What to Expect from Artificial Intelligence—and What Not To
Meteorologists can predict tomorrow's weather with 95 percent accuracy. For ten days out, that accuracy drops to the level of a coin toss. If someone is selling you AI as a tool that will transform your strategic decision-making and predict the future of your business, they should talk to a meteorologist first.
The Hype That Overtook Reality
Over the last three years, we have witnessed an unprecedented volume of predictions about what AI can do: it will replace this profession; it will uncover patterns that humans miss; it will predict who will leave, what they will buy, and how they will decide. Part of this is true. A large part is not.
The problem is not with the technology. The problem lies in what we expect from it. And this is where most discussions about AI go wrong, because they confuse two very different things: efficiency and prediction.
AI is genuinely exceptional at efficiency: handling repetitive cognitive tasks, sorting, summarizing, and generating first drafts. But as soon as the ambition jumps from efficiency to predicting complex human systems, the logic falls apart.
The Weather Problem
In the 1960s, Edward Lorenz, an American meteorologist and mathematician, noticed a curious phenomenon while running weather simulations: a tiny change in initial conditions led over time to a dramatically different outcome. Rounding off a single number by a thousandth produced a completely different forecast. He described this sensitivity to initial conditions as what we know today as the butterfly effect.
This is why meteorologists—equipped with excellent models, satellites, and supercomputers—fail on a ten-day horizon. Not because they are bad at their jobs, but because the atmosphere is a chaotic system, and in a chaotic system, small errors compound exponentially.
Weather has one distinct advantage: it is governed by physical laws that do not change. The atmosphere does not alter its laws simply because we are observing it. Human systems do not share this advantage.
Three Reasons Why AI Cannot Handle Human Systems
Correlations Are Not Causes
AI learns correlations from historical data. It observes that customers who did X subsequently did Y. But it does not understand why. Researchers working on causal inference, such as Susan Athey from Stanford, state this openly: machine learning improves prediction when the environment is stable; it does not explain why behavior changes.
Research from Microsoft Research analyzed extensive datasets on social diffusion and cultural adoption. Models predicting future human behavior achieved a correlation of at most $r = 0.2$. An AI model stating "this customer will churn" is accurate less than 55 percent of the time—barely better than a coin flip.
Tacit Knowledge Cannot Be Taught
Decades after the rise of IT automation, technology continues to deliver uneven productivity gains. Technology efficiently replaces repetitive tasks, but real productivity rises only when people adapt and change how they work.
The missing ingredient is tacit knowledge: judgment, taste, timing, political nuance, and the ability to read a situation where the rules differ from what is written. AI cannot learn what is not explicitly expressed, and the most valuable human knowledge rarely is.
Systems Change Faster Than Models
Even if we fully understood causality and managed to capture tacit knowledge, predictions would still fail. Economists know this well as Goodhart's Law: as soon as a metric becomes a target, it ceases to be a good metric. People react to incentives; they adapt to rules; they alter their behavior simply because they know they are being measured.
A one-day weather forecast is 95 percent accurate; five days out, 75 percent; ten days out, 50 percent. In human systems, degradation happens much faster and far less predictably because the system itself reacts to the prediction.
Why It Sounds So Convincing Anyway
AI possesses one extraordinary talent: it sounds confident. It answers fluently, structurally, and in a format that looks like rigorous analysis. However, this format is entirely independent of accuracy.
Marshall McLuhan warned us long before the AI era: every new medium reshapes the way we think. Television didn't teach people to process information critically; it taught them to consume images. Social media accelerated this trend by creating echo chambers. AI accelerates it further by generating confident answers to every question, framed to sound self-evident.
The greatest risk of AI, therefore, is not that it will replace us. It is that it will entice us to stop thinking.
Where AI Truly Delivers Value
Only now, after acknowledging all its limitations, do we arrive at what matters most: where AI delivers real value. It holds true where three conditions are met:
The environment is stable and the rules do not change.
Plenty of high-quality historical data exists.
The task is bounded and specific, not open-ended and strategic.
In practice, this means: automating routine documentation, sorting and routing customer inquiries, extracting information from unstructured documents, assisting with data analysis, drafting initial copy, or summarizing meeting notes and reports.
These are not minor wins. According to McKinsey data, the average manager spends over 60 percent of their time on tasks that could be fully or partially automated. Freeing up that capacity and redirecting it toward genuine strategic thinking is a valuable transformation—just not the one most often talked about.
What This Means for Your Organization
Meteorologists didn't stop making forecasts just because a ten-day horizon fails. They adapted their work to reality: short-term forecasts are accurate and actionable, while long-term scenarios are explicitly marked as indicative.
The same approach applies to AI. Be precise about what you demand from it. Deploy it on bounded problems with available data and stable environments. And leave strategic decision-making where it belongs: in the hands of people who understand context, grasp what data doesn't show, and can read a situation that is unlike anything in history.
AI is not a crystal ball. It is a powerful tool with clear boundaries. Companies that grasp this sooner won't gain an edge by trusting AI more, but by deploying it smarter.