For years I filed it under near miss. It was not one. A near miss needs a meeting, a judgment, and a wrong answer, and none of that happened. The founder never reached my calendar, so my judgment never ran. What failed was coverage, which is two things: finding the founder, and finding the hour to meet him. Here only the first happened. If a founder a colleague had flagged, in my own room and my own focus category, could fall through, the far larger field I never hear about fares worse.

Judgment is venture’s second step, the one everyone argues about when they argue about AI. But first, do the arithmetic on coverage.

Start with the pool. The databases that track startups for a living now hold tens of millions of companies apiece: Harmonic lists more than thirty million, Specter more than fifty. That is the standing stock of companies in the world, most of it noise, and it is now a single query away. What matters in a given year is the flow. In the United States alone, about fifty-five thousand companies took angel money in 2023, and roughly three times that many pitched an angel and got nothing.1 Only a fraction of the funded ones ever reach an institutional round. Across the whole country, about fifteen thousand venture deals of every stage got done in 2024.2

Now the VC coverage. A prototypical busy partner at a generic example firm takes about five first meetings a week, which after travel and vacation is something like two hundred a year, and makes about three investments. Four partners see maybe eight hundred companies between them and write twelve checks. Widen it to the whole pyramid, the associates and analysts who live on the front line and screen more founders than any partner does, and a four-partner firm with six juniors beneath it might push its coverage to three thousand first meetings a year.3 Three thousand, against fifty-five thousand that took angel money, in a year the whole industry does fifteen thousand deals. The firm still backs twelve. Not all fifty-five thousand are venture candidates, and a firm’s meetings rightly skew toward the ones that go on to raise institutional money: call it two thousand of the three thousand. Set those against the ten thousand or so seed and early-stage rounds the industry funds in a year and the firm’s reach looks better, about a fifth. The twelve checks do not get any bigger. And that is the best case in our example scenario, every seat firing at once.

We call it judgment. Mostly it is judgment applied to the few hundred companies that happened to find our calendar.

Widening with people only goes so far. Every first meeting a junior takes in a partner’s place is a lossy copy of the judgment it is supposed to feed, which puts the Copy Problem I’ve written about earlier front and center.

Accelerators show where it stops. Y Combinator winnowed a recent batch from twenty-seven thousand applications, accepting under one percent. The largest venture firm in the country runs its own accelerator that considered “over 19,000 startup pitches” for a single cohort.4 Nineteen thousand applications, read in a single application window by a small team that also has an accelerator to run. At even ten honest minutes apiece, that is over three thousand hours of reading. That is a full-time reader for a year and a half, which no firm spends on its front door, not even one with many hundreds of people.5 Something is making the first cut on those nineteen thousand applications before a human person gives any of them a fair look.

The funnel is no kinder on the way out, which is where it really matters. Of the roughly fifteen thousand deals the industry did in 2024, about fifty-five hundred were pre-seed or seed rounds.6 About a quarter of a seed cohort raises a Series A within two years in a normal year, and for companies seeded in 2022 it was closer to seventeen percent.7 More get there given longer, but two years is about what a seed round buys. That year’s other deals, other companies’ rounds and not this cohort’s survivors, ran to about forty-seven hundred early-stage, forty-two hundred later-stage, and under eight hundred venture growth.8

About a hundred and ten American companies a year crossed the billion-dollar unicorn mark over the last four years,9 and a seed cohort followed for most of a decade sends about one in a hundred to that mark. In round numbers, and with the cohorts not the same year: fifty-five hundred seed rounds in, a hundred-odd unicorns and maybe a dozen decacorns out.10, 11 Of course, it’s well known that the dollars concentrate too: about six percent of investments produce sixty percent of the returns.12 That is what makes a coverage hole expensive: the costliest event in venture is the outlier that never reaches your calendar at all. The founder I walked past was a hole in my coverage, and no amount of judgment makes whole a meeting that never happened.

