Why Early AI Advantages Can Limit Machine Learning Startups Later
Image: "Wikimedia Servers-0001 42" by Helpameout (CC BY-SA 3.0) via Wikimedia Commons, recoloured.
Do the technical choices behind a machine learning product shape a young firm's strategy? Fraser and Özcan followed six early-stage ML ventures and found that training data, learning paradigms and model outputs are tightly interdependent, and that features giving a venture its early edge can later narrow the strategic options open to it.
When the model shapes the strategy
Strategy research treats machine learning capabilities as valuable assets, but it has said little about how the technical make-up of those capabilities affects a firm's competitive position. Jack Lewis Fraser (Imperial College London) and Pınar Özcan (University of Oxford) address that gap by treating the technology itself as an explanatory variable rather than a black box.
Their design is inductive and longitudinal. The authors tracked six early-stage machine learning ventures taking part in an accelerator program, following how each moved between phases of specialization, in which a firm narrows its focus, and phases of exploration, in which it searches for new applications. The question was whether properties of the underlying technology constrained those moves.
According to Fraser and Özcan (Organization Science, 2026), three technical elements matter: the training data a venture relies on, the learning paradigm it uses and the outputs its models produce. These elements depend on one another and on the applications they serve, and the relationships among them shift as the firm grows. The resulting interdependence shapes each venture's growth path, sometimes widening and sometimes restricting the strategic options available.
The headline finding is a paradox of path dependence. Characteristics that give a machine learning venture an early competitive advantage can become constraints on later growth. A data and model configuration that fits a first market well may not transfer easily when the firm tries to expand or pivot.
As a multicase study of six ventures within a single accelerator setting, the work is designed to build theory rather than to estimate effect sizes, and its conclusions are framed as implications for strategy research on competitive advantage and path dependence rather than as general predictions for all AI firms.
Why it matters: For researchers studying AI firms, digital strategy or entrepreneurship, the paper is a reminder that technical configuration is not neutral background. Measures of AI adoption or capability that ignore training data, learning approach and output type may miss the mechanism that ties early success to later lock-in. The study also offers a template for qualitative, process-oriented work that opens up the technology rather than treating it as a single variable.
Fraser, J. L., & Özcan, P. (2026). Growth Paths in Early Machine Learning Ventures: Technical Configuration and Competitive Advantage Over Time. Organization Science. Read the original paper. Independent editorial summary of third-party research.
Comments