515 Startups Got the Same AI Tools. The Ones Who Saw a Map Generated 1.9x More Revenue.
Researchers at INSEAD and Harvard Business School published a working paper in March 2026 that may be the cleanest experiment on AI adoption to date. They enrolled 515 high-growth startups from across the globe in a three-month accelerator and gave every single one the same package: $25,000 in API credits, access to frontier AI models from OpenAI, Google, and NVIDIA, weekly technical training from MIT and Harvard’s AI clubs covering prompt engineering, agentic workflows, and rapid prototyping, plus mentorship, pitch practice, and exposure to leading venture capitalists.
Then they split them in half. Starting in week three, the treatment group attended four additional case study workshops. That was the entire intervention.
What Four Workshops Changed
The workshops were not technical tutorials. They did not teach founders how to prompt better or which API to call. Instead, they showed how AI-native companies had reorganized their entire production processes around the technology, using simple before-and-after diagrams that traced how the chain of work changed.
One case study featured Gamma, a presentations company that grew to over $50 million in annual recurring revenue with roughly 50 employees. The conventional product development chain requires a sequence of specialists, each handling one step in cycles that take months. With AI, the chain compressed: AI detects usage patterns and generates product variants directly, enabling a single product manager to continuously ship features that would previously have required an entire team.
Another case showed how a fintech startup moved KYC screening, its call center, and even hiring interviews to AI, growing its customer base by 20% while reducing costs. A third illustrated how an accounts receivable company replaced eight sequential steps alternating between software and human bridges with a fully automated pipeline, converting a labor-intensive professional service into a scalable software product.
The common thread was not which AI tool to use. It was discovering where AI could reshape how value was created, and how the rest of the organization needed to change around it. The researchers call this the mapping problem, and it is the central concept of the paper.
The Results
The numbers were striking. Treated firms discovered 44% more AI use cases across their operations, concentrated in product development and strategy rather than the obvious applications like email drafting or basic research. They completed 12% more tasks, were 18% more likely to acquire paying customers, and generated 1.9x higher revenue. All while needing $220,000 less external capital (39.5% reduction) and with no change in headcount.
This is not a survey or a self-reported benchmark. It is a pre-registered randomized controlled trial with just 1.6% dropout over the full three months, making it one of the cleanest experiments on AI and business performance to date.
The 90th Percentile Story
The headline number of 1.9x revenue is real, but it conceals an important pattern underneath. For most startups in the study, the difference between treated and control firms was close to zero. The average was driven by a sharp spike among the top 10% of performers.
This is not a limitation of the study. It is arguably its most important finding. The case studies did not turn struggling ventures into winners. They raised the ceiling for firms that were already positioned to execute but had not yet discovered where AI fit into their specific production process. When those firms found the map, the gains compounded across interconnected activities in ways that amplified rather than merely added.
The researchers’ instrumental variable analysis puts a number on the mechanism: each additional AI use case prompted by the treatment led to 0.85 more completed tasks and approximately 26% higher revenue. The treatment induced 2.7 additional use cases on average. If you extrapolate the per-use-case value across the full range of possible applications, the cumulative effect of mapping AI broadly across a firm’s production process is large, and the treatment captured only a fraction of it.
Why the Mapping Problem Is So Hard
The mapping problem is difficult for three reasons. First, AI’s capabilities are uneven and hard to predict. Even closely related tasks can differ sharply in how well AI performs them, and even experts systematically misjudge where AI will help. Second, the search space is vast. Within any firm, AI could potentially be applied to dozens of activities, but managers tend to search locally in the neighborhood of what they already know, defaulting to obvious applications while higher-value uses that require rethinking how work is organized go undiscovered. Third, complementarities compound the difficulty. When activities within a firm are tightly coupled, the value of applying AI to one depends on whether adjacent activities also adjust.
Consider a company that uses AI to write code faster. That is the most common and straightforward application. But to capture real gains, the company must also discover that AI can rapidly prototype and test features with actual users, compressing what was a months-long feedback loop into days. And if the company does not also rethink how it identifies what customers want, it simply builds the wrong product faster. The bottleneck moves. It does not disappear.
One startup in the study illustrated this perfectly. A tender-matching platform built an end-to-end pipeline from tender classification through compliance checking and bid pricing, going from zero to $40,000 in revenue and four paying customers during the program without hiring the technical talent that would conventionally be required. The gain did not come from adopting a better model. It came from reorganizing who does what.
Who Benefits - and Who Doesn’t
A natural assumption would be that technically trained founders benefit more from this kind of intervention, since they can implement what they see faster. The data shows the opposite: there is no significant differential effect by technical background. The mapping problem is not a skill gap. It is a discovery gap.
Similarly, you might expect that higher-performing firms benefit more because they have the execution capacity to act on new information. Again, no significant differential effect by baseline performance. This contradicts most task-level AI research, where lower performers tend to benefit the most. At the firm level, the binding constraint is how broadly founders search across their production processes, and that constraint appears to bind regardless of background or starting position.
The practical implication is direct. If your organization has AI tools, training budgets, and motivated employees but is not seeing the returns, the problem may not be the technology, the budget, or the people. It may be that nobody has shown your team what reorganized production looks like in a company that resembles yours.
What This Means for the Investment Question
The control group in this experiment had the same technology, the same training, the same API credits, and the same access to mentors and investors. They used the same AI tools at comparable rates. The only thing they lacked was information about where other companies had found value, and that single gap accounted for the entire difference in outcomes.
For any CEO currently evaluating AI investments, this reframes the spending question. Before the next tool purchase, platform migration, or consultant engagement, there is a simpler and cheaper question worth answering first: has your leadership team seen what reorganized production actually looks like in companies that operate in your space? The AI measurement gap showed that most organizations cannot even quantify what their current AI tools do. This research suggests that the diagnostic step before measurement may be even more fundamental: mapping where the value is before trying to measure whether it is working.
The researchers note one important caveat. The ten-week window captures short-term effects. Whether these performance gains compound or dissipate as ventures mature remains an open question. And they study early-stage firms with low organizational inertia. How the mapping problem evolves as companies grow, as hierarchies deepen and processes calcify, is a question the paper cannot answer. The core finding is universal: as AI capabilities expand, the mapping problem may become harder, not easier, as the search space explodes and organizational inertia grows.
Related: Most Companies Have AI Tools. Few Have Redesigned How They Build. examines why adding AI to existing workflows produces marginal gains while redesigning workflows around AI produces compounding ones.