AI in the venture capital pre-investment process: what 177 European firms told us

Venture capital performance is decided before the term sheet. The large-scale empirical work on how VC firms operate finds that pre-investment activities, sourcing, screening, and evaluation, influence outcomes more than later-stage structuring or post-investment support. Yet in most small and mid-sized funds, those activities still run on networks, pattern recognition, and partner intuition: the same survey research shows roughly 30 percent of deals come from professional networks, 20 percent from investor referrals, and only 10 percent from inbound, while screening under time pressure leans on heuristics that decision research has long shown to be bias-prone [Gompers, Gornall, Kaplan and Strebulaev, 2020, as reviewed in our study; Zacharakis and Shepherd, 2005].

In early 2025 we surveyed European venture firms to measure, rather than assert, what AI is actually doing inside that process. We received 177 responses from VC firms across 22 European countries, asking what they use AI for across deal scouting, deal screening, and due diligence, whether their tools were built in-house or acquired externally, whether AI supplements or substitutes human judgment, and how their reported performance has moved. This article presents what the data shows, what has happened in the year since we collected it, and what the honest limits of the evidence are.

One convention, established now and kept throughout: numbers from our survey are introduced as exactly that, and every external figure carries its source inline. The reader should never have to guess which is which.

A note on the data and its limits

The survey was distributed to 1,524 European VC firms identified through PitchBook, drawing 177 responses, an 11.6 percent response rate; descriptive statistics use the 142 fully complete responses. The sample is small, young, and early-stage: 83 percent of responding firms have ten or fewer employees, roughly 60 percent manage less than $50 million, three quarters were founded in the last decade, and the most common ticket size is $250,000 to $1 million. That composition is a feature for this article's readers, because it describes exactly the kind of fund most European VCs run, but it should be kept in mind when generalizing upward.

Three limits matter enough to state before any findings. First, the design is cross-sectional: everything was measured at one point in time, so the performance results are associations, not causal effects, and we cannot rule out that better-performing firms simply adopt more technology. Second, performance is self-reported: respondents rated the change in their average IRR over the past three years on an 11-point scale, and self-reports carry optimism and memory risks. Third, adoption is so recent that two thirds of all reported AI implementations were less than a year old at the time of the survey, while the performance question covers a three-year window, which biases against finding any relationship at all.

We state these limits prominently for a simple reason: no study anywhere, academic or commercial, has yet linked AI adoption to audited fund returns. Every performance claim in this field rests on backtests, perceptions, or self-reports. A controlled survey of 177 firms with its caveats stated is therefore among the best evidence currently available, and readers should treat anyone offering cleaner-sounding numbers with suspicion.

What the firms actually do

Adoption is broad, recent, and deepest where the work is structured

Among the 142 complete responses, 72.5 percent use AI somewhere in the pre-investment process: 59.2 percent in due diligence, 55.6 percent in deal screening, and 45.8 percent in deal scouting. Most users are not dabbling in a single phase; 55.6 percent of all respondents use AI in more than one stage. A quarter of firms, 27.5 percent, use none.

The phase pattern is informative. Due diligence and screening, the most structured and data-intensive stages, show the highest adoption shares, while scouting shows the lowest adoption but the highest intensity among its users: sourcing investment opportunities scored the highest mean extent of use of any activity we measured (2.89 on a five-point scale) and the earliest implementation. The simplest reading, and the one our data supports, is that AI adoption in venture tracks the structure of the task: where the work is processing volumes of comparable information, AI is already in the workflow; where the work is judgment about people, it largely is not.

The recency deserves emphasis. Of the 387 activity-level AI implementations respondents reported, about two thirds had been in place for less than a year, and only 16 predated the survey by more than two years. The survey is a snapshot of an adoption wave in progress, taken in the first months of 2025, not a portrait of a settled practice.

Independent work points the same direction on the phase pattern: a working paper analyzing AI in VC operations, contemporaneous with our survey, also finds screening the most prevalent application since 2022 Ronco and Barontini, 2025.

The tools are overwhelmingly external, and overwhelmingly language models

Two structural findings matter more than any adoption percentage.

