At some point the proposal arrives: a score on every company in the pipeline, updated nightly, that ranks inbound and discovered companies by how closely they resemble the firm's past successes. It might come from a data vendor, from a new hire who has built one before, or from an AI tool that already reads the firm's company records. The appeal is plain. The team sees more companies than it can evaluate, and a ranked list promises to put attention where it counts.
The decision is not whether the model is clever. It is what the model is trained to predict, what it will quietly deprioritize, and how the firm will notice if its pipeline starts to drift toward the familiar. I went looking for evidence before writing any of that down, and found a paper that measures the trade-off directly.
Source
What the study measured
Maxime Bonelli (London Business School) published "Data-Driven Investors" in The Review of Financial Studies (volume 39, issue 7, pages 1909 to 1969, 2026). The typeset article is behind a paywall. I read its published abstract and, in full, the author's manuscript dated September 23, 2024, which predates publication by about a year. The figures below come from that manuscript, and the page and table numbers refer to it; the published article may differ in detail.
The study covers 223,479 investments by 7,454 venture firms in 70,335 startups worldwide from 2000 to 2019, drawn from Crunchbase, with patents from the major patent offices and fund data from Preqin (page 12, Table 1). The average firm in the sample was 14 years old and had 11 employees at the time of an investment (page 12).
A firm counts as "data-driven" from the date it hired its first person whose job was to build data infrastructure or algorithms for screening. The author identified 247 such jobs from employee histories and checked them by hand, excluding people who advise portfolio companies on data (page 15). He notes the limit himself: a firm that outsources the work entirely, with no one on staff, would not be counted (page 16).
To measure what a model can learn from history, he built a "backward-similarity" score: how closely a startup's business description matches those of startups funded in the same industry over the previous five years (page 18). As a check, pioneers score low: Facebook and Myspace rank at 0.13 and 0.12 on a scale of 0 to 1, while later entrants such as Snap rank at 0.56 (Internet Appendix B.3).
Source
What it found
Adoption was rising. The share of firms classified as data-driven went from close to zero in 2000 to more than 4 percent in 2020, and those firms took part in 20 percent of funding rounds in 2020, against about 5 percent before 2010 (page 16, Figure 1). Firms with more early-stage deals were more likely to adopt; prior experience and past IPO success were not significant (page 17, Table 2).
Adopters scaled. After the first data hire, a firm's yearly investments rose about 30 percent relative to other firms, roughly 1.2 more deals a year on an average of 4.2, and assets under management rose by about the same share. The increase showed up in the second year (pages 21 and 22, Table 4, Figure 3).
The new deals looked like old ones. Investments in the most backward-similar startups, the top quarter, almost doubled; investments in the least similar quarter did not change (page 22, Table 5).
Screening improved where history was informative. Adopters became 2.5 percentage points more likely to back a startup that survived and raised a follow-on round, a 5 percent increase on a base of 52 percent. The gain was concentrated in the most similar startups, at 4.2 points, with no significant change elsewhere (page 24, Table 6). That matters because similar startups are, on average, less likely to succeed (page 19, Figure 2, Table 3): the models were picking the better ones from a weaker pool.
Rare successes became rarer. Adopters became 2.5 points less likely to back a startup that went public or was acquired for more than it raised, a 17 percent drop against a base of 15 percent. That drop appeared across all levels of similarity, and held when only IPOs were counted (page 25, Table 6, Table A.7). Their portfolio companies also filed about 6 percent fewer patents and drew fewer citations, mostly because of the shift toward similar startups (page 25, Table A.8).
Fund returns did not visibly change. On the smaller set of funds that report performance, net IRR and net multiple showed no significant difference after adoption (page 26, Table A.9). The author reads this as firms scaling their assets without a measurable loss in performance, which would increase fee income (page 26).
Adopters did not reach further afield. They became less likely to invest outside their home regions, no more likely to back serial founders or top-school graduates, and no more likely to take board seats (page 26, Table A.10). They also took part in fewer of their companies' later rounds (Internet Appendix C.5).
The pattern appears causal. Because firms choose to adopt, the author uses the 2006 launch of Amazon Web Services, which set off a boom in new startups in cloud-exposed industries and raised the screening load for firms that invested there. Firms most exposed were more likely to hire data staff, shifted toward similar startups, got better at follow-on prediction only in the most similar group, and became 8 points less likely to back a major success, a 38 percent drop (pages 32 and 33, Tables 7 to 9).
