Mortgage underwriting AI: Freddie Mac's $1,500 a loan and GreenState's 26%

Will Drewes

Founder, Fern Strategy · LinkedIn · 2 min read

Two verifiable numbers in a field full of press releases

Most AI coverage in mortgage lending is vendor-led and directional, with no hard figure attached to a named institution. Two examples have numbers.

Freddie Mac: $1,500 per loan, 5 days off the cycle

Freddie Mac's machine-learning automations inside Loan Product Advisor are reported to save originators up to $1,500 per loan and shorten the loan production cycle by 5 days, mostly by automating verification of income, assets and employment.

At volume, a mid-size lender closing 500 loans a month is looking at $750,000 a month in potential cost reduction from one workflow layer. The 5 days matter too. At current rates, closing speed is one of the few things borrowers can see when they compare lenders.

The setup is stable rules (GSE eligibility criteria) applied to inconsistent inputs (pay stubs, bank statements, employer records), which is the profile where automation pays.

GreenState Credit Union: 26% more approvals, no added risk

GreenState Credit Union deployed AI-assisted underwriting and reported a 26% increase in general approval rates. Approval rates for women and Hispanic applicants rose by 32%, with no added credit risk reported.

That second number deserves the attention. Fair lending is one of the highest-friction areas in mortgage operations, examiners scrutinize approval disparities across protected classes, and the working assumption has been that reducing bias risk costs approval volume. GreenState's result cuts against that.

The likely explanation is consistency. Human underwriters apply criteria unevenly across files. A model that is trained and audited properly applies the same criteria every time, so more qualified applicants get through and the approval pattern holds across demographic groups.

What this means if you run a lender

Freddie Mac's result sits in GSE infrastructure and GreenState's in a credit union's underwriting layer, so neither is a template. Both come from the same kind of workflow: high document volume, repetitive verification logic, rules that rarely change, and human review steps that add variance without adding judgment.

The harder part is governance. When an examiner asks why a loan was declined, "the model said so" fails. The institutions getting durable results built the audit trail into the workflow from the start.

If you're deciding where to begin, these two give you a benchmark. The test is whether your process can produce the same documentation when a regulator asks how the decision was made.