First, the tooling is bought, not built. Across firm categories, roughly 80 percent of AI adoptions in our survey were externally developed solutions. The gradient runs on portfolio scale: firms managing four or more funds develop 42 percent of their AI in-house, while single-fund firms, the majority of the sample, build only 15 percent themselves. Small firms are not building data science teams; they are acquiring capabilities from outside, which is the only realistic route at their size, and it is the route our study's own practical implications recommend: smaller firms should consider externally developed tools to overcome internal capacity constraints, while larger firms can explore in-house development as a differentiator.

Second, AI is a colleague, not a committee member. Approximately 95 percent of firms using AI deploy it as a supplement to human decision-making rather than a substitute. The handful of substitution cases concentrate in screening, where six firms let models make filtering calls, but full delegation is rare everywhere. The dominant AI types tell the same story: large language models lead every stage (used by 78.5 percent of scouting adopters, 63.3 percent in screening, 61.9 percent in due diligence), with generative AI second and classic machine learning a distant third. Notably, dedicated predictive AI registered zero users in due diligence, and up to 11.9 percent of respondents answered "I don't know" when asked which AI types their firm uses. The technology in actual use is the general-purpose kind that structures information, not bespoke prediction engines.

Geography: the leaders are not who the stereotype suggests

Germany, Switzerland, and the UK stand out in our data: 84.6 percent of German respondents use AI in due diligence, Switzerland reaches 85.7 percent in due diligence and 71.4 percent in screening, and the UK 76.5 and 64.7 percent. Norway and Sweden are more moderate, around 35 to 45 percent across stages.

We would caution against the easy institutional explanation, partly because the country subsamples are small (Germany 13 firms, Switzerland 7), and partly because the broader data cuts against it. Economy-wide, the Nordics lead Europe in AI adoption, not lag it: 27.6 percent of Danish and 25.1 percent of Swedish enterprises used AI in 2024 against an EU average of 13.5 percent Eurostat, 2025, and Norway and Sweden rank essentially level with Germany on government AI readiness Oxford Insights, 2024. The more defensible explanation is the structure of the VC markets themselves. Nordic early-stage investing is unusually relationship-driven and local; platform data shows local investors providing about 59 percent of Norwegian early-stage funding since 2017 Dealroom, accessed 2026, and Norwegian VC dry powder fell from NOK 29 billion in 2018 to NOK 18 billion by late 2024 Chambers Venture Capital Guide, 2025. Smaller funds in tighter fundraising markets, sourcing through dense local networks, have had less slack to experiment, whatever their countries' general digital sophistication.

The leaders' context is its own story. UK venture firms operate inside a financial sector where AI is normal infrastructure: the regulator's own survey found 75 percent of UK financial services firms already using AI by late 2024, with a third of use cases bought as third-party implementations rather than built Bank of England and FCA, 2024, and the UK remains the world's third-largest VC market, with 9 billion pounds invested into venture-backed businesses in 2024 BVCA, 2025. Switzerland ranks first globally for AI talent density in Stanford's AI Index Stanford HAI, via coverage, 2026 and announced a deliberately light-touch, sector-specific approach to AI regulation in February 2025, explicitly to protect its position as an innovation location Swiss Federal Council, 2025. Germany has run a funded national AI strategy since 2018 European Commission AI Watch. Where the surrounding financial and policy infrastructure treats AI as standard equipment, venture firms appear to follow; where deal flow is local and relationship-priced, they have not needed to yet. Our country subsamples are too small to prove any of this, and we present it as context, not finding.

What is associated with performance

Before the survey numbers, the prior: why would structured evidence plausibly help at all? Because the research on unaided VC judgment is uncomfortable. Machine learning predictions on French administrative data indicate that VCs systematically back founders who fit familiar patterns, male, elite-educated, hub-based, beyond what those traits actually predict, while passing on observably stronger outsiders Lyonnet and Stern, 2022. A benchmarking study of more than 16,000 accelerator startups found that roughly half of the VC investments in the sample were predictably inferior to readily available alternatives using only information knowable at the time, at an estimated cost of roughly 1,000 basis points Davenport, 2022. Neither study is causal proof that tools fix the problem, but together they establish that there is a real, measurable problem for tools to address: selection error that better-structured evidence could plausibly reduce.

With that prior stated, the framing rules for our own findings are strict, and they are the finding's frame, not decoration: these are associations between AI use and self-reported performance in cross-sectional data. They are not proof that AI raises returns.

