At a glance
- A cost-per-loan baseline worksheet — baseline hours per file × loaded cost × automation share, minus review time, exception time and run cost — gives you a defensible number your own pipeline supports, unlike a vendor's quoted range.
- Freddie Mac's own Cost to Originate research is the best-sourced public benchmark: a wide, dated gap between top and bottom performers, not a savings figure for any one workflow.
- Costs keep rising industry-wide even as digital-tool adoption grows, because technology spend, purchase-loan complexity and lower volume all push the other way.
- A savings-per-loan range can leave review time, exception handling and the system's own run cost out of the labor hours it removes — check for each one.
- Start the worksheet on one workflow with a written baseline before you price any vendor or build decision.
How much does it cost to originate a mortgage in 2026?
MBA publishes more recent quarterly cost-to-originate figures in its Mortgage Bankers Performance Report, but that release wasn't available for this article, so it isn't cited. The latest verifiable figure is Freddie Mac's analysis of retail-only lenders' financials: average cost to produce a mortgage was approximately $11,800 in the second quarter of 2025 [2]. That was down from $13,400 in Q1 2025 but about $200 higher than the $11,600 recorded in Q3 2023 [2]. Treat this as the latest verified reference point, not a live figure for your pipeline.
Freddie Mac's 2025 update also reported that lenders' highest pre-tax net income per loan since 2021 arrived at $900, even as production costs kept trending upward [2].
A single industry average hides how differently lenders actually perform. Freddie Mac's 2024 Cost to Originate study broke Q3 2023 costs out by performance quartile, and the spread is wide:
| Segment | Average cost per loan |
|---|---|
| Top-quartile performers | ~$6,900 [1] |
| Total market average | ~$11,600 [1] |
| Bottom-quartile performers | ~$16,500 [1] |
The gap between the top and bottom quartile is nearly 2.4 times [1]. That range is context for the worksheet below, not a number to plug into your own model. It describes the industry sampled in Freddie Mac's study, not your workflow, your volume mix, or your loan officers.
The cost-per-loan baseline worksheet
Building a defensible number starts with your own workflow, not an industry average. The worksheet below prices one workflow at a time — condition clearing, post-closing QC, loss-mitigation intake, whatever you're evaluating. It uses inputs you measure, not figures anyone else can hand you, and follows the same loaded-rate discipline as our general AI ROI framework, applied at the per-loan level a mortgage business case needs.
| Line | What it captures | Your number |
|---|---|---|
| Baseline hours per file | Hours your team spends on this workflow per file today, measured over a representative sample — not an estimate. | ___ hrs/file |
| Loaded hourly cost | Fully loaded hourly cost of the staff doing this work: wages, benefits, taxes and overhead. | $___/hr |
| Automation share assumption | Share of that time you expect the system to absorb, set from a pilot or a conservative estimate, not a vendor pitch. | ___% |
| Gross hours saved per file | Baseline hours × automation share. | = ___ hrs/file |
| Review time added per file | Hours your reviewer spends checking the system's output per file — measured, not assumed to be zero. | ___ hrs/file |
| Exception time added per file | Share of files that fall out to manual handling × hours per exception, averaged across your sample. | ___ hrs/file |
| Net hours saved per file | Gross hours saved − review time added − exception time added. | = ___ hrs/file |
| Labor saving per loan | Net hours saved × loaded hourly cost. | = $___/loan |
| Run cost per loan | This workflow's share of usage, hosting, monitoring and maintenance cost, divided by loan volume. | $___/loan |
| Net saving per loan | Labor saving per loan − run cost per loan. | = $___/loan |
How much can AI save per loan?
No verified figure covers "AI" broadly, because the workflows, vendors and baselines vary too much to average. The most detailed public number covers something narrower: Freddie Mac estimated that a lender could save roughly $230 to $570 per loan in personnel and cost-of-funds expenses by using its Loan Product Advisor tools — automated underwriting, asset and income modeling, and appraisal waivers, not generative AI. It also added one to eight days of shorter cycle time [1].
Separately, Freddie Mac's 2024 survey of 500 loan officers, originators and underwriters asked about a fully digital origination process. The majority believed it could cut cost per loan by as much as 40% [1]. That's a surveyed belief about a hypothetical future state, not a measured outcome — treat it as what the industry hopes for, not a number to build a business case on.
Why haven't origination costs fallen with AI adoption?
Three of the drivers Freddie Mac documents push against any technology-driven savings, and none of them is generative AI itself:
- Rising technology spend: technology-related expenses per loan rose from 2% to 4% over three years. Lenders Freddie Mac interviewed universally reported that vendor costs per seat or application have doubled in a short time, specifically in third-party solutions [1].
- A harder loan mix: today's volume is dominated by purchase originations, which are more complex, labor-intensive and time-consuming than refinances. Prior Freddie Mac research also found purchase loans more prone to errors and defects [1].
- Lower volume, higher per-loan allocation: industry-wide volume is lower than two or three years ago. That typically drives a higher per-loan allocation of total costs [1].
Generative AI specifically is early in this picture, not the cause of it. Freddie Mac's own 2024 research describes lenders as being "in the initial stages of investigating how AI can change mortgage production" [1]. A technology many lenders are still evaluating can't yet explain an industry-wide cost trend either way.
What costs do vendors leave out of mortgage AI ROI?
