A small fund can receive more applications in a week than its team can read properly. AI is attractive because most of the first pass is reading: what does the company do, what stage is it, what numbers does the deck claim, does it fit the thesis at all. That reading can be done by a model, as long as somebody can check what it read.
The harder question is where to stop. The sections that follow cover how VCs use AI across screening and diligence, which jobs suit it, which decisions should stay with people, and the practical controls (citations, untrusted input, override with reasons) that keep the work checkable. It uses public frameworks from NIST, the EU AI Act and OWASP as checklists, and a large survey of how venture investors actually make decisions.
How VCs use AI across the deal funnel
Start with the shape of the work. In a survey of 885 venture investors at 681 firms, Gompers, Gornall, Kaplan and Strebulaev found that for each deal a firm closes it considers about 101 opportunities, meets the management of 28, reviews 10 at a partner meeting and runs due diligence on about 4.8. The average firm in that sample screened about 200 companies a year and made four investments. Most of the effort sits at the wide top of that funnel, and most of it is reading.
That is where AI fits best: jobs where the output is a summary of a document the team already has, and where a person can check the summary against the original.
| Job | What AI can do | What a person still does | Risk to watch |
|---|---|---|---|
| Deck reading | Extract company facts, stage, round and claimed metrics, each with the slide it came from | Spot-check the citations and correct what was misread | Charts and tables read wrongly; numbers without a period or definition |
| Thesis screening | Apply written rules and preferences, and mark each criterion met, mismatched or unknown | Decide what to do with mismatches and unknowns | Treating a missing value as a failure |
| Question drafting | Suggest follow-up questions from gaps and open claims | Choose, edit and send the questions | Generic questions that ignore what the founder already said |
| Research | Summarise public sources on the market and competitors, with links | Read the sources that matter and judge their quality | Confident summaries of weak or outdated pages |
| Memo drafting | Assemble sections from the evidence already in the record | Write the judgement: why invest, the main risks, the recommendation | A draft that sounds finished before the thinking is done |
| Pipeline triage | Sort and group new applications for the right reviewer | Decide who gets a meeting and who is passed | Quiet rejections nobody looked at |
Notice the pattern in the third column. The person does not redo the AI's work, but they do own the decision that follows from it. Our guide on how to screen startup applications covers the screening steps in more detail.
What AI should not decide
The same survey asked investors what matters most when they choose a company. The management team was named as an important factor by 95% of firms and as the most important factor by 47%. Business model, product and market followed. Judging a founding team, a market's timing and whether a fund can help is exactly the kind of call a model cannot make on a fund's behalf, because it rests on context, relationships and a view of the future that no deck contains.
So draw the line in writing. AI extracts, applies, flags and drafts. People decide every stage change, every pass and every investment, and approve every message that goes to a founder.
- No automatic rejections. If a rule says a deal is out of scope, a person should still see it before the founder hears no.
- No AI messages to founders without a person approving them first.
- No scores that hide their reasons. Each criterion should show the rule, the value found, the source and the rationale.
- Overrides are allowed, and recorded with a reason, so the team can learn where the rubric is wrong.
The EU AI Act's Article 14 on human oversight is written for high-risk AI systems, and whether a fund's screening tool falls in that category is a question for your counsel. Its points still make a good checklist for any fund: the people overseeing the system should be aware of the tendency to over-rely on its output (automation bias), able to interpret the output correctly, and able to disregard, override or reverse it. If you screen founders in the EU or UK, Article 22 of the GDPR also gives people the right not to be subject to a decision based solely on automated processing that significantly affects them, which is one more reason to keep a person in every pass decision.
Make every AI output checkable with citations
An AI summary without a source is a claim you have to take on trust. A summary where every fact links to the slide, page or transcript moment it came from is a claim you can check in seconds. The difference decides whether your team actually saves effort or just moves it to second-guessing.
Ask three things of any AI tool you use for deals:
- Does every extracted fact cite where it came from, down to the slide or page?
- Are quotes checked against the source, and dropped when they are not found there?
- Does it say when it could not read something, such as a chart or a scanned page, rather than guess?