The same arbitrariness runs the other way. One of the best investments I have made in years arrived as a cold email with the subject line “Texas boy from MIT.” I opened it because the subject line was charming, and that is close to the whole reason. The company has gone from our pre-seed to a seed and then a Series A, tracking toward a real outcome, and the chain traces back to four words that caught my eye on the right morning. A slightly duller one and that founder is someone else’s win, or no one’s. The chance to judge is a lottery in both directions. My judgment did its work on that one only because four words bought it a turn.

For twenty years this was simply the water we VCs swam in. Coverage was scarce at both ends. Finding founders took a network and meeting them took hours, so the firms with the widest networks and the loudest brands held a real edge at the first. That edge is being repriced. A VC investor can now point AI-powered data-gathering software at the entire legible universe of companies, everything that has left a trace, and pull back whatever matches a thesis. Andre Retterath wrote it plainly back in 2020: “Sourcing will become less of a differentiator as investors will gradually leverage the same identification web crawlers.”

For as long as the industry has existed, a large share of what looked like judgment was really coverage wearing judgment’s clothes. The firm that saw the deal at all looked shrewd next to the firm that never did, and it banked the difference as skill. Hand every firm the legible field and that borrowed credit is repriced toward nothing. Venture has always claimed to be paid for three things: finding the company, judging it, and being let into the round. Sourcing was the first defense, and it is the first one AI takes away. The hours to meet the founders the software finds are still rationed.

Whether a machine can judge a founder as well as you can is the question everyone is arguing about, and it is the second question. The first is how many founders you were ever in a position to meet at all.

