Two years ago, when IBM's annual CEO survey asked chief executives what would primarily drive their business growth by 2026, nearly half said generative AI. The 2026 survey has arrived, and the number who say AI is primarily driving growth sits at 10%.
The IBM Institute for Business Value polled 2,000 CEOs across 33 countries and 21 industries this spring, in partnership with Oxford Economics. The results document what the report calls a "structural gap between CEO optimism and enterprise readiness." In 2024, 49% predicted advanced AI would be their primary growth engine by now. That forecast moderated to 15% a year later. The actual figure came in at 10%. Meanwhile, 53% of CEOs say they are still primarily in piloting and experimentation mode.
That is not straightforwardly a story of failure. The same survey finds 70% of CEOs say AI has already begun transforming the core parts of their business. AI now makes 25% of operational decisions without human intervention, typically in areas like pricing updates, inventory allocation, and automated incident remediation, and CEOs expect that share to nearly double to 48% by 2030. Three-quarters of organisations now have a Chief AI Officer, up from just a quarter in 2025. The technology is clearly doing things. It is not yet reliably producing new revenue at the scale that was expected.
The explanation for that gap may lie less in enterprise execution and more in what AI is actually automating. A piece published this week by Oxford economist Carl Benedikt Frey argues that AI productivity growth will likely underwhelm relative to expectations, because it automates something fundamentally different from what the computer revolution automated.
The personal computer and the internet removed friction from information: finding it, storing it, transmitting it. The gains were relatively direct. A researcher who Googled a source got the same information they would have found in a library, just faster. Computers, when they did core work, did it deterministically. A spreadsheet could propagate bad inputs, but it did not invent arithmetic.
AI automates the production of cognitive outputs themselves: writing, coding, analysis. It often does this well. But because AI can be confidently wrong in ways that look plausible, it creates what Frey calls a "verification tax." When a professional uses AI to draft a document, the time saved in generation is partly offset by the time required to check the output. In settings where accuracy is non-negotiable, that offset can be substantial, and can fully eliminate the productivity gain.
The illustration is pointed. Sullivan and Cromwell, one of Wall Street's most prestigious law firms, recently filed an emergency court motion riddled with fabricated citations and AI-generated errors. The mistakes were caught not by the firm's own review process but by opposing counsel. It is a single episode, but Frey argues it is diagnostic: the cost of AI errors is shifting. As systems become more agentic, acting autonomously across complex workflows rather than just generating text in response to prompts, the consequences of unchecked mistakes grow.
The empirical evidence on productivity is mixed and task-specific. A large study of AI in customer support found a 14% productivity gain on average, with substantial gains for novice workers and little benefit for experienced ones. A randomised trial of experienced open-source developers found that access to frontier AI tools made them about 19% slower, with the lost time going into prompting, reviewing, and correcting. The conclusion Frey draws is that AI's payoff depends heavily on task structure: where errors are cheap and outputs are easy to test, the tool accelerates work; where correctness is hard to observe, the bottleneck shifts from doing the work to certifying it.
What changes this picture is verification infrastructure: the audit trails, compliance standards, professional norms, and institutional frameworks that allow AI outputs to be trusted rather than checked from scratch each time. These do not evolve at the speed of model releases. A federal judge in Texas has already begun requiring lawyers to certify that AI-drafted filings have been verified using traditional legal research. That kind of institutional adaptation, Frey argues, is what the productivity boom actually depends on. Until it arrives, the gains that CEOs keep predicting will remain concentrated in the places where accountability is loosest and errors are cheapest to absorb.