Describe this figure
A horizontal bar chart of changes in likelihood, in percentage points, on a linear scale from minus 10 to plus 6 with zero marked. Survived and raised a follow-on round, the common outcome, on a base of 52 percent: all adopters became 2.5 points more likely to back such a startup, a 5 percent increase; among the most similar startups, the top quarter by backward similarity, the gain was 4.2 points, with no significant change elsewhere. Went public or was acquired for more than it raised, the rare major success, on a base of 15 percent: all adopters became 2.5 points less likely to back one, a 17 percent drop, the highlighted bar; in the author's causal test, firms most exposed to the startup boom after the 2006 launch of Amazon Web Services became 8 points less likely, a 38 percent drop. The study covers 223,479 investments from 2000 to 2019; the figures come from the author's September 2024 manuscript, pages 24, 25 and 33, and the published article may differ in detail.
Synthesis
What the study does not tell an investor
It measures firms that built. The classification rests on hiring data staff, so the evidence speaks most directly to a firm that develops its own screening tools. A score bought from a vendor and used without in-house staff is outside what the paper counted.
It stops before today's tools. The data run through 2021 (page 11). The algorithms in question learned patterns from structured histories of past startups. Whether a general-purpose language model reading pitch decks behaves the same way is a separate question, and I return to it below.
It does not settle returns. The performance result is a null on a self-reported subset. It does not show that data-driven screening helps or hurts a fund, only that the study could not detect a difference.
Synthesis
What that means for the firm's score
Name the outcome the model predicts. The paper's central finding is that models get good at outcomes that are common in the history they learn from and poor at outcomes that are rare. A follow-on round is common; a company that returns the fund is not. A vendor or builder should be able to say which one the score is trained on. If the answer is "raised a next round," the score is a triage tool, and the firm should call it that.
Use it to set aside, not to select. The author notes that practitioners describe the most useful function of these tools as eliminating the least promising companies (page 10). That fits the evidence. A score is on firmer ground saying "this one does not warrant a first call" than saying "this one is the next great company."
Show the similarity next to the score. The drift the paper found was a change in what firms looked at, not a failure of prediction. The firm can see the same thing coming if each record shows how much a company resembles what the firm already backs, and if the pipeline report counts how many first meetings came from companies unlike the rest. A falling count is the early warning.
Keep a lane the model does not rank. The companies the models missed were the ones with little history to compare against. I would reserve a set share of first meetings for companies the score cannot place, and record them separately so the firm can see what that lane produces over time.
Remember whose history it is. A score trained on the firm's own company records learns the firm's past choices. A small firm has seen few rare successes, which is exactly the outcome a model struggles to learn. The larger datasets a vendor holds carry the same imbalance at scale. Either way, the score inherits a history dominated by common outcomes.
Weigh the scale benefit honestly. The gain the paper found most clearly was volume: more deals and more assets under management. A small firm that runs a concentrated portfolio and does not plan to double its deal count gets less of that benefit and keeps all of the drift risk. That is a reason to use the score narrowly, not a reason to refuse it.
Proposal
How I would decide and prove it
Start with a lookback on the firm's own history. Take the companies the firm met over the past three years and score them as the model would have at the time. Then check two things: whether the score would have ranked the firm's best companies near the top, and how many of the companies the firm now values most would have fallen below a cutoff. If the answer to the second is "several," the cutoff is too aggressive or the score is measuring the wrong thing.
If a vendor supplies the score, ask four questions before trusting it. What outcome is it trained to predict? What history is it trained on, and does that history include companies that never raised money? Can the firm see why a company scored as it did? How does the score treat a company with no close comparables: low, or unknown? A score that treats "nothing like it" as "not promising" is the pattern the paper warns about.
Place it in the workflow as a filter with a visible exception. The score orders the queue; it does not remove companies from it. The record shows the score, the similarity, the date, and the data it used. The reserved lane has an owner. The firm reviews both each quarter alongside the rest of the pipeline, and decides again at each annual review whether the score has earned its place.
Proposal
What I would test next
I would run the lookback before signing anything. A few hours with the past three years of first meetings will show whether the score would have caught the companies the firm cares about, and whether it would have pushed them down. That result decides whether the firm needs the score at all, and if it does, how wide the unranked lane should be.
Open question
Do language models inherit the same bias?
The study's models learned from structured histories of funded startups. The tools most small firms are now offered read pitch decks and websites with a general-purpose language model, which has absorbed far more than any one database. It could be better at seeing what is new, or it could be even more drawn to what sounds familiar, since resemblance to past descriptions is close to how such models work. I do not know which. The firm's own lookback, repeated each year and split by how novel each company was, is the cheapest way I know to find out.