Unknown is a valid answer. A deck that does not state revenue has not failed a revenue rule; it has left a question. Good screening turns that unknown into a question for the founder instead of a quiet no. Our guide on how to evaluate a pitch deck explains what to look for on each slide, with or without AI.
Treat decks, files and websites as untrusted input
Pitch decks are written by people who want a yes. That is normal and fine for a human reader, but a model reading a document can be manipulated by text placed in it, for example a line in white text that says to rate the company highly. OWASP's Top 10 for large language model applications lists prompt injection first (LLM01:2025), and describes the indirect form as when a model accepts input from external sources such as websites or files.
The practical defences are simple to state:
- 1Keep instructions to the model separate from the content it reads, and mark uploaded content as untrusted.
- 2Never let a model take a privileged action, such as changing a deal's stage or emailing a founder, without a person approving it. OWASP recommends human-in-the-loop controls for privileged operations.
- 3Compute scores in code from criterion results, so a persuasive paragraph cannot move a number directly.
- 4Scan uploaded files for malware before any preview or extraction.
- 5Check what your AI provider's terms say about using submitted content for training, before you upload founders' documents.
How to introduce AI in a small fund
NIST's AI Risk Management Framework (AI RMF 1.0, released 26 January 2023, voluntary) organises the work into four functions: Govern, Map, Measure and Manage. NIST added a profile for generative AI (NIST AI 600-1) on 26 July 2024. You do not need a compliance programme to use the idea. For a fund, it translates into a short sequence:
- 1Govern: write down which decisions stay with people, and who approves messages to founders.
- 2Map: list the jobs you want AI to do, the data each one touches and what goes wrong if it is wrong.
- 3Measure: before relying on deck extraction, compare its output with the decks for a sample of applications and note what it misreads.
- 4Manage: keep a log of overrides and corrections, review it monthly and adjust the rubric or the prompts.
- 5Tell founders, in plain words on your application page, how AI is used in your process and that people make the decisions.
Common mistakes
- Letting a model rank founders on how polished the deck or the speaker is. Assess substance and evidence; exclude accent, fluency and presentation polish.
- Pasting confidential decks into a consumer chat tool with unclear data terms.
- Accepting AI research without opening the sources.
- Using AI only to say no faster. The same tools can flag promising deals your team would have missed in a busy week.
- Skipping the measurement step, then losing trust after the first misread number.
Fairness needs separate care. Writing criteria down and applying them the same way to every application helps, but a model can still reflect patterns in what it was given. Our guide on bias in startup screening covers the process side.
Where Delsivo fits
Delsivo is an AI investment workspace for venture funds, at the design partner stage: nothing is generally available yet, and each capability is labelled with the release it arrives in. It is built around the split described above. AI suggests and cites; people decide.
In the Pilot, Delsivo is designed to read PDF decks and cite the slide for every extracted fact (deck analysis), apply your thesis as versioned rules and preferences with Met, Mismatch and Unknown results (thesis screening), and keep claims with evidence labels. A mismatch is surfaced for review, never an automatic rejection, and an unknown becomes a question. Scores are computed in code from criterion scores, uploads are treated as untrusted input and scanned before extraction, and overrides need a reason.
In the first release, live AI founder interviews (the AI asks the fund's questions and assesses only the transcript), cited research and Ask This Deal are planned. Committee memos and packets are planned for a later release. No stage change, pass or founder message happens without a person.
Sources
- Gompers, Gornall, Kaplan and Strebulaev: How Do Venture Capitalists Make Decisions? (NBER Working Paper 22587, September 2016) (checked 6 November 2026)
- NIST: AI Risk Management Framework (checked 6 November 2026)
- EU AI Act, Article 14: Human oversight (checked 6 November 2026)
- GDPR, Article 22: Automated individual decision-making (checked 6 November 2026)
- OWASP GenAI Security Project: LLM01:2025 Prompt Injection (checked 6 November 2026)
Delsivo publishes this guide and makes software for venture funds, so we have a stake in the topic. Facts about other companies, research and regulations come from the sources listed, on the dates shown, and can change. Corrections are welcome through our contact page.