Teaching startups how to reorganize around AI nearly doubles short-run revenue and raises customer acquisition by 18%, driven by a 44% rise in discovered AI use cases; growth occurs without proportional increases in labor and with lower external funding needs.

Main Finding

Providing startups with structured examples and frameworks for how other firms reorganize production around AI (solving the “mapping problem”) causes firms to discover substantially more AI use cases and produce large, causal firm-level performance gains: treated startups discover 44% more AI use cases, complete 12% more tasks, are 11 percentage points (≈18%) more likely to acquire paying customers, and generate 1.9× higher revenue. Growth occurs without proportional increases in inputs—external capital demand falls (~39.5%, ≈$220k less) and labor demand is unchanged—and gains are concentrated at the upper tail of performance.

Key Points

• Mapping problem: a key friction is not access to AI but discovering where/how within a firm’s production AI creates value. Firms tend to search locally and miss high-value, non-obvious applications.
• Intervention: in a 3-month INSEAD accelerator (AI Founder Sprint), treated firms received case studies and frameworks showing how other firms reorganized workflows, teams, business models, and financing around AI. Control firms received standard accelerator content and technical training but not the mapping-oriented case studies.
• Effects on AI adoption: treated firms discover 2.7 additional AI use cases (a 44% increase), particularly in product development and strategy-related domains.
• Performance impacts:
  • 12% more completed tasks (measured during the accelerator),
  • +11 percentage points (≈18%) higher probability of acquiring paying customers,
  • 1.9× higher revenue on average for treated firms.
• Heterogeneity: revenue and investment gains concentrate in the 90th percentile and above—AI expands the upper bound of venture performance rather than uniformly shifting the distribution.
• Inputs and scaling: despite faster growth, treated firms do not proportionally increase labor; they request substantially less external capital (~39.5% decline, ≈$220k less).
• Mechanism evidence: instrumenting AI use cases with random assignment implies causality from mapping-induced use-case discovery to outcomes—each additional AI use case induced by treatment → +0.85 completed tasks and ≈+26% revenue.
• Robustness/signals: low and balanced attrition (1.6%); balanced baseline covariates; feedback indicates comparable engagement across groups; effects are driven by changes in AI use cases, not non-AI channels.

Data & Methods

• Context: AI Founder Sprint, a global, virtual 3-month accelerator for early-stage, high-growth startups (median founding year 2024, median team size 4). Sample includes 515 firms that consented and completed baseline data.
• Randomization: stratified random assignment (32 strata) on geography (4 regions), a traction score (0–3: product launched, revenue, external investment), and baseline AI use (0–2 vs. 3+ use cases). Final split: 255 treatment, 260 control. Baseline balance confirmed.
• Common support: all firms received identical technical resources—API credits and access to frontier models (~$25k in-kind), technical training, peer groups, demo days, and weekly progress reports.
• Treatment: starting week 3, required workshops for treated firms delivered case studies and frameworks illustrating how AI-native firms reorganize production (team structure, workflow cadence, business-model implications). Peer learning and office hours reinforced workshop content.
• Outcomes and measurement:
  • Weekly progress reports capturing AI use cases, tasks completed, traction outcomes (customers, revenue), and resource requests (fundraising, hiring).
  • Primary outcomes: number of new AI use cases; tasks completed; customer acquisition; revenue; external capital demand; labor demand.
• Identification and causal inference:
  • Intent-to-treat estimates from randomized assignment.
  • Instrumental-variable estimates using treatment assignment as an instrument for number of AI use cases to estimate returns per additional use case.
  • Pre-registered design and IRB approval; low attrition; supplementary analyses rule out alternative (non-AI) explanations.

Implications for AI Economics

• Discovery/search frictions matter. The dominant bottleneck to realizing firm-level gains from AI is managerial search and mapping (identifying which internal activities to transform and how), not merely access to tools or technical training. Policy and firm interventions that broaden the search space (case studies, analogies, frameworks) can unlock large returns.
• Complementarities and reorganization amplify returns. When firms discover AI applications that trigger coordinated changes across activities, gains compound—explaining why effects are largest in the upper tail (AI expands the frontier of achievable firm performance).
• Rethinking productivity measurement. Task-level gains can (and here do) scale to firm outcomes if mapping frictions are addressed. Empirical work on AI should measure both the mapping process (applications discovered, reorganization) and firm-level outcomes to capture aggregate effects.
• Capital and scaling dynamics change. AI can enable faster growth without proportional increases in traditional inputs; investors and entrepreneurs may need to update expectations about capital intensity and the nature of growth (greater returns at the top, lower external capital demand for some startups).
• Practical recommendations:
  • For firms: invest in structured discovery—use analogical case studies and cross-functional frameworks to search beyond obvious, local AI use cases.
  • For accelerators/trainers/policymakers: provide not only access and technical skills but also concrete examples and frameworks for organizational reconfiguration around AI.
• External validity & future research: experiment is in early-stage startups (low organizational inertia) over a 3-month window—effects in larger, more established firms or over longer horizons require study. Future work should track persistence of gains, optimal complements (organizational changes, incentives, human capital), and welfare/market-level implications as mapping frictions are addressed more broadly.