The venture capital world thrives on the notion that a handful of companies will reshape entire economies. Now the tools investors increasingly rely on to spot those companies may be working against them.
Roughly three-quarters of venture capital firms now use AI to help evaluate deals. It’s faster and more thorough than anything that came before it, processing pitch decks, sizing up competitors and flagging risks. But there is a fundamental problem with this approach: AI is trained on the past, but startups that reshape industries rarely resemble the past.
Bias Of The AI Analyst
For decades, VC investing has followed a familiar ritual. A founder presents a pitch deck. Investors meet the team and run diligence. They invite advisors and industry experts to weigh in. Together they ponder whether this startup is the next Google, Uber or NVIDIA. They ask, is this the next improbable company that could change everything? And eventually someone places a bet on a future that does not yet exist.
In many ways, AI is a genuine upgrade to the investment process. Investors paste in diligence materials and ask questions like: Is this market real? Who are the competitors? What are the risks? Is this technology actually feasible?
AI can scan thousands of research papers, patents and industry reports in seconds. It can map competitive landscapes and identify regulatory hurdles. At speeds unimaginable even a year ago, AI supplies investors with a confident, seemingly thorough and impressively well informed analysis.
But the strength of AI is also where the risk begins. The problem is not that AI gets facts wrong. Getting facts right is exactly what AI does best. In this case, perhaps too well.
Large language models generate answers by identifying patterns across enormous datasets. They predict what is most likely based on facts and past outcomes.
That structure makes them extremely useful for evaluating incremental improvements. It also makes them far less reliable for recognizing true breakthroughs.
When The Improbable Becomes Inevitable
History shows time and again that many of the most transformative companies looked improbable when they first appeared.
When Airbnb pitched investors in 2008, many rejected the idea that strangers would rent out rooms in their homes to travelers. Today the company is worth tens of billions of dollars and has reshaped the global hospitality market.
In the early 1980s, the personal computer market was tiny and fragmented. A model trained on the data of that era might reasonably conclude that PCs would remain a niche tool for hobbyists and engineers.
In the early internet era, few people trusted the internet with personal information. Privacy fears dominated public discourse. A system analyzing the sentiment of the time might conclude that users would never willingly upload their social lives online.
Yet companies like Microsoft and Facebook were built precisely on the failure of those conclusions.
Breakthroughs, by definition, do not look like extensions of what came before. They look premature at best, impossible at worst.
Which is exactly what a well-calibrated AI system would likely say about them.
This is not a flaw but rather a feature of AI. It is built to reward explanations grounded in existing evidence.
It is precisely what makes AI so useful for testing assumptions about known markets. It is also what makes it unreliable for evaluating companies that are poised to transform established markets.
Energy: The Next Paradigm Shift
Energy shows this dynamic clearly.
Consider AI evaluating a startup building small modular nuclear reactors. The historical record shows construction timelines and regulatory hurdles stretching from years into decades. Three Mile Island. Chernobyl. Fukushima. As well as a long list of failed commercialization attempts dating back to the 1950s.
But that history does not necessarily inform the future. Small modular reactors are fundamentally different from the large, bespoke nuclear plants of the past. Manufacturers design them to be factory built and standardized at scale.
The economics have also shifted. AI data centers require enormous amounts of continuous, reliable power. Companies such as NVIDIA, Microsoft, Google and Amazon are already signing agreements and making investments tied to nuclear generation.
An AI system will quite possibly see a technology that has failed repeatedly and expensively. A human investor who understands what has changed might see a technology whose time has finally arrived.
Finding Outliers
Great startups often look irrational at first. They depend on behavioral shifts and technological leaps that have not yet fully materialized. They do not confirm existing patterns but rather create new ones.
If investors increasingly rely on AI to screen opportunities, they may unintentionally tilt their portfolios toward safer bets that extend today’s markets. But that is a different proposition, a different game from finding and betting on the outliers that define entire generations of innovation.
None of this means investors should avoid using AI. It has already become an indispensable research tool. It can map industries, analyze technologies and pressure test business models. Used well, it makes the diligence process faster and more rigorous.
The challenge is to be mindful of what AI can and cannot do.
AI can tell you whether a startup makes sense given the world as it is. What it cannot reliably do is recognize the moment when the future breaks away from the past. It cannot tell you whether the world is about to change in ways that make the startup necessary.
That judgment still requires something AI does not possess: human imagination.
And for venture capital, imagination may remain the most important signal of all.
