Imagine a company whose employees suddenly become 20 per cent faster at writing reports, reviewing documents and preparing presentations.
Would you expect its profits to increase by 20 per cent?
Probably not. Yet much of the enthusiasm surrounding artificial intelligence rests on an assumption remarkably close to that.
Workers save time. Companies become more productive. Economic growth accelerates.
The theory is attractive. The evidence is proving more complicated.
A study published by the National Bureau of Economic Research in March 2026 surveyed nearly 750 corporate executives about their use of artificial intelligence.
More than half of the firms had already invested in AI. Executives reported improvements in productivity, particularly in finance and highly skilled services.
But the researchers identified an important discrepancy: perceived productivity improvements were greater than the gains companies could measure.
That finding deserves more attention than another announcement about a revolutionary AI product.
It suggests that the technology may be improving individual activities faster than businesses are learning to convert those improvements into financial results.
The missing hour
Suppose an analyst previously spent four hours preparing a report and can now complete it in two.
The immediate productivity gain is obvious.
But what happens to the two hours saved?
If the analyst uses them to prepare another report that nobody needs, the business has gained very little.
If the company can handle more clients with the same workforce, shorten delivery times or improve the quality of its decisions, the economic value becomes more substantial.
And if the analyst spends the time correcting errors introduced by an unreliable AI system, the apparent efficiency gain may largely disappear.
This is why task-level productivity cannot be treated as a reliable measure of company-wide performance.
Businesses are systems of interdependent activities.
Speeding up one stage does not necessarily improve the whole system. A faster sales proposal is of limited value if customer approval, production or delivery remains unchanged.
The economic benefit emerges when the organisation changes how work is performed.
The technology is not the entire investment
Companies often discuss AI adoption as though purchasing a software subscription marks the beginning of transformation.
In practice, software may be the simplest part.
The harder questions concern data quality, employee skills, accountability and the redesign of established processes.
Who checks the output? Which decisions can be automated? When must a human intervene? What happens when a system makes a mistake that affects a customer?
These are management questions with direct financial consequences.
Poorly organised firms can introduce sophisticated technology without substantially improving their performance.
Well-managed firms may achieve meaningful gains from relatively modest applications because they know precisely which processes need improvement.
This helps explain why adoption rates alone reveal little about the economic return from AI.
A company with thousands of employees using AI tools is not necessarily more productive than one that has automated a handful of costly operations.
The relevant measure is the value created after accounting for implementation, supervision and operating expenses.
What about jobs?
The labour-market evidence also deserves a more careful reading.
The NBER research found little evidence of an immediate aggregate decline in employment attributable to AI, although larger companies anticipated reductions and smaller firms expected modest employment gains.
The study also identified changes in the composition of labour demand, with routine administrative roles facing pressure while demand for more specialised technical skills increased.
This does not establish that widespread displacement will never occur.
It does suggest that the initial economic effects are more complicated than the familiar prediction that machines will simply replace workers.
In some companies, AI may reduce headcount. In others, it may allow existing teams to serve more customers or develop products that previously would have been uneconomic.
The outcome depends partly on how management responds to the technology.
And that is precisely why broad forecasts of AI-driven employment losses should be treated cautiously.
The productivity boom may be uneven
Another NBER study, published in September 2026, examined the potential macroeconomic impact of AI through improvements in software engineering.
Its modelling suggests that the productivity benefits could be substantial, particularly if more efficient software development also accelerates research and innovation.
But those estimates reflect a model informed by financial-market expectations. They should not be confused with realised increases in national output.
The distinction is crucial.
Financial markets can capitalise expected future productivity long before the economy records the corresponding gains.
Companies can report promising experiments long before those experiments produce stronger earnings.
That does not invalidate the technology.
It means the economic transformation is still being measured, tested and translated into business practice.
The most interesting AI companies over the next few years may not be those with the greatest number of users or the most impressive demonstrations.
They may be the businesses that learn how to reorganise work around capabilities that were previously too expensive or too slow.
The question is no longer whether AI can make an individual task faster. We have enough evidence that it often can.
The difficult question is whether companies can turn those saved hours into something customers value and investors can measure.
That is where the real productivity revolution will either happen or fail to materialise.