    1. Jeffrey Sohl, “The Angel Market in 2023: An Inflection Point for Women Angels?”, Center for Venture Research, University of New Hampshire, June 20, 2024: 54,735 entrepreneurial ventures received angel funding in 2023 (down 12.2% from 2022), $18.6B in total, from 422,350 active angels, at an average deal of about $339,000. The same report puts the angel “yield rate” at 24.2%, so on the order of 225,000 ventures pitched angels that year. It is a survey extrapolation (13% response rate) spanning every sector angels touch, and the yield rate is per opportunity presented rather than per venture, so the count of distinct companies pitching is somewhat lower. Most of these companies never appear in PitchBook: the gap between 55,000 angel-funded and 5,500 tracked seed rounds is angel and friends-and-family rounds that never became an institutional round, and companies that died before one. The database totals cited above (Harmonic 30M+, Specter 55M+) are global, all-time cumulative figures: the standing pool, not the annual US flow. ↩
    2. About 15,260 US venture deals were completed in 2024 per the PitchBook-NVCA Q4 2024 Venture Monitor, since revised to 15,250 in the NVCA 2026 Yearbook. That counts deals, not unique companies, since a company can raise more than once in a year, so the distinct-company figure is somewhat lower and the number raising for the first time (new entrants to the VC funnel) is a few thousand. Every version of the comparison leaves the firm’s twelve checks a rounding error. ↩
    3. An illustrative construction, not a cited ratio. Firm structures vary widely, but a mid-size early-stage firm with four partners might carry six junior investors (principals, associates, analysts), who screen more first meetings than partners do. At roughly two hundred first meetings per partner (eight hundred for the four) and three to four hundred per junior (about twenty-one hundred for the six), total firm coverage lands near three thousand first meetings a year, and at three investments per partner the firm writes twelve checks. The exact number is not the point; every plausible version is a small fraction of the field. ↩
    4. a16z Speedrun, “What We Look For In Applications,” speedrun.substack.com (SR006 cohort, January 2026). Speedrun does not publish a review-team headcount or the length of its review window; “a small team” refers to its named leadership. a16z has not stated that it automates triage. The impossibility of hand-review is arithmetic, and the conclusion is the reader’s to draw. ↩
    5. a16z does not publish a role-segmented headcount, so “many hundreds” is an estimate rather than a live tally. The firm describes itself as mostly a platform and operating organization (a platform team widely reported at 150 or more people, plus separate talent, policy, comms, and finance groups), with a much smaller capital-deploying investment side. Dividing the reading among more people does not change the bill: roughly 3,200 hours, inside a single application window. ↩
    6. PitchBook-NVCA count 5,510 US pre-seed and seed deals in 2024 (5,049 in 2025) against 15,250 total US venture deals (NVCA 2026 Yearbook, data as of Dec 31 2025). PitchBook’s headline counts are estimates that include deals it expects to surface later; the reported-only figure runs about nine percent lower, with the gap concentrated at seed. ↩
    7. Carta, from its cap-table data: 30.6% of companies that raised a seed round in Q1 2018 reached a Series A within two years; 15.4% of the Q1 2022 cohort did (VC Fund Performance, Q1 2024); Carta’s Feb 2025 update puts the 2022 cohort at “17% or so” and calls 2018 “something like 25-30%.” Its March 2026 benchmark treats 25% at two years as median and 15% as low. ↩
    8. PitchBook-NVCA stage buckets for 2024 (NVCA 2026 Yearbook): 5,510 pre-seed and seed, 4,711 early-stage (Series A and B), 4,200 later-stage, and 786 venture growth, which with a 43-deal residual between the stage table and the headline count make the year’s 15,250. These are counts of rounds done that year, not survivors of the seed cohort above. Within the early-stage bucket, the rounds that carry both an explicit series label and a disclosed value number 1,729 Series A and 767 Series B (1,655 and 731 in 2025), about half the bucket. The Series B step is steeper than the A step: Carta data from October 2024 has 25% of companies that raised a Series A in Q3 2018 reaching a B within two years, and 9% of the Q3 2022 cohort. Longer windows are kinder; CB Insights’ 2008 to 2010 seed cohort found 48% raised a second round and 63% of those a third, so roughly three in ten of the original cohort reached a third round over eight-plus years. ↩
    9. New US companies reaching a $1B+ valuation, per PitchBook-NVCA (NVCA 2026 Yearbook): 85 (2019), 101 (2020), 356 (2021), 199 (2022), 51 (2023), 70 (2024), 124 (2025). The 2022 to 2025 average is 111 a year, across a wide spread; 2021’s 356 sits outside that window and would have pulled it far higher. Crunchbase’s independent US count for 2024 was 65, within eight percent. Unicorn counts get revised upward for a year or more after the fact as valuations surface, so treat any current-year figure as provisional. ↩
    10. CB Insights tracked more than 1,100 US tech companies that raised a first seed round in 2008, 2009, or 2010 through August 2018: 1.07% (twelve companies) reached a $1B valuation. AngelList’s own 2013 to 2019 seed data puts it nearer 2.5%, on a deal set that skews upward. A cohort rate and a calendar-year ratio are different things, and the hundred-odd unicorns a year run above what one percent of a 5,500-round seed cohort would produce: today’s unicorns come out of the far larger 2020 to 2022 seed vintages (7,405 and 7,005 rounds in 2021 and 2022), some never had a tracked seed round at all, and 2021-era marks minted more of them than the 2008 to 2010 vintage’s rate implies. More generally: the unicorns minted in a given year were mostly seeded five to nine years earlier, so a year’s unicorn count divided by a year’s seed count is the right order of magnitude, not a probability. ↩
    11. Crunchbase counted 24 companies newly valued at $10B+ worldwide in 2025 (30 in 2021 through November, 15 in 2020, five in 2019); it does not publish a US split, and applying the US share of new unicorns (56 to 59 percent in 2024 to 2026) gives thirteen or fourteen, hence “maybe a dozen.” No one publishes a seed-to-$10B rate. Crunchbase counted 82 decacorns at the end of 2025, roughly one in twenty of the private unicorns on its board (the board is a live list, 1,640 to 1,880 depending on the date); applied to a one-percent unicorn rate that is about one seed round in two thousand, a construction rather than a measurement. The 2008 to 2010 cohort above did far better (Uber, Airbnb, Stripe), which says more about that vintage than about the odds. ↩
    12. Horsley Bridge data across hundreds of funds since 1985, as reported by a16z: about 6% of investments, representing 4.5% of dollars, generated about 60% of total returns. Cambridge Associates’ study of the top 100 value-creating deals per year (1995 to 2012) found deals returning 5x or more accounted for 85% of the gains within that top-100 set. Both datasets are a decade old; nobody has published a post-2020 equivalent. ↩