Who This Is For
Heads of retail and consumer-direct lending, marketing operations leaders who own lead spend, and Chief Compliance Officers who have been asked to approve an AI agent that calls borrowers.
The Lead Was Never the Problem
Every mortgage lender I talk to has done the same math. They spend a meaningful amount per purchased or paid-search lead, and they know their contact rate falls off sharply in the first few minutes after the form is submitted, because the borrower filled in three forms that evening and the first lender to call gets the conversation. Then they look at when their leads actually arrive. A large share come in after hours and on weekends, which is when people shop for houses, and those leads sit until a loan officer gets to them on Monday morning. By then a competitor has already called.
The usual fixes are a round-robin dialer, an after-hours call center, or asking loan officers to work their phones at night. None of those scale, and loan officers who spend their evenings dialing are not spending them on files that are already in process. That is the job we built the Loan Officer Assistant to take: answer every lead in seconds, qualify it against the lender's own criteria, book the loan officer, and keep following up on the ones who did not pick up.
What Happened When We Ran It
We put the agent on both ends of the funnel for a top-25 independent mortgage bank in California: outbound speed-to-lead on new web leads, and 24/7 inbound for anyone who called the lender's numbers. It was connected to their existing CRM and to the loan officers' calendars, so a qualified borrower got a real appointment with a named loan officer before they hung up.
Appointments set with loan officers rose 9.2 times on average, and lead-to-funded conversion went up 12%. The number I remember most is a softer one. Several loan officers told us they had to start blocking hours out of their own calendars, because the agent was booking them faster than they could take the calls. That was a good problem, but it was still a problem, and it is why booking rules are something the lender sets rather than something the agent decides: how many appointments a loan officer takes in a day, and which loan officers are licensed in the borrower's state.
The rest of this post is about what we had to get right before any of that was allowed to happen, because the compliance team signed off on the agent long before marketing saw a result.
The 9pm Problem: Calling Hours Are Local, and Some States Are Stricter
The federal TCPA rule at 47 CFR 64.1200(c)(1) bars telephone solicitations before 8 a.m. or after 9 p.m. in the called party's local time. So a lead that lands at 8:58 p.m. can be called. One that lands at 9:02 p.m. cannot, not by a person and not by an agent. "Answer in seconds" has to mean "answer in seconds if the clock allows it," and the agent has to know which clock.
That is harder than it sounds, because the called party's local time is not always the lender's, and a borrower with a 415 area code who lives in Phoenix and is buying in Tampa has three plausible time zones. We have the agent resolve it to the strictest: if any time zone tied to the lead is outside the permitted window, it waits. Then there are the state laws. Florida's Telephone Solicitation Act, section 501.059, narrows the window to 8 a.m. to 8 p.m. and caps solicitation calls to the same person on the same subject at three in a 24-hour period. Other states have their own versions. All of these go into the Rulebook as rules with the state they apply to and the date they took effect, and the agent reads them for every attempt rather than following a fixed dialing cadence.
So what does the agent do with the 9pm Sunday lead? It does not call. If the borrower's consent covers text messages, it can send a short message saying a loan officer's assistant will call in the morning, with a link to book a time directly. At 8:00 a.m. in the strictest applicable time zone, it calls.
Inbound is different. A borrower who calls the lender at 9pm on a Sunday has initiated the call, and the agent answers it. That was where much of the after-hours lift came from in the California deployment: calls that used to go to voicemail and rarely converted were answered and qualified.
Consent Has to Be Provable, Call by Call
In February 2024 the FCC issued a declaratory ruling (FCC 24-17) confirming that calls using AI-generated voices are calls using an "artificial or prerecorded voice" under the TCPA. That settled the question for voice agents: an outbound marketing call from an AI agent to a cell phone needs prior express written consent, the same as a prerecorded message.
Then, in January 2025, the Eleventh Circuit vacated the FCC's one-to-one consent rule in Insurance Marketing Coalition v. FCC, the rule that would have required consent to name a single seller and limited calls to topics "logically and topically related" to the consent. A lot of lead buyers read that decision as permission to go back to the old way. I read it more narrowly. The court held the FCC exceeded its authority; it did not say a borrower who ticked a box on a comparison site consented to every lender who later bought the lead. Your consent record still has to show that this borrower agreed to be called by this lender, and in a TCPA suit the lender carries the burden of proving it.
So the agent will not dial a lead without a consent record attached, and it logs the record it relied on with every call: the form, the disclosure language shown, the timestamp, the source. When a borrower says "stop calling me," in whatever words they use, the agent treats it as a revocation, confirms it, and suppresses that number across the CRM. The FCC's revocation rules now require lenders to honor a revocation made by any reasonable means. An agent is better at this than a tired person at the end of a dialing shift, because it does not try one more rebuttal.
The Agent Books the Loan Officer and Stops There
Regulation Z's loan originator definition in 12 CFR 1026.36(a)(1) covers anyone who takes an application, offers or negotiates credit terms, or assists a consumer in obtaining or applying for credit, for compensation. The SAFE Act licensing line follows that activity. We wrote about where an AI assistant sits relative to that line in detail; for speed-to-lead the practical rule is short. The agent does not quote a rate. It does not say whether the borrower will qualify. It does not recommend a loan product. It asks the qualifying questions the lender defined, books a licensed loan officer, and passes along a summary.
Borrowers test this. One of the most common questions on a speed-to-lead call is some version of "what rate can you get me." The agent's answer is that the loan officer it is booking will go through pricing with them, and that it can get them on the calendar today. We score every call against that line, the same way the Compliance Agent scores a loan officer's calls. When the agent's wording drifts toward anything that sounds like a term or an approval, the call is flagged and the script gets corrected before the next exam finds it.
The 9pm Lead Gets the Same Questions as the 9am Lead
The fair-lending risk in speed-to-lead is quieter than the TCPA risk and harder to see in a single call. Regulation B at 12 CFR 1002.4(b) prohibits making oral or written statements to applicants or prospective applicants that would discourage a reasonable person from applying, on a prohibited basis. Human qualification calls vary. A loan officer who is tired, or who has a quota, spends less time on the lead that seems less likely to close, and "seems" is where bias gets in.
An agent can make that consistent, but only if you design for it. Our agent asks every lead the same qualification questions for the same loan purpose, offers the same appointment availability, and does not change its tone or its follow-up cadence based on anything that could proxy for a protected characteristic. It also handles calls in the borrower's language where the lender supports it, which is one place an agent does better than an after-hours call center. Because every call is logged and scored, the lender can run the test that a human operation rarely can: compare booking rates, follow-up counts and time-to-first-contact by geography and language, and see whether any group of leads is being treated differently. In the California deployment the guardrails covered TCPA, UDAAP and Fair Housing, and every call was logged for QA, which is what made that comparison possible.
Where This Goes Wrong
Most speed-to-lead failures I have seen came from configuration: a dialer set to "call immediately" with no time-zone logic, a consent field mapped from the lead vendor's feed and never checked, a follow-up cadence built for one state and copied to another, a script approved by compliance in March and quietly edited by marketing in June. What prevents those is that the calling-hours rule, the consent requirement and the script are in one version-controlled Rulebook with an owner and an effective date, and the agent cannot act on a rule it cannot cite.
If you are evaluating speed-to-lead with AI, the first question to ask a vendor is not how fast it calls. Ask it to show you the record for a single call: the consent it relied on, the local time it calculated, the state rule it applied, the script version it used, and what it said when the borrower asked about rates. If that record does not exist, the speed will cost you more than it earns.
Pranay Shetty
CEO & Co-Founder