A field experiment with 515 startups suggests AI creates firm-level gains when teams learn where to reorganize work around it, rather than only getting access to tools.
Source note: Hyunjin Kim, Dahyeon Kim, and Rembrand Koning. Mapping AI into Production: A Field Experiment on Firm Performance. INSEAD Working Paper 2026/20/STR, March 30, 2026. https://papers.ssrn.com/sol3/papers.cfm?abstract_id=6513481
We have all seen localized AI wins: drafts completed in seconds, meetings summarized instantly, and boilerplate code generated in a flash. Yet, a paradox remains. Why do individual productivity gains fail to translate into firm-level performance breakthroughs? If a developer is fifty percent faster, why is the company not growing fifty percent faster?
A new paper, “Mapping AI into Production: A Field Experiment on Firm Performance,” by Hyunjin Kim, Dahyeon Kim, and Rembrand Koning, offers an empirical answer. The bottleneck is not access to models, budget shortages, or poor prompting, but the organizational mapping problem.
Firms fail to realize overall benefits because they do not know where and how to integrate AI across workflows. They default to obvious, isolated use cases like customer support chatbots or writing assistants, leaving high-value applications that require workflow restructuring undiscovered. When guided to search more broadly and reorganize operations around the technology, the results are clear: founders discover more use cases, complete more tasks, acquire more customers, and generate nearly double the revenue, while decreasing their need for external capital.
Why This Paper Matters
Most empirical evidence on AI focuses on individual worker productivity, showing that consultants write reports or developers code faster. But a firm is an interconnected production chain, not a collection of isolated tasks.
This study connects individual productivity to firm-level performance. Through a randomized controlled trial of 515 high-growth startups in a global, three-month accelerator, the researchers showed that task-level gains dissipate unless adjacent workflows are adapted.
The paper shows that access is not the bottleneck. Spending to provide AI tools and training is insufficient. The control group had access to the same models, training, and resources as the treatment group, yet the treated group outperformed them. The difference is managerial: the ability to reorganize the firm.
The Idea in Plain English
Imagine an assembly line. If a robotic arm installs steering wheels ten times faster, you might expect the factory to produce cars ten times faster. But if the next station, dashboard installation, still operates at human speed, total output remains unchanged. Steering wheels simply pile up. To accelerate production, you must reorganize the line, perhaps by combining both steps.
This is the mapping problem. AI is a general-purpose technology, like electricity or the internet. It is highly flexible, capable of reshaping operations, product development, strategy, and sales. However, because it is so versatile, its best applications are neither obvious nor uniform across businesses.
When organizations grant employees AI access, the default response is local search. Employees look at immediate tasks and ask how the tool can make them slightly faster, using it to draft emails or debug scripts. This is the equivalent of speeding up steering wheel installation. It feels like progress but leaves surrounding bottlenecks intact.
To capture value from AI, a firm has to map the technology across its production process. For example, if AI generates code instantly, engineering is no longer the bottleneck; the constraint shifts to customer feedback. The firm has to reorganize to build and test multiple prototypes in parallel, which requires a different team structure and operating rhythm. The paper shows that teaching firms to search for these wider applications helps them break through organizational bottlenecks.
What the Researchers Tested
To isolate the mapping problem’s impact, the researchers conducted a randomized field experiment within INSEAD’s AI Founder Sprint, a three-month global startup accelerator. The sample comprised 515 high-growth, early-stage startups. These were serious ventures: 52.6% had launched a product, 55.3% had acquired customers, and about a third generated revenue and had raised external capital.
The startups were randomly assigned to a treatment group of 255 firms or a control group of 260 firms, balanced on geography, baseline traction, and prior AI usage.
Both groups received identical baseline support. Every startup received approximately $25,000 in API credits and partner tools from Google Cloud, OpenAI, NVIDIA, and Manus AI. All attended weekly, three-hour technical training sessions covering rapid prototyping, retrieval-augmented generation (RAG), agentic workflows, and “vibe-coding.” They also competed for $100,000 in seed funding and pitched to venture capitalists.
The only difference lay in their weekly 60-minute workshops. The control group followed a standard entrepreneurship curriculum focused on customer profiles and lean validation.
The treatment group studied case-based strategies showing how AI-native firms mapped AI into production and reorganized workflows. These were strategic breakdowns, not technical prompting tutorials: * Gamma: Compressed its product development cycle, enabling a single product manager to ship features continuously without an engineering team. * Ryz Labs: Used AI to build multiple prototypes in parallel. * FazeShift: Automated an entire eight-step accounts receivable workflow, replacing human glue. * Ranger: Automated its service delivery to function like a high-margin software product, delaying the need for venture capital.
These examples prompted the treatment group to search broadly. The researchers tracked both groups through weekly progress reports, with a minimal 1.6% attrition rate balanced across groups.
What They Found
The results show that solving the mapping problem improved both operating execution and financial performance.
Expanding the Scope of AI Adoption
Treated firms discovered 2.7 additional AI use cases compared to the control group (a 44% increase) and deployed the technology more widely across functional categories.
