliance posture: an AVM is now explicitly a fair-lending surface.

The second is the interagency Guidance on Reconsiderations of Value of Residential Real Estate Valuations, finalized July 26, 2024. It describes how a lender may let a borrower request that an appraiser reassess a value when the borrower believes the appraisal is deficient or reflects discrimination, and how that reconsideration-of-value process can sit inside the lender's risk management. The guidance says on its face that it has no force or effect of law and imposes no new requirements, so it does not by itself create a borrower right or make a nonfunctioning ROV process a violation. What gives an ROV process teeth is the separate law it sits on top of, the ECOA and appraisal-bias exposure, UDAAP, and the GSE and investor requirements that do bind, plus the practical reality that examiners read a lender's valuation practices against those obligations. The collateral desk that treats the guidance as a description of sound practice and builds the process the binding rules and investors actually require is the one that holds up.

We build the agent that runs across origination, and the collateral desk is one of the places its role is most tightly bounded, because appraisal independence puts a line around valuations that the agent cannot cross. What follows is where the agent works on the collateral desk, and where the rules stop it.

What the Agent Reviews and What the AVM Rule Now Requires

The agent's collateral review is a consistency and completeness review against the loan's requirements, and it is exactly the kind of review that benefits from running on every file. It reads the appraisal and checks the Uniform Appraisal Dataset fields for internal consistency, confirms the appraisal is the current form and complete, reconciles the appraised value against the sales price and the loan-to-value the program allows, checks that the comparable-sales approach is present and the adjustments are within expected ranges, and flags the appraisal exceptions the GSE collateral tools raise for a human reviewer. It also confirms the collateral documents the file requires are present, the flood determination, the Reg B valuation-rights notice triggered at application, and the appraisal delivery to the borrower on time.

Where the AVM rule changes the agent's job is the control layer around any AVM the lender uses, and the fair-lending factor specifically. An AVM used to set or confirm value in a covered transaction now has to sit inside a control system, and an agent that consumes an AVM output has to treat that output as a governed model input rather than a fact. That means the agent's use of an AVM is subject to the random sample testing and review the rule requires, the AVM's performance has to be monitored for accuracy and for disparate results across geographies and protected characteristics, and the agent cannot select among AVMs or valuation inputs in a way that would let it shop for a value. The model-risk-management discipline the lender applies to its models now has to extend to the AVMs the collateral process relies on, and the agent has to operate inside it rather than around it.

The nondiscrimination factor is the one the collateral desk historically did not think about as a valuation-quality issue and now has to. An AVM that systematically produces lower values in certain neighborhoods is a fair-lending problem even if its overall accuracy is acceptable, and the rule now requires the control system to look for exactly that. The agent's contribution is that the AVM outputs it consumes are logged with the property and loan attributes needed to run that disparate-results analysis, so the monitoring the rule requires has the data to run on.

The Reconsideration-of-Value Process the Agent Can Run

The ROV guidance is where an agent adds the most operational value on the collateral desk, because ROV is a borrower-facing intake and coordination process that lenders have historically run inconsistently or not at all. The guidance describes a process that lets a borrower who believes an appraisal is deficient, whether from a factual error, an omitted comparable, or discrimination, request a reconsideration, and requires the lender to route that request, gather the borrower's information, and get it to the appraiser for a reassessment, all inside the lender's risk-management controls.

The agent runs the intake and the coordination. It receives the borrower's ROV request through the channels the borrower uses, captures the specific basis, the comparable the borrower believes was missed, the factual error the borrower identifies, the concern about bias the borrower raises, and structures it into the record the process needs. It checks the request against the lender's ROV policy for completeness and timing, assembles the borrower-provided information, and routes it to the appropriate reviewer and, where the policy directs, to the appraiser for the reassessment. It tracks the ROV to resolution so the request does not fall into a gap, and it documents the outcome and the communication back to the borrower.

What the agent produces is a functioning, documented ROV process that treats every request the same way and leaves a record an examiner can follow. What it does not do is decide the value or pressure the appraiser, and that is where appraisal independence draws the line.

