Last year, an MIT report that 95% of businesses got zero returns from AI made headlines. Well, the NBER has just published an even bigger survey of more than 6,000 executives in the US, UK, Germany and Australia on their AI use and what they got out of it over the last three years. Plus, they asked what executives expected for the next three years.
First, the basics. 69% of companies reported using AI at the moment, with text and image generation via LLMs the most common use case. On average, AI tools are used 1.5 hours per week, with three in ten companies reporting that they do not use any AI tools on a regular basis, and only 7% reporting AI tool use of more than five hours a week.
However, AI adoption is rising rapidly. In the first wave of the survey from February to April 2025, 61% of UK companies reported the use of some AI tools. By January 2026, that had risen to 71%.
Where it gets interesting is to look at the past and expected gains in productivity and potential losses in employment. Let’s look at productivity first, where businesses report hardly any productivity gains in the last three years. On average, productivity has increased by some 0.3%, which is very much in line with the rather pessimistic expectations of Daron Acemoglu, who thinks that productivity gains will be in the range of 0.1% per year for the coming decade.
Yet, companies still have high hopes for the future and expect productivity gains to accelerate by a factor of five to 1.4% over the next three years, driven mostly by US and UK businesses, which expect productivity gains to rise by a factor of eight to ten.
A separate survey of US employees finds reported productivity gains of 0.5% over the last three years and an expected doubling of productivity gains to 0.9%.
Productivity gains from AI
Source: Yotzov et al. (2026)
When it comes to employment, however, employers and employees start to disagree fundamentally. US employees generally expect AI to generate new jobs, while employers report no impact on employment in the last three years, but expect to see a significant reduction in employment in the future. Again, US and UK businesses seem most optimistic about the potential of AI. US businesses expect employment in their companies to drop by 1.2% in the next three years, while UK businesses expect to cut jobs by 1.4%.
Employment impact of AI
Source: Yotzov et al. (2026)
I think this survey acts like a Rorschach test on AI. If you are optimistic about AI’s possibilities, you can look at these charts and say that AI adoption is increasing rapidly, and the power of AI tools is growing so fast that productivity gains are going to be large. I guess that Fed Chair-designee Kevin Warsh would make that argument to justify more rate cuts.
AI sceptics, in contrast, can argue that after three years, when AI could have made the biggest impact on productivity since it is starting from scratch and the low-hanging fruit is plentiful, the productivity gains are still only 0.3%. Plus, use cases seem to be restricted mostly to writing texts and creating images (some would say corporate AI slop), where productivity gains will stop pretty quickly going forward. So, it looks a little bit like AI may be more a promise than a reality and productivity gains always just out of reach, but never here.
Which side do you fall on?
FWIW, in Software Engineering I would estimate the productivity gains based on my team's increased speed and quality of development more in the range of 25-50% (conservative estimate).
I'm sceptical of employee responses on AI usage surveys, because turkeys don't vote for Christmas. "Thanks for the enterprise subscription, boss; it's so good I think I'll soon be redundant!"
It reminds me very much of Nobel Laureate Robert Solow's quip "You can see the computer age everywhere but in the productivity statistics." https://en.wikipedia.org/wiki/Productivity_paradox
Another problem that the Solow paradox suggests is that even objective productivity data might understate AI's impact, at least for now. The historical parallels with steam and electricity are instructive; dividends of a productivity-enhancing disruptive technology were reaped only slowly, with an initial lag over decades, due to the time required for the technology to diffuse into common use and for organisations to reorganise around it. We're still very early in that reorganisation phase with AI.
I heard a podcast yesterday https://youtu.be/OzeJsrhh4k0 interviewing a hedge fund manager who thinks the gloom-amd-doom resulting from that viral Substack post https://www.citriniresearch.com/p/2028gic is greatly overstated; "what did all those unemployed elevator operators and door-to-door knife sharpeners end up doing?"