Salesman Hal

AI assistants and your mortgage business: a loan officer’s guide

Borrowers now ask ChatGPT, Claude, Perplexity and Google’s AI Mode what they can afford, which loan fits, and who to talk to. Here’s what those assistants look at, the fixes that help a broker or loan officer get recommended, and the compliance rules that still apply when the reader is a machine and the lead arrives through one.

Updated

How borrowers use AI assistants now

Many borrowers now start by typing their whole situation into an assistant: “I make $85,000, I have $15,000 saved and a 680 credit score. Can I buy a $350,000 house in Aurora? FHA or conventional? Who are good local lenders?”

There’s some data on how common this is. Bank of America’s 2026 Homebuyer Insights Report (a survey of 2,000 adults, fielded in April and May 2026) found that 20% of prospective buyers and current homeowners had used AI tools or chatbots for homebuying research in the past 12 months, rising to 28% of millennials and 32% of Gen Z. Among prospective buyers who used AI, the top use was estimating affordability, mortgage payments or closing costs (57%). The same survey found buyers still prefer people for high-stakes steps.

Borrowers use assistants to understand loan types (FHA or conventional, what a DSCR loan is), to compare rates, often pulled from rate pages without the assumptions behind them, to find a loan officer (“who’s good for VA loans near Fort Carson?”), and increasingly to take the next step by sending an inquiry or booking a call.

Lenders have noticed. In 2026 LendingTree launched a ChatGPT app for exploring rates; its disclaimer says it gives no lender-specific quotes, approvals or pre-qualification. Better launched one that underwrites loans, built for loan officer teams on ChatGPT Enterprise rather than borrowers. For a broker or small lender the question is simpler: when a borrower asks about your market, does the assistant know you, describe you correctly, and know how to reach you?

What assistants look at

No assistant publishes how it picks a lender, and anyone promising a formula is guessing. What we can say is what assistants can read, and what a borrower’s question needs answered:

The borrower wants to know Where the assistant looks
Do you do this kind of loan? Loan program pages on your site, your Google Business Profile, directory listings
Can you lend in my state? Licensed states on your site, NMLS Consumer Access, areaServed in your markup
Are you legitimate and any good? NMLS IDs, reviews on Google and elsewhere, loan officer bios, how long you’ve been lending
What will it cost? Rate pages (with APR and assumptions), fee explanations, payment examples
What do I need to qualify? Program pages, document checklists, an FAQ
How do I get started? A pre-approval or contact form that works, a booking link, a phone number someone answers

None of this works if assistants can’t load your site. A robots.txt rule or a security setting aimed at AI training can also block the assistants that send borrowers. Our guides to robots.txt for AI crawlers and Cloudflare and AI agents cover the settings.

Fix 1: a page for each loan program

An assistant can only say “yes, they do bank-statement loans for self-employed borrowers” if a page says so. A “Loan Programs” page with one line per product isn’t enough. Give each program you actually offer its own page:

  • Conventional, FHA, VA and USDA. Say plainly who each is for. You can say an FHA loan is insured by the federal government; don’t suggest the government endorses you.
  • Jumbo, with how you define it in your market.
  • Non-QM products such as DSCR and bank-statement loans. These are often business-purpose or investor loans, which can change which rules apply; ask counsel how to describe them.
  • Refinance, HELOC, first-time buyer and down payment assistance pages if you do them.

On each page, say who the program suits and who it doesn’t, the requirements in general terms (property types, occupancy, documents), the states you lend in, your NMLS ID, and a clear next step. Be careful with specific numbers: a down payment percentage is a Regulation Z trigger term, so it brings disclosures with it. Put it all in plain text, not a PDF or image; some AI fetchers read only the raw HTML.

Fix 2: loan officer pages and licensing

Borrowers ask for people, not companies: “Who’s a good loan officer in Fort Collins who speaks Spanish?” Give every loan officer a page with:

  • Name, title, individual NMLS ID, and the states they’re licensed in.
  • Specialties (VA, first-time buyers, self-employed borrowers, investors), languages spoken, and how to reach them directly.
  • Reviews that mention them by name, collected honestly.

Then check your company and every loan officer in NMLS Consumer Access, the public license lookup. It lists each NMLS ID with addresses, phone numbers, website and state licenses. If the phone or website there differs from your site, fix one or the other; an assistant that finds two numbers has to guess.

