The real scarcity in the age of AI isnʼt computing power or access to models but the willingness to take a position that isnʼt already reflected in the consensus. And to own it when it matters.
That observation may sound philosophical, but in practice it shows up in boardrooms and investment discussions with increasing regularity. Moreover, it has a specific mechanism behind it.
When the tools converge, so do the conclusions
The outputs AI systems generate are shaped by what has already been written. That design choice makes them remarkably effective for structured, precedented work — synthesis, benchmarking, summarisation, competitive analysis. The same design makes them unreliable for anything that genuinely hasnʼt happened yet.
The practical consequence for organisations is that AI is very good at producing something defensible: well-formed, internally consistent, unlikely to be obviously wrong. Defensible is useful. It is also, increasingly, the minimum — not a competitive advantage — when every organisation in a market has access to the same capabilities.
If two competitors derive their strategic analysis from the same underlying training distribution, neither has gained ground. They have each purchased the same view at the same price. This convergence matters most where the value of a decision lies precisely in not following the consensus — capital allocation, market entry, product bets on emerging categories.
The precedent problem
William Barnett, who researches how organisations miss transformative developments at Stanford Graduate School of Business, has put the institutional pressure clearly: people fear being a fool more than they hope to be a genius. That asymmetry has always existed. What has changed is the availability of tools that make the safe, defensible answer instantaneous and free — while the non-consensus call still requires someone prepared to articulate and defend it.
Barnett also argues that organisations are consistently poor at anticipating what ends up mattering, but exceptionally skilled at rationalising it in retrospect. His example is Google, which struggled to attract early investment because nearly every preceding search engine had failed. The data was accurate. The inference was wrong.
When AI dramatically reduces the cost of producing a well-reasoned consensus view while leaving the cost of holding a contrary view unchanged, organisations will systematically underweight the contrarian position — even when it turns out to be correct. That is not a problem of individual courage but a structural bias built into the economics of using these tools.
Intelligence without commitment
Inside large organisations, the pattern takes a particular form. Analytical capacity has expanded significantly — faster scenario modelling, broader competitive coverage, synthesised intelligence delivered in minutes rather than weeks. The volume of material available before a major decision has never been higher, and the time required to produce it has never been lower.
It has not made the decisions easier, because the difficulty was never primarily about information. High-stakes choices — entering a new market, deploying significant capital, committing to a platform before its value is proven — are hard because they require choosing one version of the future and accepting responsibility for that choice. More analysis sharpens the inputs, but it does not change the nature of the act.
AI can surface risks earlier, compress assessment timelines, and extend the range of options a team considers before committing. What remains unchanged is the moment of commitment itself — and the accountability that follows it. As AI becomes standard infrastructure, the differentiating variable shifts from analytical capability to the quality of judgment applied above it.
What actually becomes scarce
The economics of execution have shifted considerably. Tasks that previously required significant time, skill, or headcount can now be performed faster and at lower cost. That compression is real, broadly distributed, and still accelerating.
What does not compress at the same rate is the capacity to form an independent view, act on it before the evidence is conclusive, and accept responsibility for the outcome. These remain distinctly human contributions — not because AI lacks processing power, but because they are fundamentally about accountability, not computation.
The risk worth watching is not AI replacing human roles outright. It is the quieter pattern of delegation: each step where the model generates the first draft of a view, each iteration where the human role shifts from forming a position to reviewing one. Individually, each step looks reasonable. Cumulatively, they represent a transfer of judgment that productivity metrics are not designed to capture.
Hence, the question for any leader is not whether AI can handle parts of the work — in most functions, it clearly can. The more important question is whether the parts being delegated include the ones that actually determine outcomes, and whether that delegation was a deliberate choice or simply the path of least resistance.