A savings-per-loan range is only half a business case. The other half is what it costs to run the system and to check its output — ask for both before you accept one.
| Cost line | Why to ask | What to ask for |
|---|---|---|
| Review time | A savings figure nets out labor removed, not labor added to check the output. | Minutes per file a reviewer spends checking the output, measured. |
| Exception handling | A savings range may describe the files that go smoothly, not the ones that fall out to manual handling. | Share of files that fall out to manual handling, and how long each takes. |
| Run cost | Usage, hosting and monitoring cost scales with volume — ask for it separately from any savings claim. | A per-loan run-cost figure at your expected volume, not a seat price. |
| Integration and upkeep | A pilot's setup cost is visible; ongoing upkeep against LOS and platform changes isn't. | What it costs when your loan origination system's next release changes the integration. |
A May 2026 ROI analysis from Tech 42 Consulting models per-loan labor savings and denied-application cost, and tells readers to factor in implementation and ongoing costs for their own solution. It doesn't state a run-cost or payback figure for those costs [4]. Ask for both before treating any vendor's number as complete.
Where to start: a first scoped workflow
Run the worksheet above on one workflow before you talk to any vendor. Start with a scoped workflow, not a full-pipeline rollout:
- Pick one workflow
Choose a workflow with a clear queue and an owner who can measure it. Condition clearing and post-closing QC are often good candidates if your policy for them is already written down.
- Measure your baseline before any pilot
Time a representative sample of files for baseline hours, loaded cost and error or rework rate, before automation changes the number.
- Run the worksheet with a conservative automation share
Use a share your pilot data supports, not a vendor's headline figure, and include review time and run cost from the first pass.
- Compare against Freddie Mac's benchmarks for context only
Use the quartile table above to sanity-check whether your total cost per loan looks typical, not to estimate your savings.
- Decide with the underwriter, not around them
Every credit decision and condition sign-off stays with the underwriter. That's this article's assumption, and your organization's own operating choice to keep or change.
Our loan underwriting document workflow page covers one such workflow's inputs, human authority and acceptance criteria in full. When you're ready to rank a scoped workflow like this against every other candidate in your operation, do it by cost and time-to-value. A 360° AI Blueprint does that ranking.
When this worksheet doesn't apply
- You haven't measured a baseline yet. Run the worksheet after a timed sample, not from memory — an unmeasured baseline makes every downstream number unreliable.
- The workflow's volume is too thin or seasonal for a stable baseline, such as one that only spikes during a refinance wave.
- You're pricing a deterministic, rules-based check with no model calls — conventional software TCO applies, not this worksheet.
- The goal is to have the system make or recommend the credit decision itself, not prepare the file faster; that changes the risk profile entirely.
Frequently asked questions
How much does it cost to originate a mortgage in 2026?
MBA publishes more recent quarterly production figures in its Mortgage Bankers Performance Report, but that release wasn't available for this article, so it isn't cited. The most recent verifiable benchmark, from Freddie Mac's 2025 update, shows per-loan production expense still trending upward versus Q3 2023, despite improved lender margins.
How much can AI save per loan?
There's no single verified figure for AI broadly. Freddie Mac estimated a conservative per-loan savings range from its own Loan Product Advisor tools — automated underwriting, asset and income modeling, appraisal waivers, not generative AI. It also included several days of shorter cycle time. It's narrower and better sourced than the vendor range cited above, which gives no source for it.
Why haven't origination costs fallen with AI adoption?
Several documented forces push the other way: technology-related expense per loan is rising, and today's purchase-heavy, more complex loan mix is more labor-intensive than refinances. Lower industry volume also drives a higher per-loan allocation of costs. Many lenders are also still in the early stages of investigating AI, so it hasn't had time to move the trend either direction.
What costs do vendors leave out of mortgage AI ROI?
A savings-per-loan range can leave out what it costs to run the system or check its output. That includes review time on flagged files, exception handling for files the system can't resolve, per-loan run cost at your actual volume, and ongoing integration upkeep against your loan origination system's changes. Ask for all four before accepting a savings range.
How long is the payback on mortgage automation?
There's no industry-wide figure — build cost, workflow, volume and vendor terms vary too much to average honestly. Payback is build cost divided by net annual saving. Net annual saving is this worksheet's net saving per loan (already net of run cost) times your annual loan volume — the same formula our general AI ROI framework uses, applied to your own numbers.
Related reading
- The general framework this worksheet specializes: how to calculate AI ROI before you build.
- What an AI system costs to run, line by line: AI project total cost of ownership.
- The workflow this article's worked example is modeled on: loan underwriting document workflow.
- Where AI agents fit across origination, closing, post-close QC and servicing: AI for mortgage lenders.
- Ranking this workflow against every other candidate in your operation: 360° AI Blueprint.
Sources
- Freddie Mac Single-Family, "2024 Cost to Originate Study: How Technology Impacts Savings, Cycle Time and Revenue" (May 2024) (accessed 2026-09-30)
- Freddie Mac Single-Family, "2025 Updates to the Cost to Originate Study" (2025-11-18) (accessed 2026-09-30)
- Ocrolus, "Mortgage Manufacturing Rates: Bending the Cost Curve with AI" (2025-10-09) (accessed 2026-09-30)
- Tech 42 Consulting, "Intelligent Document Processing for Mortgage Originators: An ROI Analysis" (2026-05-21) (accessed 2026-09-30)
More in Lending operations
How to Automate Borrowing Base Certificate Review: The Checks to Run Before Availability Is Approved
A check table and tie-out checklist for how to automate borrowing base certificate review: ineligibles, advance rates, and what stays a human decision.
Financial ServicesCross-Document Validation in Loan Files: Rules That Catch Income and Identity Mismatches
A cross-document validation rule table for loan files: how pay stubs, W-2s, tax returns, and bank deposits should be compared, with tolerances and routing.
Loan Covenant Monitoring Automation: Build, Buy or Layer an AI Agent on Your Portfolio System
Loan covenant monitoring automation: compare buying point software, extending your LMS, or building a custom agent, with a covenant test register template.
Financial Services