Fix 3: schema.org markup

Structured data states your facts in a format machines read without guessing. There’s no mortgage-broker type in schema.org, but the types you need exist:

  • FinancialService, a kind of LocalBusiness, for your company.
  • Person for each loan officer. It supports jobTitle, worksFor, knowsLanguage, identifier and hasCredential.
  • MortgageLoan, a kind of LoanOrCredit, for each program. It inherits loanType, loanTerm, requiredCollateral, interestRate and annualPercentageRate, plus provider and areaServed.

Here’s an example for a company and one loan officer. The NMLS IDs go in identifier as a PropertyValue; replace the zeros with your real IDs.

<script type="application/ld+json">
{
  "@context": "https://schema.org",
  "@type": "FinancialService",
  "@id": "https://summitlending.example/#company",
  "name": "Summit Home Lending",
  "url": "https://summitlending.example/",
  "telephone": "+1-720-555-0188",
  "address": {
    "@type": "PostalAddress",
    "streetAddress": "1600 Wynkoop St, Suite 200",
    "addressLocality": "Denver",
    "addressRegion": "CO",
    "postalCode": "80202",
    "addressCountry": "US"
  },
  "areaServed": [
    { "@type": "State", "name": "Colorado" },
    { "@type": "State", "name": "Wyoming" }
  ],
  "identifier": { "@type": "PropertyValue", "propertyID": "NMLS", "value": "000000" },
  "employee": {
    "@type": "Person",
    "name": "Dana Ortiz",
    "jobTitle": "Senior Loan Officer",
    "url": "https://summitlending.example/team/dana-ortiz",
    "identifier": { "@type": "PropertyValue", "propertyID": "NMLS", "value": "0000000" },
    "knowsLanguage": ["en", "es"],
    "hasCredential": {
      "@type": "EducationalOccupationalCredential",
      "credentialCategory": "license",
      "name": "Mortgage loan originator license",
      "validIn": { "@type": "State", "name": "Colorado" }
    }
  }
}
</script>

And one for a program page:

<script type="application/ld+json">
{
  "@context": "https://schema.org",
  "@type": "MortgageLoan",
  "name": "FHA purchase loan",
  "url": "https://summitlending.example/loans/fha",
  "description": "A purchase loan insured by the Federal Housing Administration, for a home you will live in.",
  "loanType": "FHA",
  "domiciledMortgage": true,
  "requiredCollateral": "The home being purchased",
  "provider": { "@id": "https://summitlending.example/#company" },
  "areaServed": [
    { "@type": "State", "name": "Colorado" },
    { "@type": "State", "name": "Wyoming" }
  ]
}
</script>
  • Notice what’s missing. The program example has no interestRate, annualPercentageRate or loanTerm. A rate or repayment period in your markup is arguably advertised, and a term is a trigger term. Leave them out unless compliance has signed off on the full disclosures next to them.
  • Don’t mark up your own star rating from testimonials on your site. Google recommends review markup only for sites that review other businesses.
  • Keep the markup in step with your pages, your Business Profile and NMLS Consumer Access. Conflicting facts are worse than none.

Fix 4: a next step an assistant can complete

An assistant that has picked you still has to make contact. The easiest paths are a short form and a booking link.

  • Separate “talk to a loan officer” from “apply”. A short request form (name, contact details, goal, state, timeline) is something an assistant can fill in. The full application, with Social Security number, income and assets, belongs in your loan origination system’s secure portal.
  • Offer a booking link for a short call with a loan officer.
  • Use clear labels (“State where the property is”, not “Field 7”), only the fields you need, and a confirmation that says what happens next.
  • Watch puzzle CAPTCHAs. They can stop an agent acting for a real borrower. Server-side checks and rate limits do less collateral damage.

The MCP guide and the WebMCP guide explain how a site can offer these actions to assistants directly, instead of hoping they can work the form. If you’re on WordPress, see our WordPress guide.

Fix 5: follow up fast

A borrower an assistant sent your way has often been given two or three names. The first loan officer to have a useful conversation has an advantage. Make sure AI-sent inquiries land in the same queue as your other web leads, that someone owns them, and that you answer during business hours in minutes, not days. If your team can’t always pick up, a speed-to-lead tool helps; our sibling company Callingly calls web leads back automatically and is used by many mortgage teams. Whatever you use, make sure how you call and text follows the consent rules below.

Compliance: AI-facing content is still advertising

Pages written “for AI”, FAQ answers your chatbot gives, schema markup, and the inquiry an assistant sends you are all ordinary marketing and lead handling in the eyes of your regulators. Nothing here is legal advice. Rules differ by state, license type and charter, so take this list to your compliance counsel.

