Outside software development, the evidence that AI is improving productivity remains thin. The barriers are less technical than organisational, argues Sam Gilbert.
Huge investments are being made on the assumption that generative AI is about to deliver transformative gains in economic productivity. Between them, Microsoft, Meta, Amazon and Google are expected to spend more than $600bn on AI infrastructure in 2026, while global capital expenditure on advanced chips, data centres and power generation looks set to exceed $1 trillion – more than the capex of the entire oil and gas industry. Analysis by The Economist suggests AI is on track to be the largest investment boom in history – bigger than the dotcom boom; bigger even than the railway and canal booms in the 19th century. But to what extent are these investments justified by real evidence of productivity improvements?
AI adoption is broad but shallow
ChatGPT now has more than one billion active users, a milestone reached faster than Instagram or TikTok. But striking adoption statistics like these can obscure how much most people are using AI tools. Gallup polling suggests that 52% of US employees use AI at work, but only 15% use it every day. Similarly, a National Bureau of Economic Research survey of ~6,000 executives in the US, UK, Germany and Australia found that while more than two thirds use AI, they only use it for an average of 1.5 hours per week.
Analysis of companies’ spending on AI tells a similar story: the payments platform Ramp, which processes transactions for around 70,000 businesses, reported in July 2026 that while the top 1% of firms spend roughly $7,500 per employee per month on AI, the median spend is only $12 per employee per month.
Coding is the “killer app”
Software development is the one domain where the productivity case for AI is more or less proven. More than half of all new code is now AI-generated, and analysis by the venture capital firma16z indicates AI companies’ coding-related revenues are more than double everything else combined. Academic working studies of AI coding tools like Cursor and Codex consistently show productivity improvements in the range of 20-40% (Cui et al 2025, Agarwal et al 2026, Sarkar 2026, Becker 2026).
But even here, the gains are contingent. Drawing on commit data from around 100,000 developers, research by Yegor Denisov-Blanch and colleagues at Stanford finds that the biggest productivity benefits come from low-complexity tasks and new projects using popular programming languages such as Python. For complex tasks on legacy codebases in less common languages like COBOL, the gains from using AI shrink dramatically or even turn negative. As public sector technology development is predominantly “brownfield”, this is an important finding for policymakers to take on board.
The better the evaluation, the smaller the effect size
Outside coding, most deployments of generative AI that have been independently evaluated involve AI assistants – either public-facing chatbots or internal-facing “copilots”. In general, the more rigorously these are studied, the less impressive the results look.
The UK Civil Service is a case in point. A cross-government trial of Microsoft Copilot, involving 20,000 civil servants, claimed average time-savings of 26 minutes per day based on self-reporting. But a subsequent controlled study in the Department of Work and Pensions found time-savings were more modest, while the Department for Business and Trade’s evaluation found no evidence of improved productivity, with some tasks slowed by the need to correct poor-quality outputs. This resonates with the Bennett School finding that using AI chatbots does not make workers more productive in and of itself.
In North America, the Canada Revenue Agency reported that its chatbot “Charlie” gave accurate answers to user queries in 70-90% of cases in internal testing, but an external evaluation by the Auditor General found it gave correct answers only a third of the time. Meanwhile, New York City withdrew its own citizen-facing chatbot after investigative journalists demonstrated that it was systematically giving unlawful advice to landlords about tenants’ rights. Little wonder, then, that according to KPMG only 7% of business leaders are currently reporting a positive return on their AI investments.
The most encouraging case is instructive about why. Caddy, an internal copilot built by the UK Citizens Advice with the Incubator for AI, was evaluated through a randomised controlled trial and produced positive results: response times to citizen queries were halved, with advisers much more likely to report high confidence in the advice they gave.
Rather than seeking to replace advisers, Caddy was specifically designed to help them become more productive. It may be that in customer- and user-support contexts, realising productivity gains is dependent on a “human in the loop”. This would be consistent with research showing that AI is not a “plug and play” technology, but depends on strong management practices to deliver results.
Rising costs and externalities
Another factor that may limit productivity gains from AI is that the scope for cost-saving opportunities is sometimes smaller than assumed. In functions like insurance claims, much cost has already been stripped out through decades of business process optimisation, outsourcing, and offshoring, meaning there is not much left for AI to unlock. At the same time, AI services are getting more expensive: having subsidised usage costs to drive growth between 2022 and 2025, investors in OpenAI and Anthropic now expect margin improvements from higher prices as the companies prepare to go public in 2026.
Even where productivity gains from AI are real, they are partly offset by externalities associated with the broad adoption of AI tools. Employees who propagate “workslop” – AI-generated reports, presentations and emails, that appear polished but lack substance – create an extra burden for their colleagues. According to the Harvard Business Review, more than four in ten workers report having encountered workslop, estimating that each instance costs them nearly two hours in rework. Meanwhile, a recent survey found that almost two-thirds of professionals have uploaded work-related information to public AI tools, exposing their employers to material data protection and commercial confidentiality-related risks.
There is also a demand-side externality of AI adoption with serious consequences for the public sector. Chris Schmitz and colleagues at the Hertie School have documented 84 cases of what they call “service flooding”, where AI-assisted applications, appeals and objections overwhelm the operational capacity of support functions.
Policy implications
To paraphrase American economist Robert Solow, AI is everywhere except in the productivity statistics. There are surprisingly few instances of AI-related productivity gains being realized outside software development, while evidence of negative externalities is beginning to emerge. This should moderate policymakers’ expectations about the potential for AI to relieve pressure on public services and improve national productivity – and about the timeframe in which this can be expected to happen. Historically, major technological breakthroughs have taken decades to transform the economy, and as yet, there are few reasons to believe AI is a special case.
The views and opinions expressed in this post are those of the author(s) and not necessarily those of the Bennett School of Public Policy.