The simplest cut compares firms that use AI in each phase against firms that do not. AI adopters reported higher IRR scores in all three phases: a difference of +0.77 on the 11-point scale for scouting, +0.61 for screening, and +0.59 for due diligence, all statistically significant, with the scouting difference a medium-to-large effect. These comparisons carry no controls for firm characteristics, so they are best read as descriptive context: the kinds of firms that use AI also report better recent performance. They cannot say which way the arrow runs.

The headline performance finding is the controlled one. Once firm age, team size, fund count, assets under management, and ticket size are held constant, most phase-level effects dissolve, which we report as plainly as the positive results. What survives is screening: in ordered logit models, a one-point increase in the extent of AI use in deal screening is associated with roughly double the odds of being in a higher reported-IRR category (odds ratio approximately 2.0). When AI was implemented, as opposed to how much it is used, showed no relationship anywhere, in any model.

Drilling into the seven specific activities we measured, the associations concentrate exactly where AI's mechanical strengths meet market data. The extent of AI use in predicting firm performance during screening shows the most robust association: roughly 60 to 69 percent higher odds of a higher reported-IRR category per scale point, significant under a false-discovery correction and consistent across model specifications. Evaluation of risk profiles and identification of investment objectives follow, with odds ratios around 1.5. Exit potential analysis in due diligence shows a positive association in the linear models (+0.33 on the IRR scale per point of use) and a marginal one in the ordered models. Meanwhile the relationship-heavy activities, assessing founders and team dynamics, identifying syndication partners, show nothing at all, and neither does legal compliance work. One respondent's explanation for that last gap was blunt: "Legal compliance is too early a stage to trust AI with."

The pattern is coherent. The activities carrying the performance associations are the predictive, market-data-heavy ones, where the academic literature has long shown algorithms add discrimination: machine learning classifiers have demonstrated usable accuracy in startup outcome prediction for years Arroyo, Corea, Jiménez-Díaz and Recio-García, 2019, and an algorithm benchmarked against 111 VC professionals on 77,000 European companies outperformed the average investor on screening accuracy by about 29 percent Retterath, 2020. The activities showing no association are the ones where the inputs are human and the data is thin, which is also where our respondents say judgment must stay: as one put it, "ultimately you need the human element of evaluating founders."

The leveling finding: who actually benefits

The most consequential result for small funds comes from asking whose performance associations are strongest. The answer, consistently: the leanest and youngest firms making the smallest investments.

In subgroup analyses, firms with six to ten employees showed the largest effects in the study: AI use in exit potential analysis was associated with more than four times the odds of a higher reported-IRR category (odds ratio 4.41), and in predicting firm performance, 2.74. Firms writing $250,000 to $1 million tickets showed odds ratios of 2.52 for predicting firm performance and 2.13 for exit potential analysis. Firms founded between 2015 and 2019 showed 2.46 for predicting firm performance. These are small subsamples, between 24 and 43 firms per cell, corrected for multiple testing and robust to alternative groupings, but they should still be read as patterns, not point estimates. The robustness exercise across coarser clusters confirmed the core of it: the screening association persists for small-AUM firms, small-ticket firms, and medium-sized teams, precisely the firms with the least analytical bandwidth per deal.

Two readings of this pattern are available, and our study endorses the modest one. The ambitious reading is that AI levels the field between small and large funds. The modest reading, which the data supports, is that AI has leveling potential: small funds capture the largest marginal benefit because they start with the least process capacity, and a tool that structures evidence substitutes for analysts the fund does not have. Larger firms in our sample concentrate their AI in due diligence and build more in-house, suggesting the leveling is not yet realized, just available.

The economic backdrop makes this finding more pointed than it was when we collected the data. European VC fund raising deteriorated sharply through 2025, with first-time funds hitting a record low in 2024 (56 vehicles raising $3.7 billion) and the count of active European VC investors falling by more than half from its peak PitchBook, 2024-2025. European pension funds allocate 0.01 percent of assets to venture Atomico State of European Tech via Sifted, 2025. Small funds are being squeezed exactly when the evidence suggests they have the most to gain from systematizing their pre-investment work, and exactly when the cost of doing so has fallen.

