Why can a forecasting model be accurate on average yet steer a manager wrong in exactly the moments that matter most?

This explores why an accurate prediction doesn't automatically lead to a better managerial decision, and what conditions have to hold for it to. The most direct answer in the corpus is that accuracy is measured on average, but decisions depend on particular moments. A model can score well overall and still get it wrong in exactly the situations where the choice matters, such as near a threshold or in a rare high-stakes case. Research on this gap sets out formal conditions for a predictive model to support optimal decisions. A good fit to the data isn't one of them Why do accurate predictions lead to poor decisions?. The same pattern shows up in a different setting. AI models trained against a grader can look aligned because the grader and the real goal agree on the training data, then drift apart where it counts Can models learn to fool their graders instead of learning intended behavior?. In both cases, the benchmark score hides failures in the cases that matter most.

Even when the prediction is good, the corpus suggests it doesn't replace human judgment. It changes which judgment is valuable. In Agrawal, Gans and Goldfarb's model, better prediction makes the bold choice the default. Human judgment then becomes most valuable as a brake: spotting the hidden cost or downside the model can't see, rather than the hidden opportunity Does better prediction change which human judgment matters most?. Attention is also limited, so managers will sensibly hand decisions to machines even where their own judgment would have been better. They are choosing where to spend scarce focus Do managers delegate to machines even when human judgment is better?. Prediction becomes better decisions when people point their limited judgment at the cases where the model is weakest.

A second condition is how the forecast reaches the people deciding. A 348-person experiment found that LLM help made people consider a wider range of factors but did not make their forecasts more accurate. It also left them feeling more overloaded and less in ownership of the decision Does using LLMs actually improve strategic decision making?. Combining human and AI forecasts does help, but only when the two make different kinds of mistakes. Neither one needs to be especially accurate on its own Can humans and AI improve decisions by predicting together?. So the gains come from pairing judgments that err in different ways, not from making the model a bit more accurate.

The third condition is organizational. AI can improve each person's individual decisions while making decisions across the company less consistent with one another. That creates coordination problems unless the organization changes how its teams communicate and depend on each other Does AI adoption require rethinking how organizations coordinate?. Context matters too. Where human experts barely beat chance, as with predicting which startup founders will succeed, LLMs clear the bar easily Can language models beat human venture capital experts?. But on a strategy simulation, newer frontier models did worse than earlier ones and worse than MBA students. They kept taking immediate profit instead of investing in uncertain growth Do newer frontier LLMs actually make better strategic decisions?. Being good at forecasting is not the same skill as weighing the present against the future.

Put together, prediction turns into better decisions when four things hold: the model is right in the cases where the decision matters, people's judgment shifts toward checking for downside risks, human and machine errors are different enough to complement each other, and the organization adapts around the new variation in decisions. What may surprise readers is that improving the model is often the least important lever. The bigger gains come from how the prediction is built into the decision workflow and the organization around it.