Artificial intelligence (AI) is spreading through business at extraordinary speed. But productivity is not moving at the same pace.
Research published by the National Bureau of Economic Research in 2026, based on nearly 6 000 senior executives across the US, UK, Germany and Australia, found that 69% of businesses were actively using AI. Yet 89% reported no impact on labour productivity from AI over the previous three years.
Interestingly, those same executives expect AI to increase labour productivity by an average of 1.4% over the next three years.
That expectation may be as revealing as the absence of gains so far. Business leaders appear to believe that the productivity dividend is still ahead of them. The question is what has to change for that expectation to become reality.
AI capability, adoption and investment are accelerating, while research increasingly demonstrates productivity improvements in individual tasks. Yet much of that improvement has not translated into measurable productivity across organisations.
This does not necessarily mean AI is failing. It may mean the technology is advancing faster than the organisations adopting it.
We have seen this before
More than three decades ago, economist Paul A David examined a remarkably similar problem. In his influential 1990 paper, “The Dynamo and the Computer: An Historical Perspective on the Modern Productivity Paradox”, David considered why enormous investment in computers had not immediately translated into equally dramatic productivity growth.
AI can make tasks dramatically more efficient without making the enterprise dramatically more productive.
His comparison was electrification. Industrial factories had been designed around steam power. When electric motors arrived, manufacturers initially introduced the new technology into factories whose layouts and processes were still designed around the limitations of steam.
The technology changed faster than the organisation around it.
Larger productivity gains emerged when manufacturers redesigned factories around electricity. Production lines, factory layouts, workflows and management practices changed.
Electricity did not merely replace steam. It eventually changed the factory.
That history provides a useful lens for today's AI productivity debate.
Task productivity is not enterprise productivity
There is compelling evidence that AI can improve individual productivity. In a peer-reviewed study published in The Quarterly Journal of Economics, Erik Brynjolfsson, Danielle Li and Lindsey Raymond examined the deployment of a generative AI assistant to 5 172 customer-support agents. Access to the tool increased productivity, measured as issues resolved per hour, by 15% on average, with substantially larger gains among less experienced and lower-skilled workers.
The question is therefore increasingly not whether AI can make people more productive. It can.
The harder question is why those gains are more difficult to see at enterprise level.
McKinsey's global AI research found widespread organisational adoption, but only 39% of respondents attributed any enterprise-level EBIT impact to AI.
This exposes an important distinction: AI can make tasks dramatically more efficient without making the enterprise dramatically more productive.
If AI cuts preparation of a four-hour management report to two hours, task productivity has improved dramatically. But if the same report, approvals, meetings and workflow remain, how much of those two hours does the enterprise actually capture?
This may be why AI productivity appears simultaneously impressive and disappointing. We may be electrifying the old factory.
There are other explanations
The analogy should not be taken too far. AI is not electricity. Electricity transformed the physical organisation of production by distributing motive power. AI potentially changes the cognitive organisation of production: it can generate and interpret information, make recommendations and increasingly take actions.
There are also other explanations for today's apparent productivity gap. It may simply be early; general-purpose technologies take time to diffuse. There may be measurement problems, with improvements in quality, decision speed, innovation and customer experience not fully captured by conventional productivity measures.
And today's AI still faces limitations in reliability, verification and accountability that constrain autonomous deployment in high-value environments.
Organisational design is therefore unlikely to be the only explanation. But it is one executives can directly influence.
McKinsey found that among 25 organisational attributes examined, workflow redesign had the strongest relationship with an organisation's ability to see EBIT impact from generative AI. Yet only 21% of respondents using generative AI said their organisations had fundamentally redesigned at least some workflows.
We are adopting the technology much faster than we are changing the organisation around it.
From AI-assisted to AI-native
Enterprise AI maturity can perhaps be understood across three stages.
The first is AI-assisted work. Humans remain at the centre of existing workflows, while copilots and assistants improve individual performance. The unit of productivity remains the employee.
The second is AI-enabled processes. AI begins operating across workflows − retrieving information, making recommendations, triggering actions and coordinating systems. Productivity increasingly becomes about cycle time, quality, throughput and cost-to-serve.
The third is the AI-native operating model, where organisations stop assuming existing processes must survive.
Consider a procurement process containing eight steps, multiple systems, several hand-offs and three approvals.
The conventional AI question is: Where can we insert AI into these eight steps? The AI-native question is: Why do we need eight steps at all?
That shift − from automating existing work to reconsidering why the work is structured that way − may contain the larger productivity opportunity.
But redesign does not mean eliminating every human or control. Approvals, segregation of duties and human review can provide accountability, prevent fraud and satisfy regulatory requirements. Some apparent organisational friction exists for good reason.
The challenge is determining which constraints remain necessary when machine intelligence becomes increasingly capable.
That might mean moving people from transaction processing towards exception handling, replacing blanket approvals with risk-based thresholds, giving AI agents bounded authority, or measuring performance around outcomes rather than activity.
The objective is not automation for its own sake. It is the deliberate redesign of work between humans and machines.
A rising tide of capability
Recent MIT FutureTech research provides another important perspective. Drawing on more than 60 000 worker evaluations covering more than 6 000 text-based workplace tasks, researchers found AI capabilities improving broadly across tasks rather than arriving primarily through sudden bursts of automation. They describe the pattern as more of a rising tide than a crashing wave.
If AI becomes capable of performing thousands of activities progressively better, transformation may not arrive as one dramatic automation event. It may arrive through thousands of smaller changes to what machines can economically perform.
Management then has to determine how those task-level gains accumulate into a different operating model.
Beyond the AI use-case factory
This challenges one of the dominant measures of enterprise AI progress: the number of use cases.
Organisations proudly report dozens or hundreds of AI initiatives. Executives monitor the number of copilots, assistants and agents deployed.
These are useful measures during experimentation. But imagine measuring electrification by counting the number of electricity use cases inside a factory.
Eventually electricity stopped being a use case. It became part of the infrastructure of the enterprise.
AI may follow a similar trajectory.
The mature organisation may move from: AI use cases → AI-enabled processes → AI-native operating models.
The question then changes from where can we use AI? to something more fundamental: How would we design this organisation if intelligent machines were available everywhere?
The CIO's next AI challenge
This changes the technology leadership agenda. The first generation of enterprise AI required CIOs to answer questions about models, platforms, data, architecture, security, governance and use-case prioritisation.
Those remain essential. But the next generation introduces harder questions: Which processes should disappear? Which decisions can machines make? Where must humans remain accountable?
These are not simply technology questions. They are operating-model questions requiring CEOs, CIOs, CFOs, COOs, CHROs and business leaders to work together.
Technology can create the productivity potential. Management must determine how the organisation captures it.
The evidence does not support the simplistic conclusion that AI has failed to deliver productivity. Nor does history prove that enormous productivity gains are inevitable.
AI is already producing measurable gains in individual tasks. Executives themselves expect larger productivity gains to emerge in the years ahead. The challenge is translating that potential into enterprise-level economic outcomes.
Some of that may require better models. Some may require time and better measurement. But some of it will require organisations to reconsider workflows, decision rights and management structures designed for a world in which human intelligence and attention were scarce.
Paul David's history of electrification reminds us that transformative technologies and transformative organisations do not necessarily arrive at the same time.
Electricity eventually changed the factory. AI may change the cognitive architecture of the enterprise.
The organisations that benefit most may therefore not be those that deploy the most models, copilots or agents. They may be those willing to reconsider how work itself should be organised.
The AI productivity paradox will not be solved by better models alone. It will also require better organisations.
Share
