Meta has lost over US$80 billion on Reality Labs trying to build a virtual world from scratch. On the other hand, Roblox lets its users build the world themselves, scaling past 150 million daily active users and a market capitalisation above US$30 billion (at the time of publication). Same sector, same technology – except one saw the formula, letting users create value at near-zero marginal cost, while the other assumed the old rules still applied.

For most of modern business history, success was shaped by knowledge and execution. Knowledge was gatekept by institutions and geography; while execution hinged on capital and time. Today, an entrepreneur in Doha, Lagos or Sofia can compete for the same client as anyone anywhere, and anyone can turn imagination into execution by describing what they want in “natural language” and letting AI build it.

When everyone has access to the same tools, what cannot be commoditised is originality of thinking, judgement and the ability to see the structural formulas that others cannot. Every industry runs on these formulas: the incentive architectures that determine who gets paid and why, the power dynamics that dictate which ideas survive, and the rules that decide which practices become standard and which become liabilities. 

AI has not changed these formulas; it has made the consequences of not seeing them more costly. Here are the principles leaders need to open their eyes.

Inventory what you assume

AI adoption is not a technology challenge; it is an intellectual one. It forces organisations to take stock of everything they take for granted and ask: if we were building this from scratch today, would we build it this way?

SpaceX asked why rockets had to be disposable. Through vertical integration and rapid iterative testing, it created an entirely new category. It was long assumed that journalism requires large newsrooms, expensive bureaus and concentrated urban audiences. These assumptions are now challenged by projects like Botipedia, the AI-generated encyclopaedic knowledge portal developed at INSEAD that produces structured content at a scale that traditional methodologies (or even large language models, widely known as LLMs) cannot match, including content in underserved languages and obscure topics.

Debunking one legacy assumption yields incremental improvement. Dismantling many exposes the underlying formula and opens the door to moonshots, or entirely new categories.

Design for scale, but mind the ethics

Three design principles reshape the formula. The first is near-zero marginal cost: AI processes can be replicated across unlimited applications at negligible cost. Every digital product and workflow should be stress-tested with one question: If it works, can it scale without proportional increases in cost or headcount?

The second is to rethink who creates value. The most powerful platforms let users do the work: Uber incentivised drivers until passengers became inevitable, TikTok drew creators until the audience became self-sustaining. In both cases, the platform produces none of the content and captures most of the value. When your users become your content creators, testers and distribution channel, growth compounds in ways traditional organisations cannot match.

The third is that in an exponential world, ethical stakes compound with everything else. OpenClaw, the open-source AI assistant, attracted over 250,000 GitHub stars within months of its launch while exposing more than 40,000 unprotected instances that leaked API keys and personal data. Its security failures scaled at the same speed as its adoption. This example shows that the integrity of what you build is no longer a reputational concern but an operational one.

Question the “truth”

The LLMs organisations rely on carry structural blind spots. They are trained on information found on the internet, a representation of opinions shaped by different incentives, interests and worldviews. These models optimise for plausibility, not truth.

In the United States, 71% of healthcare deep-learning algorithms are trained on patient data from just three states. More than half of the web pages in Google’s C4 dataset – a massive, publicly available dataset – used to train LLMs are US-hosted. Over 50% of studies in some fields cannot be replicated. Yet AI ingests them indiscriminately. Chinese models may carry different blind spots, but the structural problem is the same: No model is neutral, and none should be treated as such.

As AI-generated content becomes training data for subsequent models – a phenomenon known as model collapse – quality degrades with each generation.

Audit your dependencies

AI dependency spans these dimensions. 

1. Energy sovereignty: Who powers your AI?
2. Computational sovereignty: Where does processing take place?
3. Data and algorithmic sovereignty: Who controls the data your AI uses and whose models make your decisions?
4. Cultural and epistemic sovereignty: Does your AI reflect your worldview or default to its training data?
5. Talent sovereignty: Can your people still deliver if the AI is switched off tomorrow? 
6. Governance sovereignty: Whose laws apply when your provider operates under a foreign jurisdiction? 
7. Strategic sovereignty: Is every workflow you embed building independence or deepening dependency?

Each dimension compounds the others. A geopolitical sanction or a provider’s change in terms could disrupt access across all of them at once. If your contingency plan for a loss of AI infrastructure is non-existent, you do not have an AI strategy; you have an AI dependency.

Rethink what you protect

The traditional intellectual-property playbook was built for a slower world. Today, features are copied in days and code is replicated in hours. The question is no longer how to prevent imitation, but how to secure the freedom to operate.

The most durable protection in an AI economy is a distinct, curated and growing dataset that no one can replicate. When the use of your product generates data that improves it, attracting more users who generate more data, you have created data gravity. This virtuous loop is harder to imitate than any patent and such a position is far cheaper to defend than any lawsuit.

Companies may already possess this advantage – even if they don’t recognise it. Telecoms sit on the most valuable behavioural dataset on the planet. LinkedIn is a similar gold mine, with over a billion people sharing what they know, what they do for a living and how their careers are changing.

Act quickly, bet differently

If AI amplifies everything at near-zero marginal cost in a landscape that shifts in weeks, large, concentrated bets become dangerous. Convex bets, with capped downside and open-ended upside, let you learn cheaply if they fail and scale quickly if they succeed. Asymmetry is the point.

It is easier to compete by being 10 times better than 10% better, because 10 times redefines the category. The most successful organisations make this leap with discipline: Using cash cows to fund the conviction-driven calves, and retiring the cows when the calves outgrow them. This landscape is incompatible with hierarchical structures and risk-averse cultures.

Here is what most leadership literature does not tell you: The single greatest competitive advantage now is not a framework, platform or proprietary dataset. It is curiosity, courage and the conviction that you can figure it out. 
 

Any references to companies or organisations reflect the authors’ views and do not constitute endorsement by INSEAD.