AI Agents for Mortgage Lenders: Why Making One Loan Still Costs Nearly $11,000

Every mortgage lender knows the feeling: rates move, volume comes back, and somehow the cost of closing a single loan barely budges. The numbers from the industry's own trade group make the point bluntly.

According to the Mortgage Bankers Association's second-quarter 2026 Quarterly Mortgage Bankers Performance Report, independent mortgage banks spent $10,936 in production expenses for every loan they originated in Q2 2026. That was actually good news. It was down from $11,898 per loan in the first quarter, and MBA said expenses hit their lowest level in basis points since the third quarter of 2021. But it is still far above the long-run average: from the second quarter of 2008 through the prior quarter, loan production expenses averaged $7,945 per loan.

So the cost of making a loan is coming down, but it's still well above where it used to be. That gap is where AI agents start to matter.

The Thin Margin Behind Every Closed Loan

MBA's report shows how little room there is. In Q2 2026, production revenue averaged $11,909 per loan against those $10,936 in expenses. That left a pre-tax net production profit of $973 per loan, or 25 basis points.

MBA's Marina Walsh called it the fifth straight quarter of positive production profits, "continuing the industry's turnaround from widespread losses between 2022 and 2024." She also noted that quarterly production profits "are still lower than the historical average of 39 basis points dating from 2008."

Put simply, lenders are back in the black, but the margin is thin. MBA's cost line covers commissions, compensation, occupancy, equipment, other production expenses and corporate allocations. Much of that cost is people working through a loan file: collecting documents, re-keying data, checking conditions and answering "where's my loan?" calls. When a margin is that tight, every hour spent on a file affects whether the loan makes money.

What Lenders Say They Actually Want From AI

Mortgage lenders have been fairly clear about what they want AI to do, and it isn't flashy. In Fannie Mae's 2023 Mortgage Lender Sentiment Survey special report on AI, which repeated a 2018 study, 73% of lenders said improving operational efficiency was their main reason for adopting AI, up from 42% in 2018. Improving the borrower experience as a top reason fell from 41% to just 7%.

The same survey found that adoption had barely moved:

• Only 7% of lenders said they had deployed AI or machine learning, down from 14% in 2018.
• 22% said they had started deploying it on a limited or trial basis, up from 13%.
• Among lenders not yet using it, the biggest barriers were integration complexity with existing systems, lack of a proven track record and high costs. Concerns about data security and privacy had also grown significantly.

The use cases lenders found most appealing were just as practical. Automated compliance review ranked first, followed by anomaly detection to catch fraud or defects early in underwriting. When asked what the GSEs should build, lenders pointed to appraisal automation, borrower income and employment verification, data and documentation reconciliation, and compliance management.

That is a list of back-office work, not front-office work. And it lines up closely with what AI agents, as opposed to single-purpose chatbots, are designed to handle.

Where AI Agents Fit in the Loan File

A chatbot answers a question. An agent carries out a sequence of steps across systems, checks its own work and hands off to a person when it should. The loan file is full of exactly that kind of multi-step, rules-driven work:

• Document collection and follow-up. Working out which documents are still missing for a file, requesting them from the borrower, following up on a schedule, and flagging the file when something arrives.
• Data reconciliation. Comparing figures across the application, pay stubs, bank statements and the loan origination system, then flagging mismatches for a processor rather than letting them surface at underwriting. Fannie Mae's respondents named this directly.
• Conditions tracking. Turning an underwriter's list of conditions into a tracked checklist, matching incoming documents to each item, and showing the processor what is still open.
• Pre-checks for compliance review. Running a file against the lender's own checklists before it reaches a human reviewer, so reviewers spend their time on judgment calls rather than hunting for missing fields. Lenders ranked compliance automation as the most appealing idea in the survey.
• Borrower and agent status updates. Drafting proactive status messages from the actual state of the file, so loan officers aren't pulled into routine "any update?" calls.

None of these tasks require an AI to make a credit decision. They take away the clerical work around the people who do.

Why the Human Touch Still Matters Here

Fannie Mae's researchers were careful on this point. Lenders, they wrote, "have stressed the importance of the 'human touch' in the mortgage business, particularly as it pertains to customer interactions," and borrowers have expressed a similar preference for human involvement in what is, for many, "the largest financial transaction of their lives." Lenders in the survey showed "little expectation that AI might replace or displace their employee base."

Fannie Mae's own outlook was that AI would likely "handle more of the back-end processing," while people "continue to build and maintain the customer relationships necessary to drive sales." That is a sensible way to design an agent rollout: the loan officer stays the borrower's point of contact, the underwriter keeps the credit decision, and the agent handles the follow-up work in between.

The Barriers Lenders Named Are Rollout Requirements

The barriers lenders cited (integration complexity, an unproven track record, cost and data security) are a useful checklist for any agent project:

1. Integration first. An agent that can't read from and write back to the LOS, the document system and email is just another window to copy and paste between. Ask how it connects before asking what it can do.
2. Prove it on one workflow. Pick one measurable step, such as the time from application to a complete document package or the number of conditions open at underwriting, and measure before and after.
3. Treat borrower data as the highest-risk asset in the building. Agents handling income, asset and identity documents need tightly scoped access, a full activity log and human approval for anything that leaves the building.
4. Tie cost to the per-loan number. MBA's per-loan expense figure is the scoreboard. If an agent doesn't move cost per loan or loans per employee, it isn't working.

What This Means for Mortgage Lenders Going Forward

The industry has worked its way back to profitability, with MBA reporting that 85% of the firms in its sample posted overall pre-tax profits in Q2 2026. But at $973 of production profit on a loan that costs nearly $11,000 to make, small efficiency gains add up quickly. The lenders who widen that margin probably won't be the ones with the most impressive AI demo. They'll be the ones who take the dull, repetitive steps out of every loan file and leave borrower relationships and credit decisions with people.

Frequently Asked Questions

How much does it cost a lender to originate a mortgage in 2026?

MBA's Q2 2026 Quarterly Mortgage Bankers Performance Report put total loan production expenses at $10,936 per loan for independent mortgage banks and mortgage subsidiaries of chartered banks, down from $11,898 in Q1 2026. The average from Q2 2008 onward is $7,945 per loan.

How many mortgage lenders are using AI?

In Fannie Mae's 2023 Mortgage Lender Sentiment Survey special report, 7% of lenders said they had deployed AI/ML and 22% said they had started on a limited or trial basis. Their main motivation was operational efficiency (73%).

Will AI agents replace loan officers or underwriters?

Lenders in Fannie Mae's survey showed little expectation that AI would replace their employees, and stressed the importance of the human touch with borrowers. The best near-term fit for agents is back-end work like document follow-up, data reconciliation and compliance pre-checks, with people keeping the borrower relationship and the credit decision.

What is the best first AI agent use case for a mortgage lender?

Start with something repetitive, measurable and low-risk, such as chasing missing documents or reconciling data before underwriting, and track it against cost per loan or time to a complete file.