NMLS IDs and licensing

  • Federal rules. Regulation Z requires the loan originator organization’s and the individual loan originator’s names and NMLS IDs on the credit application, the Loan Estimate and Closing Disclosure, the note and the security instrument. Under Regulation G, a registered loan originator at a bank or credit union must give their unique identifier on request, before acting as a loan originator, and in their first written communication with a consumer.
  • State rules for advertising. Many states go further for licensees. Virginia, for example, requires every advertisement by a licensed lender or broker to include “NMLS ID #” followed by the number and the NMLS Consumer Access web address. Check each state you’re licensed in; most lenders simply show the company and loan officer NMLS IDs on every page.
  • Only claim states where you’re licensed. A page that says “we lend nationwide” when you’re licensed in four states is a problem, and an assistant will repeat it.

Rates, trigger terms and the APR

Rate pages are what assistants quote most, and Regulation Z § 1026.24 governs them. The key points for closed-end loans:

  • If you state a rate, state it as an “annual percentage rate”, using that term, and say if it can increase after closing. For dwelling-secured loans, you can show the simple interest rate too, but not more prominently than the APR.
  • Trigger terms are the down payment amount or percentage, the number of payments or repayment period, the amount of any payment, and the amount of any finance charge. Use one, and the ad must also give the down payment, the repayment terms over the full loan (including any balloon) and the APR.
  • Payments in dwelling-secured ads must show every payment that will apply over the term, and for first liens, that taxes and insurance aren’t included and the real payment will be higher.
  • Prohibited practices include misleading “fixed” claims on adjustable loans, misleading comparisons, suggesting a product is a government loan program when it isn’t, claims that a loan will eliminate debt, and calling a for-profit broker or lender a “counselor”.

For AI specifically: date every rate, show the assumptions (loan amount, credit score, points, occupancy) next to it, and take old rate pages down. Assistants can quote a stale or partial number. You can’t control their summary, but you can make sure the source page is complete and current.

Fair lending

  • Regulation B bars any oral or written statement, “in advertising or otherwise”, that would discourage a reasonable person from applying because of a prohibited basis such as race, sex, age or receipt of public assistance. Its commentary gives the example of telling a retired applicant not to bother applying. That applies to your FAQ pages and to any chatbot answering for you.
  • Equal housing. Banks supervised by the FDIC must show in mortgage ads that they lend without regard to race, color, religion, national origin, sex, handicap or familial status. The Equal Housing Lender or HUD’s Equal Housing Opportunity logo satisfies that. Other lenders and brokers should ask counsel which rule applies to them; many display the logo on every page.
  • Encouraging is fine. Regulation B’s commentary allows content that encourages groups who might not otherwise apply, such as a Spanish-language page or a first-time buyer guide.

Privacy and data security

  • Data a borrower gives you, including through an AI assistant, can be nonpublic personal information under the Gramm-Leach-Bliley Act. Regulation P covers privacy notices and limits on sharing with unaffiliated third parties.
  • The FTC Safeguards Rule names mortgage brokers and lenders as covered financial institutions. It requires a qualified individual, a written security program with a risk assessment, encryption, multi-factor authentication, and FTC notification within 30 days of a breach affecting 500 or more people.
  • Add any tool that receives leads, Salesman Hal included, to your vendor review, and keep inquiry forms short.

Calling and texting leads an AI sends (TCPA)

  • The one-to-one consent rule is gone. The FCC’s 2023 rule requiring consent to name a single seller was vacated by the Eleventh Circuit in Insurance Marketing Coalition v. FCC, with the mandate issued April 30, 2025. In July 2025 the FCC restored the earlier text of its rule (order DA 25-621).
  • Written consent still applies. Under the restored rule, marketing calls or texts using an autodialer or a prerecorded or artificial voice need prior express written consent: a signed agreement, electronic signatures allowed, that names the phone number and clearly says the borrower agrees to such calls and that agreeing isn’t a condition of purchase.
  • An AI-sent inquiry usually isn’t that. An assistant can pass on “please have a loan officer call me”, but it can’t sign a consent agreement for the borrower. Ask counsel how your team may follow up: a manual call, an email, or a text only after the borrower opts in themselves.
  • Opt-outs. The FCC’s order 26-67, adopted September 30, 2026, lets callers name exclusive opt-out methods and narrows opt-outs from informational messages. An opt-out from a marketing message still covers all future marketing calls and texts from you. Some states have their own telemarketing laws too.