The Appraisal-Independence Line

Regulation Z at 12 CFR 1026.42 prohibits coercion, extortion, collusion, bribery, or intimidation of, or any attempt to influence, a person who prepares a valuation with the purpose of causing the value to be based on a factor other than the appraiser's independent judgment. The rule bars anyone with an interest in the transaction from causing or attempting to cause the value to be misstated, and it protects the appraiser's independence as the mechanism that keeps the collateral value honest. The ROV guidance is careful to say that a reconsideration of value is not a violation of appraisal independence, because it is a request to reassess based on information, not an instruction to reach a number.

Regulation Z draws the line inside 1026.42 itself, and it is narrower than a blanket ban on talking to the appraiser. The prohibition at 1026.42(c)(1) bars coercion and any attempt to cause the value to be based on a factor other than the appraiser's independent judgment. The permitted-actions provision at 1026.42(c)(3) then says plainly what stays allowed: asking the preparer to consider additional appropriate property information, to provide further detail or explanation for the value, or to correct errors in the appraisal. A reconsideration of value lives inside (c)(3). Relaying that the borrower disputes the value and supplying recent sales the appraiser did not consider is permitted, because it asks the appraiser to consider information, and the appraiser stays free to keep the original value.

What (c)(1) still forbids is pressure toward a number. The agent cannot condition anything on the value reaching a target, cannot reward or threaten based on the outcome, and cannot tell the appraiser what value the loan needs, because each causes the value to turn on a factor other than the appraiser's judgment. We build the agent's ROV routing to carry the borrower's disputed value, the identified errors, and the supporting comparables, which (c)(3) allows, while carrying nothing that conditions the outcome on a target, which (c)(1) prohibits. The value stays the appraiser's to set or keep.

This is a fine line and the agent has to sit on the right side of it, which is a matter of what the message asks for rather than whether a message goes at all. Routing "the borrower identifies the following recent sales within a quarter mile that do not appear in the comparable set and asks that they be considered" is a (c)(3) request, and it is fine even though it communicates that the borrower thinks the value is low. Routing "the borrower needs the value to come in at least at the contract price" is a (c)(1) violation, because it names the outcome the value is supposed to reach. The difference is whether the message asks the appraiser to consider information or tells the appraiser what to conclude, and building the routing to stay on the (c)(3) side of that line is the compliance control we engineer.

The Failure Mode We Engineer Against

The pattern that produces the worst collateral outcomes now comes in two shapes. The first is the lender that adopted AVMs across its valuation process and did not build the control system the AVM rule now requires, so its AVM use is ungoverned, unmonitored for disparate results, and out of compliance as of October 1, 2025. The second is the lender whose ROV process exists on paper but does not function, so borrower requests to reconsider an appraisal fall into email threads, get handled inconsistently, and surface as an examination finding when the CFPB or a prudential regulator asks the lender to demonstrate the process works.

The architecture we run addresses both. The AVM outputs the collateral agent consumes are governed model inputs, logged with the attributes the rule's monitoring requires, and used inside the control system rather than as free-floating values. The ROV process is a real workflow the agent runs end to end, every request captured the same way, routed on time, tracked to resolution, and documented, with the borrower's information passed to the appraiser in language that conveys facts and never a target. The collateral desk becomes a place where the valuation is governed, the borrower's right to challenge it functions, and the appraiser's independence is protected by design rather than by hoping no one on the desk says the wrong thing.

The Honest Read

The collateral desk changed under two interagency actions: the AVM quality-control rule that took effect October 1, 2025 and made automated valuations a governed, fair-lending-monitored model surface, and the July 2024 ROV guidance that made a functioning reconsideration process an examination expectation. An AI agent can run the collateral review at full coverage, govern the AVM inputs inside the control system the rule requires, and operate the ROV intake and coordination as a real workflow instead of a paper policy.

The agent stops at appraisal independence. Regulation Z 1026.42(c)(1) prohibits pressuring a valuation toward a predetermined number, while 1026.42(c)(3) expressly permits asking the appraiser to consider additional information or correct errors, and the agent's entire ROV role lives inside that permitted space: it carries the borrower's information to the appraiser without carrying a target. At Sei, we treat that line as the defining constraint of the collateral desk: the agent runs the process and governs the inputs, the appraiser owns the value, and the wording that keeps the reconsideration on the right side of the prohibition is something we design rather than leave to chance. The value stays the appraiser's independent judgment. Everything around it that can be made consistent and documented is where the agent works.