There's a scene in Moneyball where the old-school scouts are in a room debating which players to draft based on "gut feel" and "the look of a ballplayer." Billy Beane walks in and tells them they're solving the wrong problem. The question isn't whether a guy looks like a baseball player. The question is: what does the data actually predict?

Real estate credit underwriting has been that room for sixty years.

A loan officer at a regional bank, a mortgage broker in a strip mall, a construction lender deciding whether to fund a $40M mixed-use development — they've all been operating on gut, relationship, and a handful of blunt metrics: FICO score, debt-to-income ratio, loan-to-value, tax returns. These aren't bad inputs. They're just deeply incomplete. And for decades, the real estate industry didn't have a better option.

That's changing — fast, and in ways most people in real estate still haven't fully processed. AI is moving up the entire credit complexity stack simultaneously: from rental payments at the bottom, through consumer mortgages in the middle, all the way to complex commercial real estate debt at the top. And what it's doing isn't just "faster underwriting." It's fundamentally redefining what a creditworthy borrower looks like.

The Problem Wasn't Data. It Was Synthesis.

The classic critique of traditional underwriting is that it ignores too much. A freelancer with $180K in annual income who pays rent on time for five years and has $90K in savings looks like a credit risk to a FICO algorithm because their income doesn't arrive on a W-2. A small multifamily developer who's delivered 12 projects in 8 years doesn't show up cleanly in a bank's underwriting template because none of those projects were large enough to generate a standardized track record.

The data to make better decisions has existed for years. Rent payment history, bank transaction data, property-level operating performance, local rental market comps, project delivery timelines — all of this is accessible. The problem was always synthesis at scale: pulling together heterogeneous data signals and producing a coherent, defensible credit judgment on thousands of loans simultaneously. That's precisely what large language models and modern ML architectures are built to do.

The result is an unlock that happens at every layer of the stack.

Layer One: Rental Credit and the Invisible Borrower

Before you can get a mortgage, you need a credit history. Before you have a credit history, you're renting. And for most of American rental history, paying rent on time — arguably the most consistent financial commitment in someone's life — counted for nothing in the credit system.

This is both a data problem and an equity problem. An estimated 45 million Americans are "credit invisible" or have thin credit files. A disproportionate share are younger, lower-income, or part of communities historically excluded from mainstream financial products. These aren't risky borrowers. They're unmodeled borrowers.

MetaProp portfolio company Split Pay (by Rent App) is tackling this directly, offering renter financial tools that include rent payment reporting to credit bureaus and flexible rent payment infrastructure. The thesis is straightforward but the implications are large: when rent becomes a credit-building instrument, the baseline population eligible for mortgage underwriting expands dramatically. You're not just solving a product problem. You're expanding the funnel that feeds the entire housing finance system.

Company Spotlight

Split Pay (by Rent App) — Renter financial tools including rent payment reporting to credit bureaus and flexible rent payment infrastructure. Turning America's most consistent financial obligation into a credit-building instrument.

This matters for lenders too. Every creditworthy borrower excluded by the current system is a customer acquisition opportunity left on the table.

Layer Two: The Mortgage That Actually Knows You

The residential mortgage process is a humiliation ritual masquerading as financial diligence. Forty-five to sixty days. Hundreds of pages of documents. Multiple requests for the same information. Underwriters who call you to ask questions already answered in your application. And at the end of it, a binary decision that still gets it wrong in both directions — approving borrowers who shouldn't qualify and rejecting borrowers who clearly should.

The reason this hasn't changed isn't that the industry lacks the incentive to improve. It's that building a better mortgage product requires solving for regulatory compliance, secondary market salability (Fannie/Freddie conforming standards), fraud detection, appraisal integration, and title coordination simultaneously. Most incumbents couldn't do it. So the dysfunction became the standard.

