The FICO Heresy

Ask a senior loan officer in 1995 whether a computer could reliably decide who got a mortgage and you would have gotten one of two answers. The polite version was that an algorithm could help with the paperwork. The honest version was that credit was a craft of reading people across a desk, the kind of thing you could only learn by sitting with thousands of borrowers and developing a feel for them. Either answer carried the same unstated conviction: the part of the job that actually mattered, the judgment, was irreducibly human.

That summer Freddie Mac issued an industry letter recommending that lenders use the FICO bureau score as a supplement to manual underwriting. Fannie Mae followed soon after. Both published score thresholds: below 620, heightened review; above 660, streamlined. A year after that the National Credit Union Administration formalized the compromise in writing. Credit unions could let an algorithm approve loans, but humans were still needed “to make discretionary lending decisions on credit applications that the System’s preset lending criteria would deny.” In 1996 a quarter of mortgage lenders used automated underwriting. By 2000 the great majority of mortgage lenders used automated underwriting.

The loan officer did not vanish. The job changed. The middle of the distribution, the cleanly scorable borrowers, collapsed into a transaction. Human discretion got pushed to the edges: thin-file applicants, the self-employed, the cases the model would have denied that documentation suggested ought to be approved. Discretion was exiled to the tail.

Credit scoring is akin to the VC Analyst’s story. Autopilot is the VC Partner’s.

For most of commercial aviation’s history, pilots controlled their airplanes throughout the entire flight. But by 2018, a Cranfield University analysis of fourteen thousand flights found that in a typical airline flight the pilots had their hands on the controls for somewhere between three and seven minutes. Everything in between was now automation, with the human in the seat as monitor. This change did not happen because anyone took a vote. It happened one expanding capability at a time, against pilot pushback at every step. The earliest serious academic warning, by Earl Wiener and Renwick Curry at NASA in 1980, said plainly that “flight-deck automation may have already passed its optimum point.” Lisanne Bainbridge wrote in 1983 about what she called the ironies of automation: the role of monitor is the one humans are worst at, and the one most likely to atrophy the skills you need when monitoring fails. The pilots’ unions have repeated some version of this argument every decade since.

They were right at the margins and they lost on the trajectory. Commercial flying is the safest mode of travel in history, and much of the gain came from systems pilots once distrusted.

Today’s venture discourse has only begun to engage these patterns directly. When voices do surface, they tend to fall into one of two camps. On the skeptic side, the argument is recognizable from any of the historical analogs. One seed-stage GP recently put it plainly: “We maintain a very manual investment process, largely out of respect for founders and reverence for the difficulty of the job. Early stage investment decisions contain a tremendous amount of nuance, which AI is not well suited to manage today.” Another voice in the same camp puts the case more sharply: “AI is not a replacement for domain expertise; it is a screen for volume.” That is the 1995 loan officer’s argument in 2026 venture vocabulary. On the other side, VCs could argue that investors who dismiss AI-driven software are like the CIOs who once refused to move to the cloud. It is the comparison those CIOs never made about themselves.

The reasonable peer pushback to all of this is that venture is a power-law business. The tail is everything, the middle is noise, and so even if an algorithm compresses the middle of the funnel into a transaction, the partner still holds the value where it actually accrues. Sure, I’ll concede the first half. The tail is where the returns live. But the firm itself compresses anyway. Cost, headcount, how a deal reaches a partner, what a partner actually does day to day, all of it changes shape if and when eighty+ percent of the funnel is handled by an algorithm. A firm built around that shape looks structurally different from the firm we run today, even if the final twenty percent of decisions remain in human hands.

However, these two analogies break in one place that matters. A loan loss is bad. An airplane crash is catastrophic. Both FICO and autopilot are systems tuned to keep the downside from happening, and the algorithm wins because it is reliable enough to do that at scale. Venture runs the other direction. A bad investment loses one check; a missed great one is a fund-returning hole that nothing else fills. The catastrophic error is not the wrong yes, it is the missed yes. The tail in lending and aviation is the rare disaster you build the system to prevent. The tail in venture is the rare outcome the whole business exists to catch. Same word, opposite stakes. Which makes the question of what survives in human hands even sharper than the analogs suggest.

So we must ask the same question, which three to seven minutes of a VC Partner’s work actually generate the returns?