The divergence was largest in areas requiring process reorganization. Treated startups were far more likely to use AI for product development, product and strategy design, and business operations like workflow automation. Both groups used AI similarly for basic tasks like research, writing, and sales drafting. The intervention pushed startups to integrate AI into core operations rather than default to administrative tools.
Firm Performance and Task Velocity
AI integration into production accelerated execution. Treated firms completed 12% more tasks, driven by internal activities like product building, landing page design, and financial modeling. External tasks, like customer discovery or investor pitching, remained unchanged.
This internal speed led to commercial traction: treated firms were 4.7 percentage points more likely to launch a product and 11 percentage points (18%) more likely to acquire paying customers. Treated firms generated 1.9 times more revenue. An instrumental variable analysis estimated that each additional use case led to 0.85 more completed tasks and a 26% increase in revenue.
The Right-Tail Breakouts
Gains were not evenly distributed. While the treatment helped startups at all levels, revenue and investment gains were concentrated in the upper tail, with the largest effects at the 90th and 95th percentiles. AI mapping acts as a force multiplier for the most promising ventures.
Scaling with Less Capital
Typically, fast-growing startups raise more capital and hire more staff. Treated firms did the opposite: despite growing faster, their target capital ask fell by 39.5% relative to the control group (a reduction of $220,000), while their demand for technical labor remained unchanged. AI allowed these startups to break traditional scaling laws and achieve higher traction with less capital.
Why It Happens
The study’s core point is simple: AI search is local, but value is systemic.
The control group struggled due to cognitive bounds. Faced with generative AI, managers default to what they know and apply the tool to immediate, isolated tasks.
This localized search fails to generate systemic value because of complementarities, as illustrated by the O-ring theory of economic development. Under this theory, the overall speed and quality of a production process are constrained by its tightest bottleneck. If you use AI to accelerate step two of a five-step process while steps three through five remain unchanged, total output is unaffected. The bottleneck merely shifts, trapping the productivity gains.
The treatment succeeded by showing founders how other companies restructured their operations. Case studies of companies rewiring entire production chains broke local search patterns. Founders realized they needed to automate full sequences, like accounts receivable, rather than simple data entry. They recognized that when coding takes minutes instead of months, customer feedback becomes the new bottleneck, requiring them to prototype and test multiple versions in parallel. By mapping AI across the production process, treated firms identified actual bottlenecks and designed workflows to exploit the technology, which decoupled growth from headcount and capital scaling.
What This Means for Builders
For founders, this paper is a blueprint for outperformance. The primary takeaway is that your technical stack is no longer a sustainable competitive advantage. Everyone has access to the same models, and the cost of intelligence is trending toward zero. If you only use AI to speed up traditional, sequential software development, you remain in the control group.
To win, builders must change the architecture of their startups. They need to move from helping engineers code faster to reorganizing the product lifecycle to bypass traditional development bottlenecks. Builders should identify human glue (points where employees manually transfer data between systems) and replace those sequences with automated AI pipelines.
Furthermore, builders must internalize the capital implications. AI mapping reduces external capital needs by 39.5%. This allows founders to raise less money, suffer less dilution, and retain more control over their companies while hitting commercial milestones faster.
What This Means for Buyers and Operators
For enterprise buyers, CIOs, and operators, this research shows the limits of current AI management. The standard playbook of buying enterprise licenses, offering prompting seminars, and hoping for productivity gains does not work.
Operators must treat AI deployment as an organizational design challenge, not an IT issue. Giving AI to employees without redesigning workflows results in localized efficiencies that do not improve corporate profitability.
To get a return on AI investments, organizations have to solve the mapping problem. That means mapping production chains, from customer intake to billing, and identifying where AI can eliminate sequential bottlenecks. Managers need case studies and frameworks that help them restructure departments. The value of AI in the enterprise will not go to companies purchasing the most tokens, but to those that reorganize their operating models.
What to Watch Next
The next question is how the mapping problem scales with complexity. This experiment examined early-stage startups with little organizational inertia. In a small company, founders can reorganize workflows over a weekend. In a Fortune 500 company with deep hierarchies and established procedures, reorganizing the production chain around AI will face resistance. How large enterprises attempt to solve this mapping problem, and whether they succeed, is the next thing to watch.
As models improve, the search space for mapping will expand. Processes that cannot be automated today may become trivial next year. Firms need continuous mapping capabilities rather than treating AI reorganization as a one-time shift.
Limitations and Caveats
These findings should be interpreted within their context. The experiment was conducted on early-stage, high-growth startups in a top-tier accelerator. These startups have little technical debt and low organizational inertia. It is unclear if the same case-based training would produce similar results in a mature corporation.
The experimental window was ten weeks long, which is short for measuring long-term outcomes. While sufficient to capture early traction, it is unknown whether these gains compound or represent a one-time step-change in efficiency. The data relies in part on self-reported use cases and capital projections, though the researchers validated these inputs. Finally, because the treatment bundled multiple case studies, the experiment cannot isolate which specific case study drove the performance gains.
Source
Kim, Hyunjin, Dahyeon Kim, and Rembrand Koning. “Mapping AI into Production: A Field Experiment on Firm Performance.” INSEAD Working Paper 2026/20/STR, March 30, 2026. https://papers.ssrn.com/sol3/papers.cfm?abstract_id=6513481