AI forecasting systems are closing the gap with expert human forecasters, but the bigger enterprise opportunity may lie in identifying which variables can actually change outcomes rather than simply producing more accurate predictions.
30 September 2026 — Artificial intelligence is becoming increasingly capable of forecasting future events, with recent benchmarking showing AI systems approaching the performance of skilled human forecasters on questions spanning geopolitics, technology, economics and other real-world events.
That progress creates an obvious opportunity for businesses. Better forecasts could help companies anticipate demand, identify emerging risks, assess market shifts and make strategic decisions earlier. However, the growing accuracy of AI forecasting also raises a more difficult question: what should organisations actually do with those predictions?
IBM’s latest analysis argues that forecasting accuracy alone is not enough. Even a highly accurate model has limited value if a company cannot identify the variables that matter most, decide which of them it can influence and translate the resulting forecast into a practical strategy.
AI forecasting is getting closer to expert humans
Forecasting has traditionally depended on a mixture of statistical models, specialist knowledge and human judgement.
Businesses have long used time-series models for tasks such as sales forecasting, inventory planning and financial projections. Large language models extend that capability because they can work with much broader forms of information, including news, reports, qualitative evidence and changing geopolitical or market conditions.
The clearest evidence of that progress is emerging from forecasting competitions.
Metaculus, an online forecasting platform, operates FutureEval, a programme specifically designed to benchmark AI systems against human forecasters on real-world questions. Its methodology compares models and AI-powered forecasting bots with both the wider forecasting community and selected professional forecasters.
In its Spring 2026 analysis, Metaculus found that the gap between its professional forecasting team and leading bots had narrowed sharply and was no longer statistically significant.
The trend has continued into forecasting tournaments. A bot won Metaculus’s Market Pulse tournament in 2026, while the Summer Metaculus Cup also saw several AI systems move into the highest ranks of a competition involving human forecasters.
These results do not mean AI can reliably predict every future event. Benchmark performance depends heavily on the kinds of questions being asked, the available information and how outcomes are scored. But they show that AI forecasting has advanced beyond purely experimental demonstrations.
Being three percent more accurate may not be the real problem
The temptation for businesses is to respond by looking for whichever forecasting model produces the highest benchmark score.
Keith Dear, CEO and co-founder of Cassi AI, which developed one of the higher-performing bots on Metaculus, argues that this misses the larger challenge.
“For most customers, a 3% better forecast is not the thing holding them back.”
Instead, Dear argues that businesses gain more value by identifying the factors most likely to change an outcome and concentrating resources on those areas.
That distinction separates prediction from decision-making.
A company might predict with high confidence that customer churn will increase next quarter. The forecast itself does not determine whether management should change pricing, improve customer support, redesign onboarding or invest in a new product feature.
The commercially valuable part of forecasting is understanding what drives the expected outcome and which interventions are available.
Forecasting requires a theory of what matters
Good forecasting begins before a probability is produced.
Forecasters must decide which information is relevant, identify important causal factors and separate useful signals from noise. According to IBM’s analysis, businesses hoping to use AI effectively may need to replicate that process rather than treating the model as a machine that simply delivers answers.
This is where human expertise remains particularly important.
An AI system can process enormous volumes of information and update forecasts quickly, but businesses still need to determine which questions are worth asking.
For example, predicting whether a competitor will launch a product in the next six months may be useful only if the organisation knows how that information would affect its own decisions.
Similarly, predicting whether a certain technology will achieve widespread adoption has limited value if the company has no clear framework for deciding when or how much to invest.
Forecasting systems therefore need to sit inside a wider decision architecture.
Businesses need to forecast decisions, not just events
One practical way to think about enterprise forecasting is to distinguish between questions that are interesting and questions that are actionable.
A business may want to know whether a recession will occur, whether a regulatory change will pass or whether a new technology will become mainstream.
Those are potentially useful forecasts, but managers also need to ask a second set of questions.
What would the organisation do if the probability rose from 30% to 60%? At what point would it alter investment, hiring or inventory? Which decisions are reversible and which involve long-term commitments?
This moves forecasting closer to scenario planning.
Rather than producing a single prediction and waiting to see whether it comes true, organisations can use AI to update probabilities continuously and link different probability ranges to predefined actions.
That could make forecasting far more useful than simply improving the accuracy of an annual planning exercise.
