NewFed says Friday Harbor helped cut application-to-funded cycle time from 32 days to 26 days, lifted processor and underwriter throughput by more than 35 percent, and supported nearly $170 million more funded volume without adding fulfillment headcount. The useful read for alt-lenders is not "AI in mortgage." It is file quality before underwriting.

Most AI lending stories still land in the wrong part of the workflow. They focus on chat, marketing copy, or borrower communication. Those surfaces are visible, but they are not where the money leaks out. The expensive failures happen inside the file: missing documents, mismatched business names, stale bank statements, unclear ownership, duplicate applications, unverified collateral, and exceptions that show up only after the credit team has already spent time.

That is why the NewFed case is more useful than another generic "AI assistant" launch. Friday Harbor is being positioned as pre-underwriting, which means the software attempts to catch file-quality issues before the loan reaches the underwriter.³ In mortgage, that means conditions, documents, borrower file completeness, and post-close purchase readiness. In MCA and revenue-based financing, the translation is bank statement ingestion, cash-flow normalization, ownership verification, fraud flags, and stacked-debt signals. In factoring, it is invoice validation, debtor concentration, purchase-order support, notice-of-assignment status, and duplicate-invoice detection. In equipment finance, it is vendor verification, collateral detail, lien checks, and proof that the borrower entity actually matches the business asking for the money.

The reason this matters for a CRO or COO is simple: pre-underwriting is where you can measure AI without pretending the model is making credit judgment. Count touches per file, missing-condition reopen rates, time between application and complete file, and senior-underwriter exception hours. If those measures improve and loss performance does not degrade, the tool has earned its place. If the tool only makes the file look prettier, it has not.

The cleanest alt-lender benchmark is not "cycle time" by itself. Speed can hide risk if the lender funds faster by skipping checks. The better benchmark is speed plus file quality. A practical test would track three measures before and after implementation: complete-file rate at first underwriting touch, exception reopen rate after conditional approval, and funded-deal loss or early-default performance by cohort.

For MCA and revenue-based finance, the first AI pre-underwriting layer should answer five questions before a human credit person opens the file. First, does the legal business name, DBA, EIN, bank account name, and Secretary of State record line up well enough to proceed? Second, do deposits match the claimed revenue pattern, or are there unexplained spikes, circular transfers, or processor artifacts? Third, does the borrower show signs of stacking through outgoing ACH patterns, UCC filings, or repeated lender names in bank activity? Fourth, are there missing months, redacted statements, or PDF artifacts that should block review? Fifth, is the broker package complete enough to underwrite without another email loop?

For factoring, the AI layer should be narrower and more forensic. Does the invoice data match purchase orders and delivery evidence? Is the debtor real and active? Are invoice numbers reused? Has the debtor appeared in prior rejected files? The value is in forcing messy receivable data into a reviewable shape before an underwriter starts pricing advance rates.

For equipment finance, the relevant layer is collateral and entity consistency. The lender should want vendor identity checks, equipment description normalization, serial-number completeness, lien-search prompts, and borrower-entity verification before credit review. The NewFed case points to the same operating principle across every segment: the first AI win is not deciding better. It is making the file less expensive to decide.

The risk is that lenders confuse automation with control. A platform that speeds up document review can also speed up bad packages if it is not tied to fraud controls. MCA shops already know this. A clean-looking PDF can still hide ownership mismatch, a fabricated DBA, a borrowed bank account, or a stacked position that does not surface until collections. Factoring shops know the same problem in a different form: a perfect-looking invoice can still be duplicate, disputed, or attached to a debtor that has no intention of paying.

The second risk is auditability. If AI flags a file as ready, the lender needs a reason. If the tool misses a mismatch, the lender needs to know which source was checked and which was not. Valid Systems' Snowflake integration announcement is a useful adjacent signal here. Valid says its integration lets financial institutions run fraud and account-risk models inside Snowflake with enterprise-grade data governance, audit trails, and real-time decisioning.⁴ The platform details differ, but the governance point applies to every AI pre-underwriting workflow: the lender should be able to explain what data was checked, what model or rule ran, and who overrode it.