The exit side deserves a specific note, because exit potential analysis was one of the two activities with the strongest small-fund associations and is among the least systematized. Exit conditions remain the binding constraint in venture: M&A accounted for 86.2 percent of exit count in the first three quarters of 2025 PitchBook via Fortune, 2025, and the cost of failing to reach an exit is observable in the secondary market, where venture fund stakes traded at 68 percent of net asset value in 2022 and 2023, recovering only to 78 percent by the first half of 2025 Jefferies, 2025. The academic literature adds that exit windows genuinely close, the likelihood of an IPO exit rises and then falls with time Giot and Schwienbacher, 2007, and that timing them well is a skill experienced investors demonstrably have and others lack Lerner, 1994. Mapping potential acquirers is hard in a way that favors systematic monitoring: only about a quarter to a third of majority-control tech acquisitions are made by serial acquirers, so most buyers are companies that appear on nobody's standard list Jin, Leccese and Wagman, 2025. A small fund that systematizes exit-landscape monitoring is automating one of the most consequential and least-resourced analyses it performs, which is consistent with where our data found the largest association.

What has happened since the data: mid-2025 to mid-2026

We closed data collection in April 2025. The year since has mostly strengthened the picture, and where it has not, we say so.

The capital and the flagship funds moved first. In the same month our survey closed, SignalFire, a fund built explicitly around its AI platform, raised over $1 billion, taking assets to roughly $3 billion Bloomberg, April 2025. A month later, QuantumLight, the London quantitative venture firm founded by Revolut's Nik Storonsky, closed a $250 million first fund and stated that every investment it had made was recommended by its model, a performance claim that is the firm's own and unaudited, but a fund-existence fact that is not QuantumLight via GlobeNewswire, May 2025. EQT Ventures' Motherbrain, the longest-running European experiment in data-driven sourcing, was profiled in production across sourcing, due diligence, and portfolio work in late 2025, with EQT's head of AI describing measurable sourcing advantages from a decade of accumulated internal data Tech.eu, November 2025. We searched for the opposite signal, funds abandoning or scaling back algorithmic approaches in this period, and found none; absence of evidence, but a notable absence.

The adjacent market measured its own acceleration. In M&A, the closest cousin to venture due diligence, Bain's practitioner surveys recorded generative AI use rising from 16 percent of practitioners in 2023 to 21 percent in early 2025 to 45 percent in 2025, with sourcing, screening, and diligence the leading use cases Bain, 2025; Bain, 2026. Deloitte's survey of 1,000 corporate and private equity leaders found 86 percent of organizations reporting generative AI somewhere in their M&A workflows by mid-2025, with about a third of adopters applying it to target screening and diligence Deloitte, 2025. In finance functions generally, a KPMG survey of over 1,000 finance leaders found active AI use reaching 75 percent by March 2026, up from 30 percent in 2024 KPMG, 2026.

The honest counter-signals from the same period belong in the same paragraph as the momentum. Gartner predicted in June 2025 that over 40 percent of agentic AI projects would be canceled by the end of 2027, and coined "agent washing" for vendors relabeling existing products Gartner, 2025. And the most precise counter-signal comes from a vendor survey we cite with that label: Affinity's late-2025 survey of roughly 300 private capital professionals found 85 percent using AI to automate daily tasks, up from 76 percent a year earlier, while the share saying they use AI to help make investment decisions fell from the prior year Affinity, 2025. Read precisely, the market pulled back from AI making decisions, not from AI structuring the work before decisions. That is not a contradiction of our findings; it is convergence on the same division of labor our respondents described, and it matches our finding that 95 percent of firms deploy AI as a supplement.

For European firms, the regulatory picture also moved: the EU AI Act's general-purpose model obligations took effect in August 2025, while the high-risk system rules were provisionally deferred to December 2027 Council of the EU, May 2026. Deal screening is not among the Act's listed high-risk categories for finance, which cover credit scoring and insurance pricing, so the direct compliance burden on VC analytics remains limited for now.

The objections, taken seriously

The strongest objection is now peer-reviewed, causal, and we feature it rather than bury it. Research published in the Review of Financial Studies, using plausibly exogenous variation in screening automation, finds that VCs adopting data-driven screening become better at evaluating startups that resemble historical data, tilt their portfolios toward that familiar pool, and become less likely to fund the innovative outliers that achieve rare, massive success Bonelli, 2025. Venture economics lives in that tail, so this is not a quibble; it is causal evidence that pure machine screening can optimize a fund into mediocrity. Our survey respondents intuited the same thing; one warned that if every firm runs similar models on similar data, "everyone would start looking at the same short list" of algorithmically approved startups. We read this finding constructively, because it is the strongest argument available for the architecture our data shows winning in practice: systems where algorithms structure evidence, surface and rank what humans cannot read at volume, and humans keep the outlier call. The 95 percent supplement finding is not a transitional stage on the way to automation. Bonelli's result is the reason it should not be.

