The Lock Desk: AI for Rate Locks, Extensions, and the Pricing-Exception Trail Fair-Lending Examiners Follow

Two Different Jobs Wearing One Name

The lock desk does two jobs that look like one. The first is lifecycle mechanics: lock a loan at the right price for the right period, extend it when the loan runs long, relock it when it expires under worst-case pricing, apply a float-down when the policy allows one. That work is arithmetic against a rate sheet and a policy, high-volume and time-boxed, and it is exactly the kind of work an agent runs well.

The second job is pricing exceptions, and it is a different animal. When a loan officer asks to shave an eighth off the price to match a competitor or save a deal, someone decides whether to grant it and on what basis. That decision is discretion, and discretion applied to price is the single most examined surface in fair lending, because the Equal Credit Opportunity Act and Regulation B do not care whether anyone intended to discriminate. They care whether the outcomes fall unevenly across a protected class, and unmonitored pricing discretion is the classic way uneven outcomes happen without anyone deciding to make them happen.

We build the agent that runs across mortgage origination on lender platforms, and the lock desk is a good place to be precise about the difference between automating mechanics and governing discretion. The agent can own the mechanics outright. On the exceptions, its job is not to decide, it is to make every exception follow policy and leave the trail an examiner will ask for. What follows is where it operates and where the line sits.

The Lifecycle Is Rule-Bound Math

A rate lock is a commitment to a price for a period, and everything that happens to it afterward is governed by the lender's secondary-marketing policy and the rate sheet in effect. Lock a loan for forty-five days, and the price reflects that period. Let it run to day fifty, and the loan needs an extension, priced per the policy's extension schedule, with the cost allocated to whoever the policy says pays. Let it expire, and a relock triggers worst-case pricing, the worse of the original lock and current market, because the policy will not let a borrower re-lock into a better market for free after an expiration. Offer a float-down, and it applies only under the conditions the policy defines and only once, usually, and only inside a band.

None of that requires judgment. It requires knowing the policy, reading the rate sheet, tracking the lock's age against the loan's actual timeline, and computing the right number without transposing it. Manual lock desks miss on exactly those points under volume: an extension priced off yesterday's sheet, a relock that quietly gave the borrower the better of two markets, a lock that expired because nobody watched the calendar and turned into a repricing the lender absorbed. The agent watches every lock's clock against the loan's real milestones, computes extension and relock pricing off the current sheet and the governing policy, and flags the lock that is about to expire before it does, so the desk acts on a live lock instead of cleaning up an expired one.

That is straightforward value and it is not where the compliance weight sits. The compliance weight is in the exceptions.

Pricing Exceptions Are Where Fair Lending Lives

A pricing exception is a departure from the price the model produced, granted for a stated business reason, most often to meet a competing offer or to keep a relationship. Exceptions are legitimate, common, and necessary in a competitive market. They are also, in the aggregate, a fair-lending dataset, because the question an examiner asks is not whether a given exception was justified. It is whether exceptions, and the favorable pricing they produce, were granted at different rates to similarly situated borrowers across protected classes.

The CFPB and the prudential regulators have treated discretionary pricing and exceptions as a fair-lending risk for years, and the mechanism is always the same: discretion exercised loan by loan, without a consistent standard and without monitoring, produces disparities that no individual decision intended. The lender's defense is not that every exception was reasonable in isolation. The defense is a documented exception policy, applied consistently, with reason codes captured cleanly enough that the fair-lending team can measure exception rates and pricing outcomes by protected class and show the distribution is explained by legitimate factors. That defense lives or dies on data the lock desk either captured at the moment of the exception or did not.

The failure we see most is not lenders granting improper exceptions. It is lenders granting reasonable exceptions and capturing them as a free-text note in a loan file, so that when the fair-lending team goes to analyze exception patterns, the reason for each exception is a sentence a loan officer typed and the comparator, the competing offer that justified it, is nowhere. The exceptions might be perfectly defensible and the lender still cannot prove it, because the monitoring data was never structured.

What the Agent Enforces on an Exception

The agent handles a pricing exception as a governed workflow, not a note. When an exception is requested, the agent requires the exception to attach to a policy-recognized reason before it can move, because an exception with no category is the exact thing that becomes unanalyzable later. A competitive match carries the competing offer as a structured comparator, not a paraphrase. A relationship or retention exception carries the basis the policy defines. The agent enforces the approval authority the policy assigns, so an exception beyond a loan officer's latitude routes to the desk manager or secondary with the full context attached, and the exception cannot self-approve past its threshold.

Then the agent captures the exception the way fair-lending monitoring needs it: the reason code from a fixed taxonomy, the size of the concession, the comparator or basis, the approver, and the timestamp, all as structured fields tied to the loan. That is the difference between a fair-lending analysis the lender can run on demand and a reconstruction project it does under an exam. In our deployments the structured reason code is the field we treat as non-negotiable, because an exception without a categorized reason is invisible to monitoring, and invisible exceptions are how a clean book looks dirty when someone finally measures it.

There is a related boundary at Regulation Z 1026.36(d), the loan-originator compensation rule, which bars an originator's compensation from varying with the loan's terms and tightly limits when a pricing concession may come out of the originator's compensation. The commentary permits it only in narrow cases, such as a bona fide unforeseen increase in a settlement cost. A concession funded the wrong way is an LO-comp violation regardless of how reasonable the price looked, so the agent checks the source of a concession against the rule and blocks the funding path the rule does not allow, rather than leaving it to the loan officer to remember which pocket the eighth of a point is allowed to come from.

The Line the Agent Does Not Cross

The agent does not grant discretionary favorable pricing on its own, and the reason is the same reason discretion is examined in the first place. An agent that could hand out concessions would be exercising exactly the unmonitored discretion the rule worries about, at machine speed and machine volume, which is worse, not better. So the agent enforces the policy and runs the mechanics, and a human authority holds the discretion the policy reserves for a human. When an exception is inside a loan officer's defined latitude and carries a valid categorized reason, the agent lets it proceed and records it. When it exceeds that latitude, the agent routes it up with everything the approver needs and lets the approver decide. The agent's contribution to fairness is not that it makes the pricing calls. It is that it makes sure no call is made off-policy and no call is made without the data that lets the lender prove, later, that the pattern holds up.

What This Changes for the Desk and the Exam

For the desk, the lifecycle stops leaking. Locks do not expire unwatched, extensions and relocks price off the current sheet and the right policy, and float-downs apply inside their conditions, because the agent runs that math every time instead of when a person gets to it. That is straightforward operational recovery, and it is the part that shows up in cycle time and in fallout.

For the exam, the exception book becomes analyzable. Because every exception carries a fixed reason code, a structured comparator, the approver, and the concession size, the fair-lending team can measure exception rates and pricing outcomes across protected classes whenever it wants, not only when an examiner asks, which means a developing disparity is something the lender sees early and can explain or correct rather than something it learns about from a supervisory finding. At Sei we build the lock agent to own the mechanics and govern the exceptions, because the lock desk is where price meets discretion, and the lender that can produce a clean, structured record of every exception is the one whose fair-lending posture survives the sample. The math the agent runs. The discretion a human keeps. The trail the agent guarantees, on every exception, is the part that decides how the exam goes.