Reviews

Assistants lean on reviews. The FTC’s Consumer Reviews and Testimonials Rule, in force since October 2024, bans fake reviews, paying for reviews that must be positive, undisclosed reviews from employees and their relatives, and intimidating people to remove reviews. Ask every closed borrower for an honest review, and reply to the critical ones.

How to check what AI says about you

Set aside an hour. Use ChatGPT, Google’s AI Mode, Gemini, Perplexity and Claude, signed out or in a private window where you can, and ask the way a borrower would:

Who are the best-reviewed mortgage brokers in [your city] for first-time buyers?
I’m a veteran buying in [your town]. Which local lenders do VA loans, and how do I get pre-approved?
What are [your company]’s rates on a 30-year fixed today? What’s the APR?
Is [loan officer] at [your company] licensed in [state]? What’s their NMLS ID?
Can [your company] do a DSCR loan on a rental property in [state]?
Set up a call for me with a loan officer at [your company] this week.

For each answer, note:

  1. Were you mentioned? If not, who was, and what do their sites have that yours doesn’t?
  2. Were the facts right? Programs, licensed states, NMLS IDs, phone number. Click the cited pages to find where a wrong fact came from, and fix it at the source.
  3. What did it say about rates? If it quoted your rate without the APR or an old rate, fix or remove the page it came from, and tell compliance.
  4. Could it act? Ask it to set up a call. Note where it got stuck.
  5. What did it say about reviews? If “slow to respond” comes up, that’s an operations fix.

Answers vary between runs, so repeat the same prompts monthly and look for trends. To see which assistants visit your site and what they read, see measuring AI agent traffic; to tie it to funded loans, see AEO reporting.

How Salesman Hal handles mortgage businesses

Hal gives your business its own MCP server, a connection AI assistants can use directly. Assistants can read your business info and search and read your pages, so they describe your programs from what you’ve written. Its optional chat can answer from your site with your approved answers taking precedence, from approved answers only (word for word, anything else referred to you), or be off; approved-only lets compliance review every sentence it can say. When a borrower wants a loan officer, Hal sends the request through your existing contact form (Gravity Forms, Contact Form 7 or WPForms through the WordPress plugin, or a HubSpot or plain HTML form), or emails it to you. With Calendly, Cal.com, SavvyCal, Acuity or HubSpot Meetings, assistants can book a call, tagged as AI-sourced. Assistants must confirm the borrower asked and confirmed their details before Hal sends anything, and requests are rate-limited. Hal never ticks a consent box, so an AI-sent lead carries no calling or texting consent, and a form that requires that box can’t be sent through Hal. Hal keeps a copy of each lead in your dashboard. It doesn’t connect directly to your LOS, CRM or pricing engine, can’t quote rates, pre-qualify or take applications, and doesn’t make your content compliant; that stays with you and your counsel.

See how Hal works

FAQ

Can ChatGPT pre-approve a borrower?

Not on its own. Some lenders have built ChatGPT apps: LendingTree’s lets borrowers explore rates but gives no lender-specific quotes or pre-qualification, and Better’s underwrites loans for loan officer teams on ChatGPT Enterprise. Most borrowers still get a shortlist and are handed to a loan officer.

Do I have to show my NMLS ID on my website?

Federal rules require NMLS IDs on application and loan documents, and many states require them in advertising, sometimes in a set format. Your website is advertising. Most lenders show the company and loan officer IDs on every page; check each state’s rule with counsel.

Should I publish rates?

Only if you can keep them accurate and compliant: the APR alongside any rate, the assumptions, a date, and full disclosures if you use a trigger term. Assistants quote rate pages, sometimes out of context, so a stale page does more harm than none.

Does a lead from an AI assistant come with TCPA consent?

Usually not. An assistant can pass on a request to be contacted, but it can’t sign the written agreement the FCC rule describes for autodialed or prerecorded marketing calls and texts. Ask counsel how your team may follow up.

Is schema markup worth it for a mortgage business?

Yes, for the facts that don’t change often: your company, loan officers, NMLS IDs, licensed states and programs. Leave rates and terms out of markup unless compliance has approved the full disclosures next to them.

How do I know if AI is sending me borrowers?

Ask new borrowers how they found you, and add “ChatGPT or another AI assistant” to the options. Check analytics for referrals from chatgpt.com, perplexity.ai and similar, and log AI agent visits to your site.

Sources

Further reading

Let AI assistants connect borrowers to your loan officers.

Hal answers assistants’ questions from your own site and approved answers, sends inquiries into your forms and books calls on your calendar. Free during early access.

Or run the free AI-readiness scan to see what assistants can find on your site today.