The third risk is integration debt. A pre-underwriting tool that sits outside the lender's CRM, LOS, document repository, bank-statement parser, and decisioning workflow can become another manual handoff. That is the old problem with a new interface. The NewFed numbers are interesting because the productivity claim is measured at the team level, not just at the screen level. Alternative lenders should demand the same proof: show the touches per funded deal before and after implementation.

The test does not need to be complicated. Pick one product line and one team. For four weeks, measure baseline intake-to-complete-file time, document-chase touches, underwriting reopen rate, and funded volume per operations employee. Then run the pre-underwriting layer on the same product line for a controlled cohort. Do not change pricing, approval policy, or broker routing during the test, or the result becomes unreadable.

The pass condition should be specific. A useful AI pre-underwriting layer should reduce manual document-chase touches by at least 20 percent, improve complete-file rate at first review, and keep early-default and fraud indicators flat or better. If the tool reduces cycle time but increases exceptions after approval, it is moving work downstream rather than removing work. If it reduces human touches but increases fraud misses, it is dangerous. If it reduces touches, preserves controls, and gives the credit team cleaner files, it is the first AI layer worth paying for.

The budget argument should also be operational, not abstract. A 35 percent throughput gain like NewFed reported is not just a staffing story.¹ It changes the shape of growth planning. Instead of hiring one more processor for every volume increase, the lender can ask whether better file prep lets the existing team handle more applications without pushing errors into underwriting. That is the right AI conversation for lending executives in 2026.

_{Source
1 NewFed Boosts Loan Volume, Cuts Cycle Times With Friday Harbor AI (National Mortgage Professional)
2 NewFed Mortgage Case Study (Friday Harbor, 2026 case study PDF)
3 Friday Harbor Platform (Friday Harbor product page)
4 Valid Systems Launches Integration on Snowflake AI Data Cloud (Business Wire via StreetInsider)
5 7(a) Working Capital Pilot Program (U.S. Small Business Administration, last updated February 6, 2026)
6 Freedom Bank Receives Preferred Lender Status for SBA's 7(a) Working Capital Pilot Program (PRNewswire, May 18, 2026)
7 MHC and Odessa Announce Partnership to Help Asset Finance Lenders Modernize Customer Communications at Scale (PRNewswire, May 18, 2026)
8 Auto Lenders, Consumers on a Tightrope (F&I and Showroom, May 13, 2026)}

Industry-leading credit platforms cannot build on stale business data.

Bectran needed Secretary of State verification that was powerful enough for risk management and simple enough to integrate into their existing AR platform.

Cobalt delivered real-time SOS data, all 50 states, timestamped screenshots, structured API responses, and a partner model built for enterprise-grade platforms.

Schedule a FREE demo call

Freedom Bank announced May 18 that it received Preferred Lender status for SBA's 7(a) Working Capital Pilot program. The SBA program supports monitored lines of credit up to $5 million, with 75 percent guaranty above $150,000, maturity up to 60 months, and pricing caps that fall to base rate plus 3 percent for loans of $350,001 and greater.^{5 6} This is a direct competitor to cleaner factoring, ABL, and working-capital deals where the borrower can tolerate SBA documentation. It will not replace speed-sensitive MCA, but it does raise the bar for lenders trying to win bankable borrowers with price alone.

Valid Systems says its Snowflake AI Data Cloud integration lets financial institutions run machine-learning fraud and account-risk models inside Snowflake, with audit trails and data governance from Snowflake infrastructure. The company says it processes more than 70 million transactions monthly and guarantees more than $6 billion in immediately available funds every month.⁴ For alt-lenders, the lesson is not Snowflake itself. It is that fraud models are moving closer to governed production data. Any AI underwriting tool that cannot show an audit trail will look weaker as this standard spreads.

MHC and Odessa announced a partnership May 18 to embed customer communications management into Odessa's asset finance workflows for auto and equipment finance lenders. The release frames the problem as high-volume communications, regulatory changes, disconnected systems, and manual processes.⁷ The read for MCA, factoring, and equipment finance is that servicing communication is no longer back-office admin. Notices, payment reminders, payoff letters, hardship communications, and regulatory disclosures are becoming workflow software. The lender with cleaner servicing communication has fewer manual touches and a better audit record.