AI and Early Stage Venture Formation

The experiment randomly varies the degree of AI knowledge and encouragement to integrate AI into venture-building during an eight-week “AI Founder Sprint.” All founders participate in a common accelerator-style program that includes weekly group meetings, access to a global mentor pool, interactive workshops, master classes with leading entrepreneurs and faculty, AI engineering classes, and the opportunity to pitch to top global VCs at the end of the Sprint.

In the treatment arm, founders additionally receive targeted interactive workshops and case studies that explicitly encourage them to apply AI across their firm, embed it in their product, and reimagine their venture’s production function with AI. The control arm receives the common program without these AI-intensive components.

2025-09-09

2025-10-17

We will measure the following primary outcomes:
1. Number of AI use cases: The count of distinct applications of AI reported by the venture during and after the Sprint.
2. Index of venture progress: A composite index based on four binary indicators: (i) whether the product is launched (i.e., publicly available and usable by customers), (ii) whether the venture has customers, (iii) whether the venture is generating revenue, and (iv) whether the venture has raised investment.
3. Venture survival: A binary measure of whether the venture is still operating six months after the Sprint.

We will also measure the following secondary outcomes:
1. Number of activities taken each week to advance the startup
2. A binary indicator of whether the venture has pivoted from its original idea by the end of the Sprint.
3. Number of active users
4. Financial performance: revenue, costs, and additional investment raised
5. Labor demand as measured by job postings expectations (i.e., distinct key skill sets sought by the venture)
6. Capital demand as measured by investment expectations

Exploratory Analyses
In addition to confirmatory outcomes, we will conduct exploratory analyses using text-based responses collected at baseline and follow-up. Specifically, we will analyze open-ended responses on:
- The problem the venture is solving
- The solution being developed
- The business model
- Customer acquisition approaches
- The biggest challenges and uncertainties the venture faces

These analyses may involve developing coded measures (e.g., breadth of customer acquisition channels, type/severity of challenges) or using natural language processing methods. Because these analyses are exploratory, we do not commit to specific hypotheses or measures in advance, and we will treat any results as hypothesis-generating rather than confirmatory.

This study tests whether increasing founders’ knowledge of how to apply AI within their business -- and explicitly encouraging them to reimagine their venture’s production function with AI -- changes (i) how they use AI, and (ii) the rate and direction of venture progress.

Not available

We implemented stratified randomization to assign ventures to treatment and control. Stratification was based on three elements: Continent × Traction Score (0–3) × AI Use Cases (Low vs. High).
- Traction Score is constructed from three binary indicators: whether the venture has revenue, has raised investment, and has launched a product. Scores range from 0 (no traction) to 3 (all three conditions met).
- AI Use Cases is a binary measure constructed from founders’ baseline reports. Ventures with 0–2 use cases (below the median) are coded as “Low,” while ventures with 3 or more use cases (at or above the median) are coded as “High.”

Randomization was implemented by computer, assigning ventures to one of two arms, treatment, and control, within each stratum.

We will assess baseline covariate balance between treatment and control groups using pre-treatment variables, including year founded, founder gender, venture stage, binary indicators for having investment, revenue, customers, and product launched, as well as continuous measures of investment amount, monthly revenue, team size, and years of work experience. To ensure comparability across ventures and reduce the influence of extreme outliers on the right hand side of the distribution, we will also winsorize traction and financial metrics at the 90th and 95th percentiles. At the time of registration, we have not identified any meaningful imbalances, and we will formally report covariate balance in the final paper.

The unit of randomization is the venture. All outcomes are measured at the venture level.

No

N/A

Currently, we have 585 ventures confirmed and enrolled. We expect the final sample size to vary due to attrition given the virtual nature of the Sprint, but anticipate it to be approximately 500 ventures.

Assuming a final sample of approximately 500 ventures, we anticipate an approximately 1:1 allocation between treatment and control arms.