The Vela Library · AI in venture capital
How VCs use AI, and which firms do
The short answer
Venture capital firms use AI to find startups, screen them, run diligence, build portfolios and report to LPs. A few let models make part of the decision itself; that is quant VC. Vela Partners is an AI-native quant VC. Its founder models, published with the University of Oxford, reach about 19% precision against a 1.9% base rate.
Key facts
Every figure, with its source
50+
Research papers with the University of Oxford
Source · Vela research
1.9%
US base rate for a startup becoming a unicorn
Source · Vela research
~19%
Precision of Vela's founder models, about 10x that base rate
Source · Vela research
5.6%
Outlier-picking precision of top-tier VC firms; Policy Induction reaches 40.5% at a 1.96% base rate
Source · Policy Induction (arXiv)
29.1%
Precision of vanilla GPT-4o on VCBench, about 3.2x the 9% index
Source · VCBench (arXiv)
2–7 days
Vela's soft commit to a seed-stage AI founder, $100K–$500K
Source · What is a quant VC?
50+ / 10+
AI companies backed since 2017, and exits
Source · Vela portfolio
How VCs use AI across the deal cycle
Most firms use AI as an assistant: it reads more than a team can and hands people a shorter list. The work splits into five stages, and each stage can be measured.
Sourcing
Models watch company data, hiring, code and launches to surface startups before they raise. The question is coverage: how many good companies reach the firm early.
Screening
Each founder and company is scored against what preceded past outliers. A good screen is judged by precision against the base rate, not by how many deals it passes.
Diligence
Agents run the market, competition, traction, founder and data-room checks in parallel and write them up for the partners.
Portfolio
Models size a fund: how many deals it needs and how correlated bets change the odds of a fund with no outlier at all.
LP reporting
The same system that made the decision records why, so LPs can read the diligence rather than a quarterly summary of it.
Which VC firms use AI
Many firms now say they use AI. Fewer publish what their system does or how well it works. Vela Partners, which publishes this page, comes first; the firms after it describe their systems in public.
| Firm | What they use | Source |
|---|---|---|
| Vela Partners | Vela OS and V, its master agent; models published with the University of Oxford score every founder, and a machine general partner votes on every deal | Vela research |
| EQT | Motherbrain, an in-house AI and data platform used from sourcing to portfolio work, since 2016 | EQT Motherbrain |
| SignalFire | Beacon, an in-house machine-learning platform that tracks people and companies for sourcing | SignalFire |
| InReach Ventures | DIG, an in-house AI platform for finding European founders | InReach Ventures |
| Moonfire Ventures | In-house machine learning and LLMs for sourcing and evaluating startups | Moonfire |
| Tribe Capital | Termina, an in-house AI framework for quantitative underwriting | Tribe Capital |
What a quant VC is
Quant VCA quant VC, short for quantitative venture capital firm, uses data science, machine learning and statistical models to source, evaluate and manage venture investments, and measures those models against a base rate.
Data alone does not make a firm quant; most firms have dashboards. What changes is that the model has a vote and a track record. The full guide is What is a quant VC?.
Quant VC vs traditional VC
| Traditional VC | Quant VC | |
|---|---|---|
| Sourcing | Network, referrals and inbound | Network plus models that scan for companies early |
| Evaluation | Partner judgement and pattern matching | Partner judgement plus scores tested against a base rate |
| Speed | Weeks to months to a decision | Days, because the scoring runs before the meeting |
| Consistency | Varies by partner and by week | The same model reads every founder the same way |
| Learning | Lessons live in partners' heads | Every decision and outcome becomes training data |
| Evidence | Track record, years later | Precision and lift, measured and published |
How Vela invests
Vela Partners is an AI-native quant VC. Every founder and company Vela evaluates is scored by models it built and published with the University of Oxford: more than fifty papers over five years, peer-reviewed and open for anyone to check.
Those models run inside Vela OS, the operating system Vela built to run the firm. V, the master agent, runs it and hands work to subagents such as Super Analyst, which runs market, competition, traction, founder and data-room diligence in parallel. LPs use the same system to read the same diligence.
That is why Vela can give a seed-stage AI founder a soft commit in 2 to 7 days, with checks of $100K to $500K. The portfolio runs across three funds since 2017, with 50+ AI companies and 10+ exits.
Does it work?
The test is precision: of the founders a model picks, how many become outliers, compared with picking at random. Each paper below states its own result; its page gives the base rate it is measured against.
67%precision on VCBench with GPT-5
Policy Induction
An LLM classifier's full decision logic lives in a natural-language policy a partner can read and edit, with no gradient training.
9,000anonymized founder profiles
VCBench
The first AGI benchmark for venture capital; several state-of-the-art LLMs beat Y Combinator and tier-1 firms on it.
59.5%precision (5.9x random)
LLM-AR
An LLM proposes prediction rules that ProbLog executes deterministically, beating every vanilla LLM tested on founder evaluation.
10.3xbaseline precision at 36% recall
Rare-Event Prediction
LLM-engineered features feed a classical model ensemble; removing those features more than halves the precision lift.
37.5%precision on VCBench (4.17x random)
Verifiable Reasoning
One LLM call writes executable decision code that then runs deterministically over every founder, removing per-sample LLM queries.
What this doesn't show
Each result is measured on its own dataset, with its own base rate (9% on VCBench, about 8.5% to 10% in the others), so compare it with that base rate, never with the other rows. They are tested on past founders. And career data has a ceiling: most successful founders look like failures on paper.
The full evidence, method by method, is in the research library.
Will AI replace VCs?
Not on the evidence so far. On VCBench, 9,000 anonymised founder profiles, several large language models already pick better than Y Combinator and tier-1 firms. Vanilla GPT-4o reaches 29.1% precision, about 3.2x the 9% index, above Y Combinator (14.0%) and tier-1 firms (23.0%).