Human forecasters still offer valuable lessons
One of the human forecasters highlighted by IBM is Yann Rivière, who ranks among the strongest competitors on Metaculus and has placed predictions on subjects ranging from geopolitical events and rocket launches to cryptocurrency prices. IBM reports that his historical tournament performance places him in the 99.99th percentile among thousands of forecasters.
Rivière acknowledged that AI competitors have become increasingly difficult to beat, particularly in tournaments.
That competition illustrates an important point for businesses: AI may be improving rapidly, but expert human forecasters remain useful partly because their methods are not limited to producing probabilities.
Experienced forecasters constantly update beliefs, challenge assumptions, consider alternative explanations and examine why a forecast might fail.
Those habits may be as important for organisations as access to the forecasting models themselves.
Human-AI forecasting may be stronger than either alone
The next phase of forecasting is therefore likely to involve hybrid systems rather than a straightforward replacement of human judgement.
A 2026 pilot study using Polymarket as an externally resolved benchmark found that human-AI collaboration produced very different outcomes depending on how people used the model. Many participants either deferred to AI or used it to reinforce existing beliefs, while a smaller group combined machine output with genuinely complementary reasoning and achieved much stronger results.
The study linked stronger hybrid performance to traits including intellectual humility, curiosity and perspective-taking rather than simply raw cognitive ability. The researchers describe the findings as preliminary, so they should not be treated as a definitive model for enterprise deployment.
Still, the findings reinforce the idea that giving employees access to better AI forecasts does not automatically improve decision-making.
Organisations also need processes that encourage people to question model outputs rather than either accepting or rejecting them instinctively.
Forecasting at AI speed changes organisational behaviour
AI also changes forecasting through speed and scale.
Human teams can track only a limited number of questions. AI agents can potentially maintain hundreds or thousands of forecasts simultaneously and update them whenever new information arrives.
That could allow businesses to monitor questions such as whether a supplier will miss a deadline, a product launch will be delayed, a market will enter recession or a regulatory proposal will become law.
But more forecasts create another problem: attention.
An organisation that receives thousands of updated probabilities every day may simply replace information scarcity with information overload.
The value of an enterprise forecasting platform therefore depends on prioritisation. Systems need to surface predictions that are both important and actionable, while suppressing changes that have little practical consequence.
AI forecasting could become part of enterprise risk infrastructure
If forecasting systems continue improving, their role may eventually resemble risk-monitoring infrastructure.
Instead of being used only by strategy departments, forecasting could be embedded across finance, procurement, cybersecurity, supply chains and operations.
A procurement team might monitor the probability of disruption across suppliers. A finance team could track changing probabilities around interest rates or currency conditions. A technology company could monitor whether particular regulatory or competitive developments are becoming more likely.
The common requirement is that forecasts must be connected to decisions.
Without that connection, increasing accuracy risks becoming another technical metric that looks impressive on a dashboard but has limited operational impact.
Prediction is becoming cheaper; judgement remains scarce
The rise of AI forecasting reflects a broader pattern in enterprise AI.
As models become capable of producing analysis, content and predictions at lower cost, the scarce resource shifts elsewhere.
In forecasting, that scarce resource is increasingly judgement: knowing which questions matter, determining which evidence is trustworthy and deciding how to respond when probabilities change.
AI can make organisations faster at estimating what might happen.
The companies that benefit most, however, may be those that become equally good at deciding what they can do about it.
The central question for enterprise leaders is therefore not whether AI can forecast better than humans.
It is whether organisations can build decision-making systems capable of turning those forecasts into better outcomes.
Sources
- IBM Think — AI forecasters are catching up to humans. It’s a new opportunity for businesses—if they hire the humans too, Nick Gallagher, 28 September 2026. https://www.ibm.com/think/news/ai-forecasters-new-opportunity-business-hire-humans-too
- IBM Think Newsletter — Why accuracy in AI forecasting isn’t enough, 29 September 2026.
- Metaculus — FutureEval AI forecasting benchmark. https://www.ibm.com/think/news/ai-forecasters-new-opportunity-business-hire-humans-too
- Metaculus — FutureEval methodology and human-versus-bot benchmarking. https://www.metaculus.com/futureeval/methodology/
- Metaculus — Spring 2026 AI Forecasting Benchmark Series. https://www.metaculus.com/aib/2026/spring/
- Ming, V. — Human Capital, Not Model Benchmarks, Predicts Hybrid Intelligence in Forecasting, 2026. https://arxiv.org/abs/2607.02467
