Three Documents That Look Identical and Are Not
A borrower shopping for a house asks for a letter. The letter says, in some arrangement of words, that they can borrow up to a number. To the borrower and the real estate agent, a prequalification, a preapproval, and a conditional approval are the same piece of paper, and digital lenders have spent a decade making that paper faster to produce. Put an AI agent on the top of the funnel and it will produce them faster still, on the borrower's channel, in seconds.
The three documents are not the same under the law, and the differences decide what the lender owes the borrower the moment the letter goes out. One of them, issued carelessly, starts clocks and creates obligations the growth team never sees and the compliance team inherits two years later in an exam. So the first thing an AI agent at the point of sale has to know is not how to write a persuasive letter. It is which of the three it is about to write, and what each one triggers.
What Regulation B Does With the Distinction
Regulation B defines an application at 12 CFR 1002.2(f) as an oral or written request for credit made in accordance with the procedures a creditor uses for the type of credit requested. The official commentary to that section is where the pre-approval problem actually lives, because it draws the line between a prequalification request and an application, and it draws it based on what the creditor does, not on what the borrower asked for.
A prequalification request, under the commentary, is a borrower asking how much they might get and the creditor giving general information. That stays outside the application definition only as long as the creditor does not evaluate the specific borrower's information, reach a decision, and communicate a decline. The moment the creditor evaluates the file, decides the borrower does not qualify, and tells them so, the request has become an application, and the adverse action machinery at 12 CFR 1002.9 attaches: notice within 30 days, with the specific reasons. Preapproval programs that issue a written, time-limited commitment after a creditworthiness review are treated as applications in the first place under the same commentary, which means a declined preapproval is a declined application and owes the borrower a notice.
This is the part digital lenders cross without noticing. An AI agent that pulls credit, runs the file against a ruleset, and tells a borrower "you do not qualify right now" has not been helpful and informal. It has evaluated, decided, and communicated, which is the exact sequence the commentary says converts an inquiry into an application. The borrower is now owed an adverse action notice, and if the agent was built to be encouraging and move on, that notice never got sent.
The TRID Clock Hiding in the Same Conversation
Regulation B is not the only rule watching this conversation. If the pre-approval flow collects the six pieces of information that complete an application under 12 CFR 1026.2(a)(3), the borrower's name, income, Social Security number, the property address, an estimate of value, and the loan amount, then the Loan Estimate clock starts, and the lender has three business days to deliver it. A pre-approval tied to a specific property is the dangerous case, because a specific address is one of the six, and a thorough agent collecting income and pulling credit has likely gathered the rest.
So the agent that issues a property-specific pre-approval can trip two triggers in one session: the Reg B application definition and the TRID six-piece set. We treat both as events the agent has to recognize at the instant they fire, which is the same discipline we build into point-of-sale intake. The letter is not just a letter. It is a document the agent can only issue after it has checked what the file now obligates the lender to do.
How We Build the Agent to Stay on the Right Side
The agent classifies every request before it writes anything. It decides whether the interaction is a general prequalification, a preapproval under the lender's program, or a full application, using the lender's own encoded definitions of each, not a generic idea of what those words mean. That classification is logged with the facts that drove it, because the classification is the thing that determines every downstream obligation and it is the thing an examiner will test.
From the classification, the obligations follow mechanically. If the agent evaluated the file and the answer is no, it does not send a cheerful "let's reconnect later" message and close the thread. It routes the file into the adverse action workflow so the notice goes out within the Reg B window with the real reasons, drawn from the actual factors the decision rested on. If the interaction crossed the six-piece TRID trigger, the agent timestamps the application-received event and hands the file to disclosure so the Loan Estimate deadline is a system fact rather than a loan officer's memory. If the request was a true prequalification where no specific borrower evaluation occurred, the agent documents that too, because "we did not evaluate" is a claim the lender has to be able to prove, not assert.
The letter itself is a compliance-approved, versioned artifact, the same way every disclosure the agent reads is. The agent fills the template. It does not improvise the commitment language, because a pre-approval letter that overpromises is both a sales problem and, if the borrower relied on it, a potential unfair-or-deceptive-practice problem. The words that state what the lender is and is not committing to are written once, reviewed, and reused.
The Fair-Lending Exposure Nobody Prices In
Here is the risk that does not show up on the conversion dashboard. A pre-approval decision is a credit decision, and a credit decision made at volume by an agent applying criteria is a fair-lending surface. If the criteria the agent applies to decide who gets the strong letter and who gets "not yet" produce different outcomes across protected classes, the lender has a disparate impact problem that began at the top of the funnel, before anyone thought of this as underwriting.
A human loan-officer funnel has this exposure too, and it is harder to see because it is diffuse across people and judgment calls. An agent applying one documented ruleset is more testable, not less, which is the point we make to compliance teams who worry that automating pre-approvals widens fair-lending risk. It widens the risk only if the criteria are hidden. When the criteria are the lender's, written down, and applied identically to every borrower, the lender can run the test, see the result, and fix the criterion, which is a thing a thousand individual loan-officer judgments never let it do. The automation is the thing that makes the control possible.
Where the Growth Team and the Compliance Team Stop Fighting
The reason pre-approvals are a perennial source of findings is that they sit on the seam between two teams measured on opposite things. Growth is measured on how many strong letters go out and how fast. Compliance is measured on whether the letters that went out created obligations nobody met. Built separately, the two systems fight, and growth wins the argument because its metric is on this quarter's dashboard and compliance's cost is a future exam.
We build the conversion flow and the obligation logic as one path, so the agent that produces the letter borrowers love is the same agent that recognizes an application, sends the adverse action notice, and starts the disclosure clock. The letter the borrower wants and the record the examiner wants come out of the same code, which means the question of whose metric wins never has to be asked. That is the version of a fast pre-approval worth building, because it is still fast after the exam, and it is the posture we hold to at Sei: the borrower's experience and the lender's defensibility are not a trade, as long as they are built together.
Pranay Shetty
CEO & Co-Founder