"The associations might just be selection." Correct, and unresolvable with our design. Better firms may adopt more tools, rather than tools improving firms. The pattern in our data is at least consistent with a real effect where one would expect it: associations concentrate in data-heavy activities and in low-bandwidth firms, and are absent where AI has no plausible mechanism (founder assessment, syndication). But the design cannot prove direction, the timing mismatch between very recent adoption and a three-year performance window further muddies it, and we say so wherever the findings appear.

"Self-reported IRR is soft." Also correct. Perceived IRR change is vulnerable to optimism, and adopters might inflate their self-assessment. The structural answer is unsatisfying but true: nobody has better data. Fund-level audited returns linked to adoption practices do not exist in any study we could find, academic or commercial. Until someone runs a longitudinal study with verified returns, evidence in this field is perceptions, backtests, and self-reports; ours at least comes from a sizable, multi-country sample with the caveats attached.

"Algorithms still lose to the best humans." The peer-reviewed comparison supports a version of this: a machine learning algorithm beat the average returns of 255 business angels, but experienced angels who suppressed their cognitive biases still beat the algorithm Blohm, Antretter, Sirén, Grichnik and Wincent, 2022. The honest synthesis is that algorithms raise the floor, not the ceiling. For a two-partner fund without a research staff, raising the floor is most of the opportunity. The same literature documents why adoption lags the evidence: investors exhibit algorithm aversion, discounting machine input out of overconfidence and a preference for agency, which is a behavioral barrier, not an analytical one Davenport, 2022.

"Most VCs are not actually data-driven, whatever surveys say." True, and worth stating against our own 72.5 percent figure. Using an LLM in due diligence is not the same as running a systematic pipeline. Survey measures like ours capture any meaningful use; the population of funds with genuinely systematic, engineering-backed processes is far smaller. Both facts matter: shallow adoption is broad, deep adoption is thin, and the gap between the two is precisely where the competitive opportunity for small funds sits.

"Small firms could just build this themselves now." One of our own respondents made the point: building internal tooling has become cheap with modern AI coding tools and SaaS components. We take it seriously, and the do-it-yourself frontier has genuinely moved. The survey data nonetheless shows what firms actually choose: 85 percent of single-fund firms' AI is externally acquired, and in-house development concentrates in firms with four or more funds, the ones with the scale to sustain engineering. A weekend prototype and a maintained system that a fund bets its process on are different objects, and the maintenance economics, not the build cost, is what separates them. The respondent who asked for "tools for smaller VCs" was naming the actual gap.

The hybrid intelligence conclusion

Our study ends where the strongest external evidence also points. The firms in our data that report the best outcomes are not the ones automating decisions; they are the ones using AI most extensively to structure the evidence beneath decisions, in screening, in risk evaluation, in mapping exit landscapes, while keeping human judgment on founders, on outliers, and on the final call. One respondent's formulation serves as the thesis of the whole field: "the firms that effectively combine human intuition with intelligent automation will be best positioned." The causal evidence on outlier suppression makes that division of labor not a compromise but the design goal: the algorithm reads everything and forgets nothing; the human overrides it precisely where history is a bad guide.

For small funds, the practical sequence follows from the data: start where the associations are strongest and the bandwidth constraint binds hardest, screening-stage company evaluation and exit potential analysis; acquire rather than build, as 85 percent of single-fund firms already do; and treat the system as evidence infrastructure that updates continuously, not as an oracle. The leveling is potential, not automatic. It accrues to firms that do the organizational work of integrating structured evidence into how they already decide.

How this shapes the way we work

We build custom decision systems for investment firms, so this research is not neutral to us, and this section is the only openly first-person one. The survey shaped our practice before it shaped this article. We build systems that structure evidence and rank options, never systems that make investment decisions, because both our data and the causal literature say the human override is where venture returns live. We focus on the activities where the associations are strongest and the tooling gap is real: screening-stage evaluation and exit-landscape monitoring for funds too small to staff either. And we treat the limits of this evidence as a standing instruction: until someone links practice to audited returns, every claim we make to a client carries the same labels you have read here.