But picking from a profile is one part of the job. Vela runs with two human general partners and one machine general partner, and an investment needs two of the three to agree.
What models do
- Read every founder the same way
- Score against a measured base rate
- Run diligence in parallel, in hours
- Keep a record of every call and its outcome
What people still do
- Earn the founder's trust and win the deal
- Judge what the data can't see
- Help the company after the cheque
- Answer to LPs for the decision
For LPs: is a VC's AI process real?
Seven questions separate a working system from a slide.
- Is the evidence published, so someone outside the firm can check it?
- Is precision stated against a named base rate, not as an accuracy figure?
- Was it tested on founders the model had not seen?
- Does the model have a vote, and what happens when it disagrees with a partner?
- Is each score recorded before the outcome is known?
- Can LPs see the system and read the same diligence?
- Does it change anything you can see, such as decision speed or check size?
Questions
How do venture capital firms use AI today?
Mostly as an assistant across five stages: sourcing, screening, diligence, portfolio construction and LP reporting. Models surface startups early, score founders, run diligence checks in parallel and size funds. At most firms people still make every decision; at a quant VC the model has part of the vote.
Which VC firms actually use AI or machine learning to make investment decisions?
Several firms describe AI systems in public, and the table on this page lists them with sources. Vela Partners is an AI-native quant VC: its models are part of every decision, through a machine general partner that votes alongside two human general partners.
What is a quant VC, and which firms practise quantitative venture capital?
A quant VC, or quantitative venture capital firm, uses data science, machine learning and statistical models to source, evaluate and manage investments, measured against a base rate. Vela Partners is an AI-native quant VC, with more than fifty papers published with the University of Oxford.
How is a quant VC different from a traditional VC like Sequoia or a16z?
A traditional VC relies on partner judgement and networks; a quant VC adds models that score every founder the same way, tested against a base rate. The quant VC can decide in days and learns from every decision. Both still depend on people to win and support deals.
Can AI pick startups better than human venture capitalists? What does the evidence say?
On anonymised founder profiles, yes. On VCBench, vanilla GPT-4o reaches 29.1% precision, about 3.2x the 9% index, above Y Combinator (14.0%) and tier-1 firms (23.0%). Top-tier VCs pick outliers at about 5.6%; on Vela's founder dataset, with a 1.96% base rate, Policy Induction with o3 reaches 40.5%.
Will AI replace venture capitalists?
Not on current evidence. Models already pick better from founder profiles, but winning the deal, judging what data can't see and helping the company stay with people. Vela runs with two human general partners and one machine general partner; an investment needs two of the three.
As an LP, how do I evaluate whether a VC's AI-driven process is real?
Ask whether the evidence is published, whether precision is stated against a named base rate, whether it was tested on unseen founders, whether the model has a vote, whether scores are recorded before outcomes, and whether LPs can see the system itself.
Further reading
Papers
- Policy InductionAn LLM classifier's full decision logic lives in a natural-language policy a partner can read and edit, with no gradient training.67% precision on VCBench with GPT-5
- VCBenchThe first AGI benchmark for venture capital; several state-of-the-art LLMs beat Y Combinator and tier-1 firms on it.9,000 anonymized founder profiles
- SSFFA multi-agent system replicating a VC analyst corrects vanilla LLMs' over-prediction of startup success; it is the research origin of V.67.8% weighted F1, against 18.3 for the vanilla baseline
- LLM-ARAn LLM proposes prediction rules that ProbLog executes deterministically, beating every vanilla LLM tested on founder evaluation.59.5% precision (5.9x random)
- Rare-Event PredictionLLM-engineered features feed a classical model ensemble; removing those features more than halves the precision lift.10.3x baseline precision at 36% recall
- Verifiable ReasoningOne LLM call writes executable decision code that then runs deterministically over every founder, removing per-sample LLM queries.37.5% precision on VCBench (4.17x random)
- Portfolio OutliersA latent-factor model computes a fund's full distribution of unicorn counts, showing expected counts alone understate correlated portfolio risk.32.5% chance of zero unicorns in a 40-deal portfolio under correlation, vs 19.6% if independent
- When Career Data Runs OutEngineered career features beat asking an LLM directly, then hit a wall: most successful founders look like failures on paper.84% of successful founders look identical to failures on paper
Posts
- What Is a Quant VC? A Practical Guide to Quantitative and AI-Native Venture CapitalA plain-English guide to quant VCs, AI-native VCs, and scientific venture capital — how Vela Partners uses data and AI to fund seed-stage AI founders.
- Contrarian vs Consensus in Venture Capital: A Quant VC FrameworkMost legendary venture deals (OpenAI, Wiz, Anthropic) were not contrarian. A quant VC framework for contrarian vs consensus investing as math, not romance.
- How Top VC Firms Evaluate Founders: The Taste Matrix of Sequoia, Benchmark, a16z, Founders Fund, Thrive, and LightspeedData from ~70 seed rounds (2024-2026) shows how Sequoia, Benchmark, a16z, Founders Fund, Thrive and Lightspeed pick founders, and which firm fits you best.
- Can AI Predict Startup Success? What Two Months of a Live Forecasting Arena Taught UsVela ran a live AI arena scoring GPT, Claude and Gemini on real startup outcomes for two months: four early signals, and a preview of VC Arena and VCRouter.
- YC Summer 2026 Batch, Ranked. The B2B Startups Worth a First MeetingThe YC S26 Startup Radar from Vela, a tier-1 quant VC: all 122 B2B companies in the YC Summer 2026 batch, scored with quantitative models. Filter, sort, decide.