Nearly every lender surveyed has at least one AI use case in production and only about a quarter have fully scaled even one, according to a joint survey from AARMR, the Mortgage Bankers Association and BCG covering roughly 40 percent of US origination volume. The average respondent had around 10 of 38 surveyed use cases running. That gap between starting and scaling is the real state of ai for loan officers in 2026, and the reason for it is not the technology.
It is that the places AI helps most in a mortgage file are also the places most tightly regulated.
~25%
of lenders have scaled one use case
10 of 38
use cases in production, on average
55%
of brokers use AI daily or regularly
That last figure comes from a separate AD Mortgage survey of more than 250 brokers, which also found 72 percent expect their usage to grow over the next three years. Broker-level adoption runs ahead of enterprise scaling, which makes sense: a broker can adopt a tool on a Tuesday.
Where ai for loan officers is actually working
Three categories, in descending order of how well they hold up.
Document handling. Pulling figures out of statements, pay stubs and tax returns into a structured file. This is the single most reliable use. The documents are standardized, the task is extraction rather than judgment, and a mistake is visible when a human checks it.
Borrower communication. Status updates, document requests, chasing the thing that is holding up the file. Loan officers lose hours to "where are we" calls that an automated update prevents entirely.
Lead qualification and response. Answering an inbound inquiry quickly, capturing the basics, booking the call. This is the one with the clearest revenue link and the one that needs the most care, for reasons below.
Document extraction
Standardized inputs, checkable output. Start here.
Proactive status updates
Removes the call nobody wants to make or receive.
Fast lead response
The inquiry at 9pm that is gone by morning.
File completeness checks
Flagging what is missing before underwriting does.
Where it stops, and why
This is where mortgage differs from every other vertical, and it is not a matter of taste.
Anything that states a rate
If your marketing states a rate, Truth in Lending requires the annual percentage rate to appear too. Certain terms trigger further mandatory disclosures: a payment amount, a down payment figure, a number of payments, a finance charge. An AI drafting marketing copy does not know this and will produce a clean, persuasive sentence that is a Reg Z violation.
This is a flag for review, not legal advice. Any AI-generated marketing copy that mentions a rate, a payment, a term length or a down payment needs compliance review before it goes out. The liability sits with the originator, not the tool, and the NMLS ID belongs on the material.
Anything that implies approval
"You are approved," "guaranteed approval," "no credit check," or any claim about approval odds. An AI agent qualifying a lead will drift toward reassurance because reassurance converts, and reassurance about approval is the exact thing that cannot be said.
The instruction has to be explicit in the prompt. A model told to "be helpful and encouraging to borrowers" will produce approval language without being asked to.
Anything that is advice
Which product suits them, whether to buy points, whether to refinance. That is a licensed conversation.
Safe for AI
Collecting documents
Explaining what a document is for
Status updates on a live file
Booking the call
Flagging a missing page
Needs a licensed human
Recommending a product
Advising whether to refinance
Discussing rate or payment specifics
Anything about approval likelihood
Interpreting a credit situation
The qualification problem
The highest-value use is also the one with the narrowest safe path, so it is worth being precise.
A compliant AI qualification conversation collects facts and books a call. It does not quote, promise, estimate or reassure.
- 1
Disclose that it is an AI assistant
First line. Several states require it, the FCC has proposed requiring it federally, and a borrower who finds out later stops trusting the firm.
- 2
Collect, do not evaluate
Property type, occupancy, rough timeline, whether they are working with anyone else. Facts the borrower already knows.
- 3
Never state a number
No rate, no payment, no estimate, no range. "That is what the call is for" is a complete answer.
- 4
Escalate on anything unusual
Credit concerns, a complaint, a request for a human, or three failed attempts to understand.
- 5
Book the call and hand off
Name, source, the facts collected, the recommended next step.
That is a real improvement over an inquiry sitting in an inbox overnight, and it does not require the AI to say anything a regulator would care about.
Consent, before any of it
Automated calls and texts to a borrower carry the same rules as any other automated outreach. The FCC confirmed in February 2024 that AI-generated voices fall under the TCPA's artificial-voice restrictions, with statutory damages of $500 to $1,500 per call. Our post on AI cold calling laws covers the consent levels in detail.
Mortgage adds a wrinkle: a lead bought from an aggregator has not consented to you specifically, and consent does not transfer between companies.
Why scaling stalls
The AARMR, MBA and BCG survey identified regulatory uncertainty and limited measurable business results as the two barriers slowing broader implementation. Both are worth taking seriously.
Regulatory uncertainty is real and partly unresolvable right now, because the FCC's proposed AI-specific rules from August 2024 have not been finalized. Planning around a proposal is uncomfortable.
The measurement problem is more fixable. Most AI deployments are measured in time saved, which is the easiest number to produce and the least convincing one. Hours saved do not show up anywhere a CFO looks. Applications per loan officer per month does. Pull-through rate does. Cost per funded loan does.
Pick one use case, not ten
Measure it against a revenue number, not hours saved
Have compliance review the prompt, not only the output
Write the escalation triggers down before launch
Read a sample of transcripts weekly for the first month
Every respondent in that survey said they expect to increase AI spending over the next year. Spending more on ten half-finished use cases is how the next survey reports the same gap.
Frequently asked questions
What can AI do for a loan officer today?
Document extraction, borrower status updates, lead response and file completeness checks. Those four are reliable. Anything touching rate, payment, product recommendation or approval likelihood is a licensed conversation.
Does AI replace loan officers?
No. It removes the document chasing and the after-hours response gap. The parts of the job that require judgment, a license and a relationship are the parts that stay.
Can AI quote a rate to a borrower?
It should not. Stating a rate in marketing triggers Truth in Lending disclosure requirements including the APR, and other terms trigger further disclosures. Keep numbers in the human conversation.
Is ai for loan officers worth it at a small shop?
Often more than at a large one, because a broker can adopt a tool without an enterprise project. The AD Mortgage survey found 55 percent of brokers using AI daily or regularly, ahead of enterprise scaling rates.
What is the biggest mistake?
Measuring in hours saved. It is the easiest metric and it never survives a budget conversation. Tie the use case to applications, pull-through or cost per funded loan.
Do we need consent before an AI calls a borrower?
Yes. AI-generated voices fall under the TCPA, informational calls need prior express consent and marketing calls need prior express written consent. Consent given to a lead vendor does not transfer to you.
Where to start
One use case, measured against a revenue number, with compliance reading the prompt rather than only the output. Document extraction is the safest first move and lead response is the most valuable one.
If you want a look at where your inquiries are going cold, book a free strategy call. We go through response times, the follow-up path and what your CRM is doing with a lead after hours. It is a working session, not a pitch.
*Sources: AARMR, Mortgage Bankers Association and BCG joint AI survey, covering approximately 40 percent of US origination volume, 2026. AD Mortgage broker survey, 250+ respondents. FCC Declaratory Ruling 24-17, February 2024. Truth in Lending Act, Regulation Z. Last reviewed October 2026.*