Tomo is the MetaProp portfolio company building what a mortgage company actually looks like if you start from scratch in 2024. AI-native from the ground up, Tomo is compressing timelines and improving accuracy by treating mortgage underwriting as a data orchestration problem rather than a document collection problem. The company has backed this with real results for real estate agents and home buyers — faster closings, fewer fall-throughs, better borrower experience — and has built its model specifically around the purchase mortgage market, where timing risk is highest and incumbent failure is most painful.

Company Spotlight

Tomo — AI-native mortgage company compressing timelines and improving accuracy by treating underwriting as a data orchestration problem. Built for the purchase mortgage market, where timing risk is highest.

The key insight: a better mortgage product isn't a feature. It's a structural advantage in a market where the average purchase mortgage takes 47 days and a 3-day improvement in close timeline meaningfully affects whether a deal survives.

Layer Three: Commercial Real Estate's Last Manual Frontier

If residential mortgage underwriting is inefficient, commercial real estate lending is a different category of problem entirely. A $35M construction loan for a multifamily development requires synthesizing project feasibility, developer track record, market absorption rates, construction cost risk, zoning and entitlement status, interest rate sensitivity, and lender concentration limits — often across multiple tranches and capital providers. There's no FICO score equivalent. The data is fragmented across dozens of sources. And the stakes of a bad decision are enormous.

This is why construction lending dried up after 2008 and never fully recovered in many markets. The analytical burden of underwriting complex CRE deals correctly — at the speed developers actually need capital — has been too high for most institutional lenders. Regional banks step in but carry concentration risk. Private lenders charge for the premium their speed commands.

CoFi is solving this at the commercial layer: AI-powered underwriting for construction and development lending that synthesizes the heterogeneous data inputs CRE credit requires and produces decisions faster and more defensibly than traditional processes allow. For a developer who's been waiting 90 days for a term sheet, the difference isn't marginal. It can determine whether a project happens at all.

What makes CoFi's position interesting is that it's not just competing on speed. A faster bad decision is still a bad decision. The value proposition is accuracy at scale — being able to process deal flow that would require 10 underwriters and produce consistent, data-backed credit assessments that hold up to secondary market scrutiny.

The Stack in Full

What Rent App, Tomo, and CoFi represent isn't three separate bets on mortgage tech. It's a thesis about what happens when AI finally touches the full real estate credit stack — from the renter building credit to the developer accessing construction capital.

The through-line is the same in each case: an underwriting process that was designed around the limits of human cognitive bandwidth and 1980s-era data infrastructure gets rebuilt around what's now computationally possible. More signals. Better synthesis. Faster decisions. More accurate outcomes.

The incumbent response will be predictable: "This is fine, but real estate credit requires relationship and judgment." That's true at the margins and increasingly irrelevant in the middle of the distribution. The bread-and-butter deals — the conforming purchase mortgage, the standard construction loan for a track-record developer, the apartment rental application — don't need a relationship. They need an accurate model. And an accurate model at scale beats an inaccurate relationship any day of the week.

The better-informed version of the incumbent concern is regulatory: real estate lending sits under a dense overlay of fair lending regulation, secondary market standards, and state licensing requirements that any AI system has to navigate. This is real. It's also a moat for the companies that solve it — and a temporary deterrent, not a permanent barrier.

What This Means for Founders and Investors

If you're building at any layer of the real estate credit stack right now, the questions that matter are: Where is the data currently fragmented? Where does human synthesis create bottlenecks and errors? Where does the cost of a credit decision — in time, in personnel, in fallout — significantly exceed what a well-trained model would cost?

These are not edge cases. They are the core of how capital flows through the built world.

The 2030 version of this industry looks different than the 2020 version in one specific way: the credit decision itself is no longer the bottleneck. Data aggregation, model architecture, regulatory compliance infrastructure — those are the hard problems now. The companies that solve them are building infrastructure for a $7T+ annual market.

Billy Beane's insight wasn't that data was more important than judgment. It was that better data produces better judgment. Real estate credit is finally catching up.

Founders building at any layer of this stack — my DMs are open.